Information Fusion 55 (2020) 220–244

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

Full Length Article

Image denoising review: From classical to state-of-the-art approaches

Bhawna Goyal a,∗, Ayush Dogra b, Sunil Agrawal a, B.S. Sohi c, Apoorav Sharma a
a
UIET, Panjab University, Chandigarh, India
b
CBME, IIT Ropar, Punjab, India
c
Chandigarh University, Gharuan, Punjab, India

a r t i c l e i n f o a b s t r a c t

Keywords: At the crossing of the statistical and functional analysis, there exists a relentless quest for an efficient image
Denoising denoising algorithm. In terms of greyscale imaging, a plethora of denoising algorithms have been documented in
Spatial the literature, in spite of which the level of functionality of these algorithms still holds margin to acquire desired
Transform
level of applicability. Quite often noise affecting the pixels in image is Gaussian in nature and uniformly deters
Hybrid
information pixels in image. Based on some specific set of assumptions all methods work optimally, however
Filters
PSNR they tend to create artefacts and remove fine structural details under general conditions. This article focuses on
classifying and comparing some of the significant works in the field of denoising.

1. Introduction 1.1. Noise suppression

Due to inherent physical limitations of various recording devices, im- An image in its digital form can be assumed to be an encoded form
ages become prone to manifestation of some random noise during image of a matrix with grey-level or colour pixel intensity values as its ele-
acquisition. Noise can be understood as a basic signal distortion which ments. Of course, this matrix has the third dimension of time in case of
hinders the process of image observation and information extraction. a video. In case of grey scale images it can be referred to as a signal in
Suppression of image noise forms elementary basis of image analysis two dimensions i.e. (x, u(x)), where u(x) is the pixel intensity value at
and processing, hence any progress made in image denoising domain location x. In colour image, value u(x) assumes a triplet of values for
helps strengthening our understanding of basic image statistics and pro- each of red, green and blue components. In order to maintain simplicity
cessing [1]. this review shall remain confine to grey-scale images except for brief
With massive increase in generation of digital image often captured mention of denoising in colour domain [1, 8]. The methodology opted
in poor atmospheric/illumination conditions, image restoration meth- for basic elementary grey-scale image denoising can be explicitly inter-
ods have become indispensible tools in present era of computer aided polated to colour scale images and 3-dimensional videos. In an orderly
analysis. From numerous types of noise prevailing in different kinds of fashion an image noise model can roughly be approximated as:
images Additive White Gaussian Noise (AWGN), impulse noise (salt and
𝑓 (𝑥) = 𝑢(𝑥) + 𝑛(𝑥) 𝑥 ⊂ 𝑋, 𝑋 ⊂ 𝑍 2
pepper), quantisation noise, Poisson noise and speckle noise are most
frequently discussed noises in literature [2]. AWGN primarily occurs in where u(x) denotes true signal value and n(x) denotes noise at location
analog circuitry during image acquisition and transmission. The preva- x. Noise can be additive or multiplicative in nature. Most often it is found
lence of other types of noises such as quantisation noise, impulse noise, to independent and identically distributed with zero mean and standard
speckle noise and Poisson noise occurs mainly due to faulty manufac- deviation 𝜎. In CCD (charged coupled devices) cameras noise may be
turing, bit error and inadequate photon count during image acquisition. present in electron circuitry (thermal noise), due to inadequate photon
There are a variety of digital images available which provide valu- count (photon noise) or it may be quantisation noise. However AWGN
able information in diverse fields of application like medical imaging, is mostly commonly encountered noise in real-time applications, hence
remote sensing, military and surveillance, robotics and artificial intelli- the AWGN model is the primary focus of this article.
gence. The contamination of these images irrevocably destroys image in- An image is an extremely interesting form of data which comprises
terpretability. Therefore image denoising methods find widespread use of several entities, for instance smooth homogenous regions, edges, con-
in the field of medical imaging, remote sensing, military and surveil- tours and spikes all resulting from gradually varying grey level val-
lance, biometrics and forensics, industrial and agricultural automation. ues of pixel intensity. The low intensity pixel values are important to

∗
Corresponding author.
E-mail address: [email protected] (B. Goyal).

https://doi.org/10.1016/j.inffus.2019.09.003
Received 5 March 2019; Received in revised form 11 August 2019; Accepted 6 September 2019
Available online 6 September 2019
1566-2535/© 2019 Elsevier B.V. All rights reserved.
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

determine the overall shape and formation of objects in an image, This article is structured as following:
whereas high intensity pixel values or high level variation in pixel in-
• Section 2 covers the various spatial domain filtering methods.
tensity values of image forms edges and features. The fine feature de-
• The transform domain methods and various thresholding schemes
tails and edges are significant for the higher level information analysis,
comprehension and extraction. The random manifestation of noise irre- are discussed in Section 3.
• The methods in other domains including statistical and probability
vocably corrupts this valuable pixel information [9,10].
In last two decades, the researchers continuously thrive to develop distribution based methods are discussed in Section 4.
• Section 5 discusses several hybrid domain methods.
efficient denoising algorithms which aim at restoring a reasonable esti-
• The sparse representation and dictionary learning methods are dis-
mate from distorted image while preserving fine features and edges. Due
to rising level of technical sophistication in image acquisition devices, cussed in Section 6.
• Sections 7 and 8 constitutes the recent trends and applications for
digital sensors tend to increase pixel resolution; however, with similar
aperture sizes these sensors become more susceptible to noise making image denoising respectively.
• Section 9 entails the experimental setup for various state-of-the-art
the role image denoising algorithms more important. Software-oriented
approaches are mostly device independent hence widely employed as methods evaluated.
• The article is concluded in Section 10.
compared to hardware and optical systems [3].
While exploring and exploiting the digital signal processing theory
and tools, researchers have experimented with various diverse tools of 2. Spatial domain filtering
mathematical sciences, in order to fabricate an algorithm to remove
noise while maintaining as many feature details as possible. There ex- Ideally, one thinks as filtering to be the foremost solution to image
ists a number of denoising methods originating from various disciplines noise removal i.e. suppression of the unwanted variation in pixel inten-
like spectral and multi-resolution analysis, partial differential equations, sity values. Indeed, filtering of signals is fundamental to basic image
probability theory and statistics etc. On the basis of the type of denois- processing and has long been in application for smoothing, sharpen-
ing algorithms, they can broadly be referred to as: spatial domain filter- ing, edge detection and contrast enhancement. In an algorithm in spa-
ing, transform domain thresholding, random fields, statistical models, tial domain, to denoise a single patch the correlation amongst the pixel
anisotropic diffusion methods, dictionary learning methods and hybrid candidates are exploited. In spatial domain, the set of operations is di-
methods [4]. Besides these other major denoising measures include spa- rectly employed on image matrix, whereas in case of transform domain
tial adaptive filters, statistical estimators of all sorts, stochastic analysis, filtering, image matrix is first mapped into corresponding transformed
morphological analysis and order statistics [5]. coefficients and then further thresholding is carried out.
The primitive and well-established spatial domain methods include According to manner of selection of pixel candidates used in filter-
local and non-local filters which exploit the similarities between pixels ing process, a filter can be categorised as a local filter and a non-local
or patches of an image. Both dictionary learning based and transform filter. The underlying principle of an image denoising algorithm is that
domain methods consider transforming images into frequency domain, noise is uncorrelated amongst pixels and true signal pixel intensities are
in which similarities and dissimilarities of transformed coefficients are correlated to each other [1,8,11].
exploited. The difference between them is that transform domain ap-
proaches apply fixed basis functions for the representation of images, 2.1. Local filters
whereas learning-based methods use sparse representations on a re-
dundant dictionary [6,7]. In this article, an attempt has been made to A filter is considered to be local if the filter support for denoising
abridge the hierarchical layout of the development and establishment of a pixel is restricted by spatial distance. A non-local filter exploits the
various denoising schemes in case of removal of Additive White Gaus- correlation amongst the entire range of pixels in an image. The popular
sian Noise from basic averaging of the pixels in the neighbourhood to local filters designed for noise reduction are Gaussian filter, Least Mean
more sophisticated hybrid techniques. The broad categorisation of var- Square filters, Bilateral filter, Weiner filter, SUSAN filter, Anisotropic
ious image denoising methods is given in the Fig. 1. Diffusion filter, Rank filter, Steering Kernel Regression (SKR), Metric
Steering Kernel Regression and Trained filters [6].
The simplest type of a Linear Translational Invariant filtering is av-
eraging or mean or box filtering which generates output at each pixel
as the average of neighbouring pixels in a given window. These linear
filters tend to convolve image matrix with filter mask to generate a lin-
ear expansion of neighbourhood values. This type of spatial filtering is
the most primitive and easiest form of noise removal, however they of-
ten generate undesirable amount of smoothing of edges, poor feature
localisation and loss of details [1,6,8]. Fig. 2 depicts the layout of basic
averaging filter.
Another types of linear filtering methods are Gaussian Filter (GF)
[12] and Weiner filter [13,14] which do not employ mean of neigh-
bouring pixels. Wiener filters are a class of optimum linear filters which
involve linear estimation of a desired signal sequence from another re-
lated sequence. GF are typical linear filters which fall under the cat-
egory of local filters and are isotropic in nature and have long being
applied in the image denoising. Gaussian filtering is of particular sig-
nificance in literature as the Fourier transform of Gaussian functions
are real and their shapes are easily specified. GF invariably assumes
that pixels in a neighbourhood are homogenous and have smooth spa-
tial variations. However, in GF filter weights attenuates with increasing
distance from centre and results in blurring of edges. Most of the lo-
cal denoising filters have primarily been the improvisations of Gaussian
Fig. 1. Categorization of image denoising methods. filtering which have been proposed since then to provide better edge

221
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Fig. 2. Basic methodology of averaging filter.

preserving ability [6]. These LTI filters though form an easy and time It is a typical non-linear feature-preserving denoising algorithm,
efficient solutions, tend to remove important structural information and which is a standalone technique to address the issue of imaging arte-
edge details which are vital to image analysis. Henceforth, various other facts. According to Dogra et al. [22], prevalence of artefacts which can
non-linear filters have been proposed in order to address their limita- disguise image interpretation especially in case of medical diagnosis can
tions in a manner of their improvisations and innovative ideas. The age be effectively removed with anisotropic diffusion filtering.
old median filtering [8,15], weighted median filter [16] and Rank fil- Since its introduction by Perona and Malik, an extensive amount
ter [17] are basic and most trivial examples of non-linear filters which of literature has existed presenting a number of PDE-based anisotropic
were formally developed and documented in literature. In median fil- models which offer diverse modifications in order to attain a steady
tering, the neighbouring pixels are ranked according to brightness (in- state solution and to address the issue of staircase effects [23–30]. The
tensity) and the median value becomes the new value for the central staircase effect presents a visually unpleasant artefactual display in im-
pixel. ages and often perplex as false edge. For instance, Qui et al. proposed
With an aim to preserve edges, image details, image geometries and a robust ridge detector, non-local derivative based feature preserving
to overcome Gaussian blurring, the anisotropic diffusion filtering (ADF) diffusion filtering process [31]. Besides this, in [32] a Gaussian filtering
was proposed, the idea of which goes back to Perona and Malik (PM) technique based on pre-denoising beforehand each iteration is proposed
[18]. The ADF interpolated Gaussian filtering, by convolving the image by Catte et al. as the modified PM model. The higher order partial differ-
u at x only in the direction which is orthogonal to its gradient Du(x). ential equations have also been explored by scientists and have proved
In a simpler manner, ADF is a process of solving the anisotropic heat to be efficient in minimizing the staircase effects. Some of the work
equation which is a second order partial differential equation. It is a have also been carried out using fourth order partial differential equa-
non-linear diffusion process, which was proposed in order to overcome tions (FPDE) and have produced satisfactory denoising results [33–35].
blurring and poor localisation issues posed by linear diffusion filtering Lu and Tan in 2013 gave a fourth-order partial differential equation
[19,20]. denoising model adopting a variety of conduction coefficients which re-
ADF is an edge preserving image denoising algorithm which applies sults in better preservation of edges and features as compared to conven-
an inhomogeneous process to lessen the diffusivity at locations of edges tional second-order PDE based models [36]. Another extremely pioneer-
and carries out a diffusion process in homogenous regions or regions ing work based on FDPE was given by You–Kaveh employing a piecewise
with slight intensity variations. The modified heat equation (AD) given harmonic function to approximate noise-free image from noisy image
by Perona and Malik takes the following form [21]: [37]. Though a lot of research has been carried out based on FPDE, they
𝜕𝑢 often tend to generate speckle noise in filtered images and over-smooth
= 𝑑𝑖𝑣(ℎ|∇𝑢(𝑥, 𝑦, 𝑠)|)(∇𝑢(𝑥, 𝑦, 𝑠)) the step edges as FPDE based models decay high frequency coefficients
𝜕𝑠
of the image much faster than PDE based models.
where 𝑢𝑠 = 𝑢(𝑥, 𝑦, 𝑠), is the image obtained at an instance s, div is the
A PDE algorithm based on combination of local gradient, grey level
divergence operator, ∇ is the gradient operator, and h(.) is the diffusion
variance and edge stopping function was given by Chao and Tsai in
coefficient or the edge-stopping function. The diffusion coefficient h(∇u)
[38] to derive better edge preserving capability. Gilboa et al. [39] in-
is a local image gradient.
troduced an adaptive diffusion process which employed forward and
The magnitude of amount of diffusion is inversely proportional to
backward diffusion process in order to introduce the aspect of better
gradient values or edges. Within homogenous or uniform regions, gra-
edge enhancement while suppressing noisy pixels in smoother sections.
dient value is small, hence h(∇u) approaches 1, hence removes or
It gave an interesting review of the hierarchy of ADE based methodol-
smoothen away the uniform regions whereas, in case of near edges or
ogy along with its extended application in the field of image denoising
boundaries, diffusion coefficient approaches zero and hence preserves
and also demonstrated regularised ADE model which was well posed
the edges [21]. The diffusion coefficient can take either of the two fol-
and has a steady-state non-trivial solution.
lowing forms:
In [40] an adaptive thresholding based ADE model has been pro-
( )2
− ∇𝑟𝐼 posed. A modified version of gradient vector flow was given by Ghita
ℎ (∇ 𝐼 ) = 𝑒 et al. [41] to further improve denoising performance. Another domain
of work, based on ADE filtering has gained a lot of importance which
1 employs directional Laplacian and causes edge-directional diffusivity in
ℎ (∇ 𝐼 ) = ( )2
1+ ∇𝐼 original image [4,42].
𝑟
An extensive amount of work has been documented in literature
h(.) is positive and non-increasing function, r is the contrast parameter which derives its inspiration from anisotropic diffusion filtering-Perona
or gradient modulus threshold which controls the amount of denoising. Malik models in one way or other, few of which have been briefed out in
The diffusion coefficient acts as a typical edge stopping function, which this section. They are non-linear set of filtering techniques which design
assumes zero value around edges or gradients. set levels of directions and gradients resulting in better edge preserving

222
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

ability as compared to other regularisation based filters. Despite the fact, information from two correlated images [64]. Another filter, which was
that a lot of improvisations of ADE-PM model have established their based on ideal interpretation of the bilateral filter, was guided filter
applicability they still tend to sharpen boundaries with loss of textural [65]. This filter implementation was faster than conventional bilateral
details [4]. filter and derives its edge stopping function from a guidance image.
In 1992, Rudin and Osher proposed the total variation minimisa- This filter had an automatic time implementation, independent of win-
tion method which is another regularisation filtering algorithm to fil- dow radius k and it freely chose arbitrary sizes of kernels in real time
ter out noise from homogenous regions but not its edges. For a given applications. It was also able to mitigate the issue of gradient reversal
noisy image v(x), authors proposed to recover original image u(x) as artefacts to some extent.
solution of a constrained minimisation idea such that, total variation Besides these filters, various other filters namely steering kernel regression
of image is minimised, in turn generating a constant piecewise solution (SKR), Metric-SKR, KSPR and Trained filters will be briefed in the later
[43]. The restored image is simply obtained by minimizing the energy section of this article.
variation between the original and restored images. Some algorithms This preceding section gives a brief insight into local spatial domain
namely Chambolle’s algorithm [44] and forward-backward algorithm filters and their gradual progression in the field of denoising has been
[45] have been given as solutions to Total Variation (TV) based minimi- traced out. Since basic idea for implementation of local filters is exploita-
sation problems. However this technique was able to preserve informa- tion of correlation amongst pixels, they tend to perform poorly at high
tion along straight edges but presented two major drawbacks: textural noise levels, as the correlation amongst pixels is severely corrupted by
information was not preserved and it introduced staircase effect. elevated noise levels.
Most of the spatial domain filters are derived on basis of spatial prox-
imity, whereas the neighbourhood filters considers grey level similar- 2.2. Non-local filters
ities to define a neighbourhood patch. The famous Yaroslavsky filter
considers both spatial distance and grey-level similarities for averaging In [2], Buades et al. besides classifying the then existing image de-
[46,47]. The Bilateral Filter (BF) was given by Tomasi and Munich as noising methods, based on experimental methodology introduced a fil-
a modified form of neighbourhood filter, which weighs the distance to ter, in order to exploit similarity amongst pixels in a non-local manner
reference pixel instead of following a fixed neighbourhood [48]. or in entire image, popularly known as Non-local mean (NLM) filter.
Similarly, another spatial filtering approach based on neighbour- This pioneering work clearly established that the self-similarity amongst
hood filter, for providing increased edge retention is known as Smallest characteristics of an image in a non-local manner is the biggest poten-
Univalue Segment Assimilating Nucleus (SUSAN) [49]. It constructs the tial basis in the field of image denoising. NLM filter exploits presence of
average of all pixels which are equidistant from central pixel. similar features or patterns in the image. In BF a weighted average of
BF considers both grey level similarity and spatial closeness in the pixels in a given patch is generated and weights depend not only geo-
denoising process. It filters out noise in the regions belonging to ref- metrical distances but also on contextual similarity. The method of non-
erence pixels and preserves boundaries and sharp edges of the image. local filtering can alternatively be understood as exploiting Euclidean
It has been established in [4,50], that BF is a form of discretisation of distance between two patches around reference and the selected pixel.
specific type of PDE-AD model. The pixels which are spatially close and NLM algorithm is typical point-wise denoising approach, which gener-
have similar pixel intensities are averaged in a manner with weights de- ates a noise-free pixel for a single point only and similarity between the
caying with the increasing distance from the reference pixel. Recently reference pixel neighbourhoods directly affects filter weights.
a lot of work is concentrated on providing improvisations of BF such After preposition of original NLM filter in [2], several improvisa-
as Weighted Bilateral filter, Robust Bilateral filter, Fast Bilateral filter, tions have been developed afterwards, in order to accelerate filter im-
Multi-Resolutional bilateral filter [50]. plementation and to improvise the qualitative and quantitative results.
It has been stated by Elad in [51] that bilateral filtering is similar to Zhang et al. introduced a two-directional non-local variation model for
a Jacobi iteration of a weighted least squares minimisation. The process image denoising which exploits an idea that if similar patches are ar-
of bilateral filtering was simple and non-iterative, but its direct imple- ranged as matrices then rows and columns exhibit similarities [66]. In
mentation was slow as BF has brute force implementation where compu- order to accelerate NLM, refined set of candidate patches was generated
tational complexity increases exponentially with increase in neighbour- by determining mean and gradient for pre-selection of the contributing
hood radius. In 2011, Chaudhary et al. gave a faster approach for BF by neighbourhoods [67]. In another work by Wang et al. in 2012, Gabor
employing trigonometric range kernels instead of considering Euclidean feature non-local algorithm was proposed in order to denoise textural
spatial distance [52]. In 2013, another approach for an accelerated BF images [68]. An improved NLM (INLM) has been given in [69], which
was given in [53]. Chaudhary et al. gave a sure and fast approach for considerably improves efficacy of original NLM by assuming symme-
image denoising using bilateral filtering in which Weighted Bilateral Fil- try in the weighting function and by computing the Euclidean distance
tering (WBF) is obtained by combining the original bilateral filter and via symmetrical movement of recursive average filter (bi-square robust
its modification in a weighted fashion with an aim to minimise the MSE function). The execution of the method in an iterative manner results
(Mean Square Error) [54]. Besides these several other accelerated forms in better grouping of patches, based on pre-stage processing. This mod-
of bilateral filtering have been given in [55-57]. ified bi-square based robust weight calculation results in reduced loss
According to Frabman and Durand, BF tends to create gradient re- of details. Though this filter removes extra noise, it tends to blur details
versal artefacts in high dynamic range detail decomposition and com- due to post-processing.
pression of images [55,58]. Also it can be clearly seen in [59] that BF Inspired by some modified versions of NLM, Xu et al. proposed a
does not work well at low noise levels. Though bilateral filtering is able patch-grouping based NLM (NLMPG) algorithm for denoising of remote
to preserve edges, but they tend to retain some of the unwanted noisy sensing images [70]. This algorithm mainly exploits Block-wise NLM
coefficients as well. Also when standard deviation of noise exceeds edge algorithm given by Coupe et al. [71] by exhibiting improvisation in two
contrast this method gives a poorly restored image [48]. aspects: it cuts out the redundancy by selecting most similar patches and
In [60], Jin et al. introduced an algorithm which employs a weight- customises filtering constant for centre patches in accordance with the
ing function to BF, where either centre pixel or vector median is chosen variance of image patches. This algorithm showed elevated performance
to be the part of filtering process. An empirical study based model for than some existing NLM improvisations and K-SVD (k means singular
bilateral filter was given as multi-resolution bilateral filter in [61,62]. value decomposition).
In a work given by Peng et al., an optimised approach for selection of Various other improvisations of NLM have been given in [72,73].
vector parameters was given based on minimisation of Stein’s unbiased Though NLM is the first filter which exploits self-similarity in the en-
risk estimate [63]. The popular joint bilateral filtering exploits global tire range of pixels, it introduces artefacts and performance degradation

223
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

when the image lacks in similar patches. The conceptual exploitation of could be represented by fewer number of non-zero coefficients or an
non-local image features to preserve edges and details while removing image could be represented as a linear expansion of few high valued
noise have remained the potential idea for most of the state-of-the-art coefficient. This property has made them an extremely applicable digital
image denoising algorithms which will be discussed in the later section signal processing tool both in 1-D and 2-D domain. The representational
of this article. attributes like localisation, isotropy, multi-resolution and orientation of
Besides this, there are many other mathematical concepts which basis functions at a variety of directions are major properties of image
have been widely employed in literature in context of spatial domain transform methods. There are a large number of variations available
image denoising. Morphological filtering is one such topic. Morphologi- in this category such as Fourier transform (FT), Fast Fourier Transform
cal operations deal with tools to extract image features that are useful in (FFT), discrete cosine transform (DCT), wavelets, curvelets, steerable
description and representation of a shape. The morphological filtering wavelets, ridgelets, ripplet, curvelet, contourlets, wedgelets, bandlets,
comprises of opening and closing operations for smoothing of images directionlets, shearlets and directional filter banks [1,6,8].
and removal of noise. The opening operation smoothes contours of an FT and DCT have been widely employed in image denoising and
object, breaks narrow dips and suppresses bright peaks, whereas closing loss less compression. Though image restoration process was simpler
generally fuses the narrow dips, eliminates small holes and diminishes with the help of DCT, the reconstructed images showed blocking arte-
darker details [74]. facts and failed drastically at preserving edge features and details of
However, a very recent filter based on morphological filtering has images. The major disadvantage associated with FT was its long basis
been developed by Graham Treece in order to adaptively remove the functions, whereas the multi-resolution tool- Wavelet transform exhibits
amount of noise present in images. The bitonic filter follows the prin- sparsity (few number of non-zero coefficients). The major reason which
ciple of bitonicity i.e. it preserves the signal content which is locally made wavelet more popular than FT is that wavelets provide localisa-
bitonic (contains only one maxima or minima in a given range) and re- tion both in space and time, whereas FT is localised in frequency only.
moves continuously altering noise pixels. The filter has shown promis- Also in comparison to DCT, Discrete wavelet transform (DWT) employs
ing results in low to high noise values. Bionic filter is a typical non-local higher number of optimal set of functions to represent sharp edges than
filter which constructs the weighted Gaussian output of opening and cosines. Besides this in case of high pass sub-bands, DCT is able to pro-
closing filter weights. It is an adaptive image denoising which preserves vide higher frequency resolution and lower spatial resolution. A consid-
edges while removing noise, without need of prior knowledge of amount erable amount of literature shows the superior performance of wavelet
of noise. This filter is shown to have better denoising performance than transform as compared FT and DCT, in terms of efficient image restora-
Gaussian, Median and Opening–Closing and Closing–Opening filters in tion and lossless compression [76].
case of AWGN and Impulse noise [75]. The advent of theory of multi-resolution signal decomposition with
The primary objective of any denoising algorithm is retention of fine wavelet transform by Mallat in 1989 brought a revolution in the field
feature details as much as possible, while denoising the image to a sig- of image processing and computer vision [77].
nificant level of clarity, representation and interpretation. Most of the A hierarchical framework for interpretation of image information
filters work on the principle of simply disabling the mechanism of noise is efficiently provided by multiresolution representation tools (MRT).
reduction, at the point of detection of edges. However, these filters tend At coarser resolutions, image can be viewed through larger structures
to follow similar criterion at fairly higher noise values, where intensity which mainly identify the context of the image. With increasing resolu-
variation in noise exceeds edge discontinuity. Therefore, though spa- tions, fine detailed structures are analysed. This particular property of
tial domain filters have achieved high levels of performance, there is MRT is useful in pattern recognition and computer vision [6].
always preservation of features at the cost of prevalence of noise. The DWT (Discrete Wavelet Transform) is a mathematical tool which is
noise is random and has been presumed to have higher frequency, and obtained by discretisation of continuous wavelet transform and is con-
true signal has lower frequency and is repetitive in nature, therefore stitutes a translation and dilation factor thereby encompassing entire
the higher frequency can easily be removed in order to achieve denois- image at different scales in a sliding window manner. The energy com-
ing. However the edges and features are often not repetitive and have paction property of wavelet transform i.e. most of the information in
higher frequency and their removal causes edge corruption and signal image is encoded in few high valued coefficients is one of the major
blurring. Therefore, these spatial domain filters continuously strive to reason for its widespread application [78,79].
remove high frequency noise and not high frequency signal, which re- DWT constructs an approximation level of the images at coarser lev-
mains the single most challenging problem in spatial domain filtering. els using low-pass filter banks and detailed level by employing high pass
Besides this, spatial domain filters are prone to gradient reversal and filter banks. Each level of approximation is further decomposed into im-
halo effects near the edges. Also due to large window size of filters and age representation at next scale
convolution process, spatial domain filters are slow and computational The DWT uses a number of mother wavelets which provide basis
ineffective. In some cases they cause staircase and rare-patch effect and function for signal decomposition. The well-known wavelet families are
it some cases they tend to cause over blurring of edges. Therefore there Biorthogonal, Coifflet, Daubechies, Discrete Meyer, Haar wavelet, re-
is a need to design a generalised framework which combines more than verse biorthogonal wavelet and Symlet [80–84]. The efficiency and suit-
one technique which not only solves the problem of image denoising ability of mother wavelets are governed by factors like length of filters,
but also is able to retain contours, edges, features and details. orthogonality, compact support and smoothness. Daubechies or DbN, is
Due to high computational complexity and inability in differentiat- most commonly employed mother wavelet function as they prove to be
ing between noise and sharp features in spatial domain, researchers have most suited for textural and feature analysis.
moved to transform domain techniques for image denoising. The denoising methods based on the wavelet usually transform the
The transform domain filters depend on various types of thresholds image content into multiple sub-bands at different resolution and scales.
in order to differentiate between edge and noise and have sustained Larger frequency coefficients contain the low frequency image informa-
paramount importance in the field of digital signal processing and multi- tion (approximation level) and noise and details exist in high frequency
resolution analysis. The next section is dedicated to various transform sub-bands. By thresholding smaller coefficients noise can be removed
based image denoising methods and their merits and demerits. and finally the image is restored by inverse transforming the coefficients
into spatial domain. This can further be understood by Fig. 3.
3. Transform domain filtering In the multi-resolution subspace, dissimilarity between noise and
true signal is accentuated on the observation that at low resolutions,
In contrast to approaches targeted in the spatial domain techniques, signal components i.e. the edges or significant geometrical structures
transform domain methods exploit the property of sparsity i.e. the signal over-exceeds noisy pixels [85]. The basic idea behind removal of noise

224
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Fig. 3. Image denoising with wavelet transform.

is that smaller wavelet coefficients in the high frequency band can be Table 1
thresholded and high magnitude coefficients in the lower frequency List of thresholding schemes.
band are preserved which are lesser likely to be noise. The performance S. no. Thresholding methods
of wavelet thresholding depends on two major criteria: one is the selec-
1 Universal Threshold
tion of the wavelet basis or mother wavelet i.e. Haar, Daubechies etc.
2 VISUShrink
and other is the choice of threshold at all sub-bands for all levels. 3 SURE-Shrink
4 BayesShrink
3.1. Thresholding 5 Prob-Shrink
6 SURELET
Rather bigger challenge in field of image denoising resides with the 7 Neigh Shrink Sure
9 Block Shrink
choice of threshold. For instance, a larger threshold value will shrink
the signal features, resulting in over-smoothing or blurring of signal,
whereas a smaller threshold value might leave higher number of coeffi-
cients associated with noise information. There are mainly two types of 3.1.1. Universal threshold
thresholding functions given by Donoho and Johnstone in 1995: hard To begin with, a universal threshold (UT) can be given as:
thresholding (HT) where each value is compared against threshold value √
𝜆 = 𝜎 2 log (𝐿)
and lower value is replaced by zero and other one is soft threshold-
ing (ST); where coefficients larger than threshold value are modified by where L is length of the signal, 𝜎 is the noise variance. In case of large
subtraction with the threshold value [86,87]. The HT leads to abrupt number of samples UT gives a fair estimate of the restored image [86].
changes which leads to the generation of artefacts whereas ST some-
times tends to over smoothen the restored image. As far as visual ap- 3.1.2. VISUShrink
pearance of the images is concerned, ST is better than HT. However it VISUShrink is a non-adaptive universal threshold given by Donoho
has been reported in literature than in some cases hard thresholding and Johnstone in 1995. The threshold value can be given as:
gives better results by preserving edge information [88]. √
ST often induces error and bias in the restored images. To address 𝑇𝑣𝑖𝑠𝑢 = 𝜎𝑛 2 log (𝑘)
this problem the literature has contained several improvisations for the where 𝜎 n 2 is the noise variance of the AWGN, and K corresponds to
design of optimal thresholding function. For instance, hyperbola func- range of the image. For larger values of K, Tvisu is also high, therefore it
tion [89], optimal linear interpolation shrink algorithm [90], were pro- generates a fairly smooth estimate. This type of threshold does not adapt
posed in order to overcome limitations posed by ST and HT. Besides well to the discontinuities in images however, it throughputs better per-
this, a thresholding similar to ST, semi-soft thresholding and Garrotte formance than universal threshold. In case of high probability where the
thresholding was proposed by Fodor and Kamath in 2001 [91]. These number of pixels reaches infinity it gives the best asymptotic MSE (Mean
thresholds have a higher degree of sensitivity towards small variations Squared Error) performance. However it often leads to over-smoothing
in signal and generated results with lesser mean squared error. With the of the images resulting in higher MSE [87].
selection of suitable threshold value, wavelet coefficients are modified
according to the shrinkage function. The restored image is obtained by
inverse transforming the thresholded coefficients. 3.1.3. SURE-Shrink
To analyse the progressive advancement in the field of design of SURE-Shrink was the very first attempt in terms of design of an adap-
threshold values, the existing methods are being categorised in three tive threshold where a different threshold is calculated for each of the
types [92,93]: detail sub-band. It is a hybridisation of Universal and SURE threshold
and depends on energy of the particular sub-band. It employs SURE
• Non-Adaptive Universal threshold: A unique threshold is selected for threshold in case of high energy coefficients and UT in case of sparse
all the sub-bands sub-bands and gives better performance than VISUShrink. The thresh-
• Sub-band Adaptive thresholding: A threshold is selected for all the old value at each resolution level is given by:
coefficients in the sub-band √
• Spatially Adaptive thresholding: Each wavelet coefficient has its own 𝜆 = min(𝑡, 𝜎 2 log (𝑘)
threshold.
where K refers to the number of wavelet coefficients in the specific de-
Under these categories the various types of thresholding methods tail sub-band, t is employed to minimise SURE and 𝜎 is the noise vari-
existing in the literature are listed in Table 1. ance. SURE-Shrink minimises MSE, delivers a better performance than

225
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

VISUShrink and is smoothness adaptive: i.e. it preserves the boundaries wavelet sub-bands same neighbouring window size was used. However,
and discontinuities in the images. NSS derived an optimal threshold value and SURE based neighbouring
According to the principle of wavelet energy compaction, most of the window size for each of the sub-band. The mathematical formula for
signal’s energy is concentrated in few wavelet coefficients and noise is NSS can be given as:
spread over all the coefficients and 𝜆 depends on the number of wavelet ( )
coefficients in the specific band. In case of extreme sparsity, the informa- 𝜆𝑠 𝐾 𝑠 = arg min 𝑆𝑈 𝑅𝐸 𝑊𝑠 , 𝜆, 𝐾
tion contributed by the coefficients where signal is zero, divulge the in- 𝜆s is the optimal threshold for sub-band S, Ks is the optimal window size
formation extended by the non-zero signal wavelet coefficients [94,95]. for sub-band S and Ws governs the minimisation of SURE [99].

3.1.4. BayesShrink
BayesShrink was given by Chang et al. and is a mathematical frame- 3.1.8. Block Shrink
In this approach an optimal size of block and threshold value is cal-
work which proposes a data driven adaptive threshold for generalised
Gaussian distribution in each of the detail sub-band. The threshold so culated for each wavelet sub-band in order to minimise SURE. This type
devised aims at minimizing the Bayesian risk. of thresholding preserves or eliminates the coefficients at the same time.
Hence it is easier to implement and gives better performance than SURE
According to its mathematical framework [88,92,93].The variance
Shrink and Neigh Shrink [100].
of a noisy image can be given as:
The thresholding schemes play a vital role in image denoising pro-
𝜎𝑦2 = 𝜎𝑥2 + 𝜎𝑛2 cess and each one has its own pros and cons. This is a brief account
of various thresholding methods existing in literature. By exploiting the
where 𝜎𝑦2 is the variance of the noisy image, 𝜎𝑥2 is the variance of the
interscale and intrascale correlations, these methods have been further
original signal and 𝜎𝑛2 is the noise variance. improvised. BLS-GSM is also one such example which is discussed in an
Then Bayesian Threshold is defined as [88]: elaborative way in the later section owing to its widespread acclamation
𝜎𝑛2 in image denoising. The transform domain methods have been widely
𝑇𝐵𝑎𝑦 = employed in image denoising, as they provide favourable assumptions
𝜎𝑥2
√ for filtering out noisy information from image features and details.
𝜎𝑥 = min(𝜎𝑦2 −𝜎𝑛2 , 0) The following properties of wavelets have made them an extremely
popular method for image transformation [6]:
𝐾
1 ∑ 2 • Sparsity: most of the information is contained in few high magnitude
𝜎𝑦2 = 𝐵
𝐾 𝑘=1 𝑘 coefficients and rest of coefficients are small.
• Multiresolution: Different sub-bands at different levels, contains var-
Bk denotes the wavelet coefficients at each scale and K is the total num- ied type of information. This facilitates different operations on sub-
ber of sub-band coefficients and 𝜎 n is computed with the help of Robust bands for increased precision.
Median Estimator.This type of threshold generates better visual results
than SURE-Shrink. In spite of the fact that wavelets have achieved remarkable success in
providing optimal approximation for one dimensional piecewise contin-
3.1.5. Prob-Shrink uous functions they are not very effective in optimally representing mul-
Given by Pižurica and Philips, Prob-Shrink is a probalistic shrinkage tivariate functions in an image that are typically governed by anisotropic
function which calculates the probability of the event: the given coef- features such as edges [101]. WT (Wavelet transform) is better to rep-
ficients contain signal features of interest. This method assumes a gen- resent 1-D signals but when it comes to 2-D signals where edges and
eralised Laplacian prior for the noise free data and computes the prod- contours are significant image features vital to comprehension and read-
uct of the probability and wavelet coefficient. In Prob-shrink framework ability of images, these transforms did not work well.
the signal of interest or signal wavelet coefficients exceeds the threshold The limited performance of the DWT is attributed to its lack of shift
which leads to the shrinking of the smaller coefficients and preservation invariance i.e. a small variation in input signal can develop significant
of larger coefficient values [94,96]. variations in distribution of energy amongst wavelet coefficients at dif-
ferent scales; poor directional sensitivity towards edges as the DWT coef-
3.1.6. SURELET ficients are separable and real. Furthermore, in case of DWT, if number
In a work in [97] an orthonormal wavelet approach was given to- of pixels in an image patch exceeds the number of coefficients in the
wards denoising. In this method, the denoising criterion was directly patch, ringing phenomena would occur.
parameterised as non-linear expansion of the elementary processes with To overcome these issues associated with DWT, popular un-
weights unknown. decimated discrete wavelet transform was proposed in [102] for de-
This technique aimed at minimizing MSE between the clean and noising. This method was able to alleviate issues posed by DWT to some
noisy signal. SURE (Stein’s Unbiased Risk Estimator) is a statistically extent, but it resulted in increased redundancy in throughputs and was
unbiased MSE estimator which only depends on the noisy image and computationally inefficient.
minimisation of which leads to solutions of linear set of equations. Luiser Further, keeping in mind the shift variance and reduced directional
et al. further improvised the work given in [97] by appending the prin- sensitivity of DWT, Kingsbury in 1998, proposed the Dual-Tree Com-
ciple of SURE. It was suggested that a typical denoising process can be plex Wavelet transform (DT-CWT) which provided shift-invariance and
expressed as linear combination of elementary denoising processes- Lin- directional sensitivity with the help of Gabor filters. Besides this it gave
ear Expansion of Thresholds (LET). This approach is popularly called as better reconstruction, limited redundancy and was computationally ef-
SURELET. This method has been applied for grey scale and colour im- ficient [103].
ages and for mixed Poisson–Gaussian denoising [98]. On similar lines an adaptive thresholding of wavelet packet was per-
formed in order to suppress the noise in medical images by Fathi et al. in
3.1.7. Neigh Shrink Sure (NSS) [90]. The technique showed an increased amount of performance at var-
Neigh Shrink (NS) is a data driven sub-band adaptive thresholding ied noise levels for medical images as compared to basic wavelet thresh-
in which a squaring window is centred on the noisy coefficient to be olding methods. Besides this, in adaptive directional lifting schemes, the
shrunk. This method was further improvised to introduce Neigh-Shrink ADL transform is applied in predicted directions besides horizontal and
SURE. NS employed a non-optimal universal threshold and for all the vertical directions [104].

226
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

In a rather interesting denoising approach an idea of selective radon transform [113] which resulted in wrap around effect or aliasing
wavelet shrinkage was employed. It determined that selection or re- and geometrical infidelity. The issue of wrap-around effect was effec-
jection of wavelet coefficients is better than probabilistic methods as tively removed in a work given in [114]. The author simply employed
former can statistically facilitate identification of a narrow interval for the finite radon transform with size equal to three. A multi-scale im-
estimated parameter which eventually adjusts wavelet coefficient with age denoising algorithm was developed in this work which effectively
a high level of certainty. This method chooses wavelet coefficients based preserved edges while removing noise. This method employed a mov-
on image features like their absolute value, regulatory across multi- ing window pyramid which yielded better approximation of lines and
resolution scales and spatial regularity. The few large magnitude coeffi- curves.
cients are grouped together. The first threshold chooses the coefficients Further, an adaptive digital ridgelet transform (ADR) was proposed
of large magnitude and second chooses a subset which exhibits spatially in 2016, as a multi-scale image decomposition algorithm for image de-
regular behaviour. This two-threshold method being non-iterative in na- noising. The improvised methodology resulted in better representation
ture provided a computationally simple solution facilitating real time of line and curve information by taking into consideration the under-
processing [248]. lying structures of image. The experimental results for this algorithm
The traditional orthogonal, maximally decimated wavelet based de- were able to depict PSNR improvement as compared to ridgelet and
noising methods exhibited visual artefacts like Gibbs phenomena due to curvelet transform [115]. An orthonormal finite ridgelet transform was
lack of shift invariance. Coifman gave the concept of “cycle spinning” also proposed in literature for image denoising [116]. Besides this,
in [105], to average out translational dependence thereby, addressing there were other image denoising algorithms documented in literature
visual artefacts. For a given set of shifts, once the data is shifted, it is de- which employed ridgelet transform in conjunction with other methods
noised and it is then unshifted. By carrying this for entire range of shifts in order to increase its efficacy. These methods fall into typical hybrid
and averaging the shifts, results so obtained exhibited much lower num- domain methods, which will be discussed in the later section of this
ber of visual artefacts. Another interesting improvisation in the wavelet article.
domain was introduction of steerable pyramids [106]. Contemporary to ridgelet, curvelet transform has also been perva-
The agenda of improvisation of wavelet performance was three-fold: sively studied in literature for image denoising. The curvelet and ridgelet
First to design adaptive thresholding methods in order to preserve more transforms were infact the major breakthroughs in the field of contour
number of high magnitude signal coefficients, second was to design di- based image representation.
rectionally sensitive image transforms which could boost denoising per- The prime motivation behind development of curvelet and ridgelet
formance and third was to hybrid wavelet transform with various meth- was to expand inherent limit plaguing the denoising of images with
ods in other domains in order to make optimal use of non-local image wavelet transform. This limit arises from the fact that along the sig-
features. nificant edges of image, wavelet exhibits large magnitude coefficients
The various threshold based methods have been discussed above. even at finer scales, such that within the map of large magnitude coef-
Besides this there are various other approaches in order to enhance the ficients, one may find edges repeated at each and every scale. This in-
performance of wavelet transform by employing data-adaptive and edge tuitively means that a large number of wavelet coefficients are required
preserving thresholding methods. in order to represent image edges in an efficient manner. This leads to
For instance, in a work given by Silva et al., image is initially di- higher computational complexity as there are so many wavelet coeffi-
vided into a set of blocks and then coefficients were obtained employing cients to process. Besides this the MSE achieved by wavelet based de-
wavelet transform. An edge strength based adaptive thresholding was noising methods was of the order €, noise parametric measure. However,
used to suppress noisy coefficients [107]. In 2013, Qiu et al. proposed the theoretical limits suggested it could be achieved uptil €4/3 .Therefore
explicit filter known as LLSURE. In this method a simple affine transform it was required to achieve this ideal MSE by developing new expansions
for input signal was considered for the filtered output in a local window. which could efficiently represent smooth functions as well as edges us-
Further, optimal transform coefficients are constructed by minimizing ing few non-zero coefficient values.
SURE. This method was motivated from SURELET and Guided Filter It is a general understanding that edges are more of ‘curved’ parame-
[108]. Further there were methods developed on statistical noise mod- ters rather than being straight! At very fine scales, curved edges appear
els, MAP and Markov random fields which will be discussed in another to be straight, so one should deploy ridgelet in a more localised manner
section of this article. in order to capture edges. Therefore, Curvelet transform, a multi-scale
In order to address the limitations posed by wavelet at higher dimen- variant of ridgelet transform was developed by Starck et al. based on
sions; a new representational system called Ridgelet was pioneered by isolation of varied scale using sub-band filtering [117]. The authors re-
Candes and Donoho in 1999, which would effectively represent the sin- ported that likewise ridgelets, curvelets occurred at all scales and loca-
gularities associated with lines in 2 dimensions [109,110]. The idea be- tions as well as orientations. Besides the properties of variable length
hind ridgelets was to enforce the wavelet with effective handling of point and width, curvelets possessed patchy anisotropy. The length and width
singularities in Radon domain. The implementation in radon transform were related by parabolic scaling and anisotropy increases at higher
domain facilitates mapping of the line singularity into point singularity scales. It could be witnessed from experimental analysis given in [117],
[111]. Initially, ridgelets were proposed for continuous space, however that thresholding of curvelet coefficients results in near-optimal solu-
the discrete implementations were a challenging problem, as straightfor- tions of smooth regions as well as discontinuities along curves.
ward implementation by enabling the discretisation of the continuous In work given in [118], curvelet coefficients were thresholded using
framework, would result in interpolation in spatial coordinates thereby soft-thresholding based on Bayesian theory in order to reconstruct iono-
leading to redundancy and imperfect reconstruction. grams from the noisy ones. In another piece of work, a new transform
Therefore, in 2003 [112], a finite discrete ridgelet transform (DRT) based on combination of wavelet, curvelet and ridgelet transform known
was proposed that was both invertible and redundant. The proposed VST (Variance stabilizing transform) was given for removal of Poisson
transform was found to have numerical stability and included adaptive noise. This transform is also popularly called extension of Anscombe
schemes along with directional and orthonormal bases for digital sig- transform [119].
nal in two dimensions. The properties of this transform were validated The primary limitation of Curvelet transform is that it is not built
in several applications including image denoising. The DRT had built- directly in the discrete domain and also is not able to provide a mul-
in linear geometrical structures, which facilitated direct thresholding tiresolution geometry. Also mathematical analysis and implementation
of the ridgelet coefficients in contrast to wavelet domain where adja- in discrete domain are more involved and less efficient as compared to
cent coefficients are initially chained together before thresholding them other directional transforms. Besides this, curvelets were lacking in or-
over the contours. In this algorithm as well, the basic building black was thogonality and weren’t critically sampled. Also thresholding in curvelet

227
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

domain results in visual artefacts in the manner of ghost effects along- are confined to resolve discontinuities along smooth curves. But these
side strong linear features. transforms are incapable to resolve discontinuities along edges and con-
While invigorating battle to design multi-scale transforms prevailed tours. While devising a transform one has to use a basis element with
in beginning of 20th century, some imaging transform like Brushlets much higher directional sensitivity, having more shapes and directions
[120], Wedgelets [121], Beamlets [122] benefited the field of image than the classical wavelets. In this relentless progressive evolution in the
processing in an immense manner by enabling the efficient singularity designing of new transform some of the major transform are bandlets
analysis. Further ahead, by considering the noteworthy chronological [133], haarlet [134], ranklet [135], morphlet [136], tetrolet [137] and
order, after curvelet transform, arose the era of contourlet transform. many more which could be found in [138]. Also to remove Poisson
While witnessing, the failed performance of wavelet transform for rep- noise, Luiser et al. proposed PURELET (Poisson unbiased risk estimate
resenting 2-D discontinuities, it was progressively thought that besides linear expansion of thresholds). It was devised to do two functions, one
precise localisation and multi-resolution analysis, new 2-dimensional is signal preservation and other is noise suppression in presence of shot
should be designed with elongated shapes oriented at different direc- noise. PURE denoising function uses unnormalised Haar wavelet trans-
tions and with varied aspect ratio. form [139]. Many methods have been reported in literature which target
The curvelet transform was able to achieve all this, but the discreti- both Gaussian and Poisson noise simultaneously.
sation of curvelet transform for critical sampling was rather challeng- In the process of continuous progression of LET’s (Linear expansion
ing. Therefore, the improvement of the curvelet transform over wavelet of thresholds) several image transforms were devised. The significant
to represent 2-D piecewise smooth signals with smooth and discontinu- transforms in the field of denoising have been discussed at some stretch
ous curves inspired Donoho and Vetterli, to design Contourlet transform in this article, while some are just mentioned.
(CT). The contourlet transform is a double iterated filter bank struc- Having transit through a gulf of transforms, there were a whole other
ture which solves the anisotropy scaling equations for curves resulting era of highly directionally sensitive transforms in literature, few of them
in a fast curvelet like implementation. The main idea behind construc- are Shearlet transform, Directionlet transform and Framelet transform.
tion of contourlet was to design a sparse representation, thereby pro- Since these transforms have exhibited exceptional performances in im-
viding smooth curves across discontinuities. In the basic understanding, age denoising in terms of PSNR, SSIM and visual quality, therefore they
edges are referred to as gathering of smooth boundaries along shapes of are being discussed in brief in the following section.
physical objects and discontinuities are sharp transitions in intensities. In the linear expansion of thresholds; multi-directionality (M-DIR)
The contourlet transform gave improved performance over wavelet and and Directional vanishing moments (DVM) (zero at w = o; sparse repre-
curvelet transform in terms of denoising. CT can be understood as de- sentation) are major properties which make them suitable for any image
ployment of wavelet-like transform for edge detection, and localisation processing task. The efficiency of 2-D wavelet transform is limited by its
of contour segments using directional transform. In double filter bank spatial isotropy and process of convolution and non-separable filter in
structure, Laplacian pyramid is used to acquire discontinuities and fur- both horizontal and vertical directions. Curvelets and contourlet pro-
ther directional filter banks are employed to expand basis images as vided anisotropy to some extent but they require oversampling during
contours or in other words to frame point discontinuities into linear implementation and are highly complex. Velisavljevic et al. [140] pro-
structures [123,124]. posed a multidirectional anisotropic transform based on integer lattices
According to Do and Vetterli [123], Contourlets are constructed with named it Directionlet. DT is an anisotropic, multi-scale geometric trans-
elongated basis function have lesser directional features and denoising form with directional vanishing moments along any two directions. Be-
with contourlet transform tends to introduce artefacts and Gibbs phe- sides being computationally efficient, it has perfect reconstruction and
nomena as they are translation invariant. Henceforth in a work given critically sampled basis functions. In a work given in [141] several
in [125], Eslami and Radha, employed cycle spinning based contourlet thresholding schemes have been exploited in Directionlet domain to ex-
transform to induce directional sensitivity in terms of image denoising. ploit intrascale and interscale dependencies.
This technique gave promising results for removal of AWGN as com- In order to substantiate the need of efficient edge representation
pared to contourlet transform over extended range of standard deviation while denoising of images, an increased sparsity based algorithm was
of noise. proposed in [142] using directionlet transform. For each pixel value
Due to sub-sampling at varied resolution levels, there is a shear lack multiple plausible estimates based on constructed DT representation
of translational invariance, which is vital to image denoising. Therefore of directions of edges were obtained. By computing weighted average
in another attempt by Eslami and Radha, translational-invariant con- of each of these estimates final denoising image was constructed and
tourlet transform was designed which employed modified versions of has shown to depict better performance that other multi-directional
directional filter banks and algorithm a trous for 1-D wavelet transform. wavelet based transforms [142]. Besides this a novel denoising algo-
Along with bivariate shrinkage thresholding their scheme presented an rithm based on Cauchy’s density function and direction-let transform
incredible denoising approach [126]. was designed to remove speckle noise from SAR (synthetic aperture
In a work given by Cunha et al., the concept of non-sub-sampled con- radar) images [143]. Another exemplary and rather recent work on re-
tourlet transform (NSCT) was floated which was fully multi-scale, shift moval of speckle noise from ultrasound images is given by Shahdoosti
invariant and had a fast directional expansion. It employed non-sub- can be found in [250] and is proposed in 2018. In this method general
sampled pyramid banks and filter banks which dispensed the need of speckle method is employed and log likehood function is maximised
2-D factorisation thereby eliminating Gibbs phenomena. The NSCT was in order to obtain noise free pixels. Further, distribution of each noisy
coupled with hard thresholding and local adaptive threshold to gener- pixel is weighted in accordance with statistical similarity between the
ate considerably higher PSNR than contourlet transform [127]. Some patch of noisy pixel and that of the pixel being processed. The sub-
other significant works based on contourlet transform for image denois- jective as well numerical observations depicted the superiority of algo-
ing could be found in [128–130]. rithm over state-of-the-art using spleen, pancreas and liver ultrasound
The study of generalisation of parabolic scaling law result in design- images.
ing of another interesting MRA (multiresolution analysis) tool i.e. Rip- In the era of multi-scale directional representation of images, a fairly
plet Type-1 transform. This transform is a higher order generalisation new representation scheme has been presented by Easley et al. and
of curvelet transform with the addition of two more parameter; sup- is called Shearlet. It is found to be very effective in providing opti-
port c and degree d. RT converges to curvelet transform when c = 1 and mally sparse representations for a variety of multidimensional data and
d = 2. Ripplet transform was shown to have exhibited promising results has showed improvements in variety of image processing applications
in image denoising at par with curvelet transform [131,132]. Ridgelet such as image enhancement, image fusion and image denoising. Shear-
resolves the discontinuities along lines where as Counterlet and Ripplet lets have been extensively used in context of image denoising [144]. It

228
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Table 2
Classification of image denoising methods.

Domain Spatial domain filtering Transform domain filtering Other domains Hybrid methods Sparse representation

Image Local Filters & Non-Local Fourier Transform [76], Bivariate Laplacian Bilateral Filtering+ Short K-Singular Value
denoising Filters: Wavelet transform [77], Distribution [159], Term Fourier Transform Decomposition [197],
methods Average filter, LTI Filters DWT [6], Dual Tree Maximum a prosteriori [172], Dual Domain Image K-LLD [201], LSSC
[1, 6, 8], Gaussian Filter complex wavelet transform [160], Markov Random Denoising [173], Progressive (learned simultaneous
[12], Weiner Filter [103], Ridgelet [115], Field [163], MSB-CRF Image Denoising [3], sparse coding) [202], CSR
[13,14], Rank Filter [17], Curvelet [117], Ripplet [166], Sylvester Lyapunov Perona-Malik-Isotropic (convolution sparse
Anisotropic Diffusion [131], Contourlet [123], Equation [168], RADL Diffusion [176], Trivariate representation) [4],
Filtering- PM-Model [18], NSCT [128–130], Shearlet, [169], Hidden Markov wavelet+ Spatial Domain M-SKR (metric steering
Yaroslavasky Filter [47], NSST [144], Directionlet Models [170] filtering [177], Non-local kernel regression) [203],
Bilateral Filter [48], [140], Framelet [150], mean +Partial Differential PLOW (patch Based
SUSAN filter [49], WBF, BLS-GSM (Bayesian Least equation [179], locally optimal wiener
RBF [54], Guided Filter Square Gaussian Scale BM3D+Brushlet [180], filter
[65], Joint bilateral Filter Mixtures [155], BM3D NLTV + BM3D [181],
[64], fast bilateral filter (Block Matching 3-D TV+NLM [182], DTCWT+
[55–57], Bitonic filter collaborative filtering) Ridgelet [192]
[75], Co-occurrence [217], [151], LPG-PCA (Local Pixel
NLM [6], INLM [69], Grouping-Principal
Morphological filters [74]. Component Analysis) [153],
Thresholding methods
Applications Natural Image denoising, medical Imaging, Remote sensing, Infrared Image denoising, Under water image denoising, SAR de-speckling

provides a more flexible tool and is more natural in implementation. The techniques. The hierarachy of various methods under specific domains
major underlying property of Shearlet is its shift invariance. It is band is summarised in Table 2.
limited and compactly supported and is basically a type of composite
wavelets having a ‘;directional parameter as well. It helps to provide 3.2. BLS-GSM (Bayesian Least Square Gaussian Scale Mixtures). It is a
the multi-resolutional and multi-scale representations due to the intro- well-established fact that there exist interscale and intrascale depen-
duction of shear parameter and anisotropic dilation. However Shearlet dencies amongst wavelet coefficients. It can be defined that if a (par-
based denoising algorithms should be optimised using adaptive thresh- ent) wavelet coefficient is small or large, its adjacent (child) coefficient
olding schemes which tend reinforce the image features and details ini- is also small or large and these coefficients tend to spread across scales
tially preserved by Shearlet transform [145]. Non-sub sampled Shearlet and hence can be modified. Since the wavelet coefficients were criti-
transform is a special type of discrete Shearlet transform. The major un- cally sampled and can lead to visually unpleasant artefacts there was
derlying property of NSST which has made it very popular for image need to develop overcomplete wavelets. This idea was exploited in BLS-
denoising is its invariance to shift of the input signal. GSM, where wavelet coefficient’s neighbourhoods at different positions
In context of denoising it is well known that shift variance intro- and scales are modelled as Gaussian Scale mixture (GSM). The noise
duces Gibbs phenomena around singularities i.e. artefacts which are free coefficients were computed using Bayesian least square estimator.
considerably reduced by NSST. The NSST is implemented using Non- In this algorithm, first the AWGN contaminated image is transformed
sub sampled Pyramid filter banks (NSP) and Non sub sampled shearing into wavelet coefficients. The each local neighbourhood is modelled as
filters (NSS) [146]. With an aim to obtain near optimal estimates and GSM and centre of each of the neighbourhood is estimated by employ-
to remove Gibbs type artefacts, shearlet transform was coupled with to- ing local Weiner estimator. In order to obtain the full optimal estimate
tal variation method for denoising of images. The shearlet coefficients of coefficients BLS solutions are computed. The denoised image can
were subjected to adaptive total variation minimisation which resulted be obtained by performing the inverse wavelet transform. This denois-
in outperforming related methods in curvelet domain form highly com- ing algorithm is by far the most proficient image denoising algorithm
plex images [147]. While witnessing the edge-preserving multi-scale and whose efficiency is matchless. However, since this method requires
multi-directional representation of the NSST, a technique based on NSST strict estimate of power spectrum density, this makes it non-adaptive
and twin support vector machines (T-SVM) along with sub-band adap- [152].
tive thresholding was proposed in [4]. This work besides outperform- Some of other methods based on exploitation of interscale and in-
ing the techniques like probshrink and BLS-GSM (Bayesian Least Square trascale dependencies are as follows: Bi-shrink is a wavelet shrinkage
Gaussian Scale Mixtures), presents a very brief, concise and informative technique designed by Sendur et al. using MAP (Maximum a Posteriori)
literature survey on image denoising. The technique was typically able and is based on a non-Gaussian bivariate distribution for modelling in-
to preserve edges while suppressing the noise. Some of other effective terscale dependencies [154]. In another work intrascale variability of
image denoising algorithms exploiting the edge preserving properties the wavelet coefficients is taken into account [155]. A method based on
of NSST are documented in [148,149]. After extensive literature sur- Gaussian Mixture Model (GMM) and Generalised Gaussian Distribution
vey, it is found that NSST is a prospective image denoising algorithm (GGD) for wavelet coefficients was given in [156].
when coupled with effective thresholding schemes and various spatial
domain filters. 3.3. BM3D (Block Matching 3-D collaborative filtering). Motivated by the
A major pioneering work in the field of image denoising includes idea of non-local grouping in NLM filter and redundant or overcomplete
adaptive directional lifting schemes for WT. This method is proposed representation in BLS-GSM, Dabov et al. designed an effective filtering
as translational invariant directional framelet transform which provides in 3D transform domain in 2007. It was designed by combining sliding
substantially superior redundant representation along with novel orien- window process with block matching. BM3D can be denoted as mul-
tation estimation based on Gabor filters [150]. tipoint approach which comprises of three steps: The single noisy im-
Besides these above mentioned transform domain techniques, the in- age is processed in a sliding manner, where blocks that exhibit similar-
terscale and intrascale correlations of the wavelet coefficients have been ity (minimum Euclidean distance) with currently processed block are
exploited in BM3D [151], BLS-GSM [152] and LPG-PCA [153]. These searched and grouped together. The resulting matched blocks are piled
three techniques infact beg the medal for the state-of-the-art denoising together to form a 3D array. The 3D groups so computed are redun-

229
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

dant and hence facilitates more effective representation than WT. The transforms and their combination with adaptive thresholding meth-
correlation amongst the blocks is exploited by applying a 3D decorre- ods, the non-local pixel similarity clustering like BM3D, Local pixel
lating unitary transform which results in a sparse representation of the grouping, BLS-GSM and NSST and so on. Various conventional proper-
original signal in 3D transform domain. After the efficient and sparse ties like multiresolution, sparsity, edge detection and anisotropy made
representation of noisy image, noise attenuation is done by applying transform domain methods an extremely popular choice for image
hard thresholding or Weiner filtering on transform coefficients. BM3D denoising.
comprises of these two steps in a sequential manner except for the fact However transform methods also had some drawbacks. The wavelet
that hard thresholding is carried out in the first step and Weiner filtering transform was ineffective in representing the smooth transitions and
is carried out in the second step. The local estimation of matched blocks DCT could not represent sharp transitions or singularities. The charac-
constructed by inverse 3D transform, results in improved denoising per- terisation of natural images with different patterns became difficult as
formance and effective detail preservation. The grouping provides ex- most of the transforms had fixed basis functions. The core idea of the
tremely reliable statistical data and Weiner filtering of 3D arrays is quite wavelet denoising stemed from the fact, that most of the signal energy is
effective. compacted in few wavelet coefficients while the noisy pixels are spread
The similarity amongst overlapping patches and correlation between throughout the wavelet coefficients. However the efficiency of these
wavelet coefficients tends to deliver high performance with BM3D. How- methods is strongly limited by the estimation of appropriate threshold.
ever, in case of presence of unique patches, BM3D ceases to provide Besides this, some methods resulted in ringing artefacts. Methods like
optimal solution. At low noise levels, BM3D has been established to BLS-GSM overcome these issues by exploiting redundancy in local neigh-
provide optimal denoising solutions; however, at standard deviation bourhoods. BM3D accomplishes better denoising results by exploring
above 40, performance of this algorithm drops as correlation amongst redundancy in the non-local manner. However when the assumption of
blocks is highly disturbed by noisy pixels, even though pre-filtering is self-similarity amongst the image pixels is not guaranteed well, the per-
carried out. Also sometimes instead of wavelet, in order to enhance per- formance of these methods is affected by occurrence of image artefacts
formance at standard deviation 40, DCT transform is employed which and residual noise.
have fixed basis functions but results in periodic artefacts. However in
an article [158], it has been commented that this replacement is unnec- 4. Methods in other domains
essary and performance improvement can be established by adjusting
some numerical parameters and same has been shown in that work. The Other than spatial and transform domain methods, researchers have
BM3D approach was primarily designed taking into consideration the designed image denoising methods based on statistical models of multi-
contamination of the images by Additive White Gaussian Noise [151]. scale coefficients and random fields. In statistical model schemes, the
To further increase sparsity of true signal in 3D domain Dabov et al. pro- noiseless coefficients are estimated from prior information based on
posed a generalisation of BM3D filter by making use of mutually similar Bayesian estimator. These methods aim at constructing spatially ho-
shape adaptive neighbourhoods. These adaptive shape neighbourhoods mogenous models adaptive to varying levels of signals with help of pa-
display local adaptivity to image details and features so that the true rameterisation of the variables which are self-random in nature. For in-
signal becomes mostly homogenous. Further ahead in 2009 to enhance stance, Yin et al. proposed an algorithm for removal of noise with the
adaptive capacity of pre-existing shape adaptive BM3D Dabov et al. in- help anisotropic bivariate Laplacian distribution function based mod-
troduced PCA (principal component analysis) as a part of the 3D domain. elling of wavelet coefficients [159].
The overall 3D transform is a separable composition of PCA and a fixed In another interesting set of works, Rabbani proposed a noise sup-
orthogonal 1D transform in third dimension. The technique can be fur- pression method which constituted of thresholding of coefficients using
ther improvised by addressing novel shrinkage criteria which are highly steerable pyramids and Laplacian probability distribution function. This
adaptive to the utilised transform [157,158]. method was further improvised in [160]. In [161], an approach towards
denoising of images is given by maximizing the MAP (maximum a pros-
3.4. LPG-PCA (Local Pixel Grouping-Principal Component Analysis). Us- teriori) measure with help of Gaussian Scale mixtures. Besides this, Fathi
ing classic idea of exploitation of non-local features and sparsity, Zhang and Nilchi developed an adaptive wavelet coefficient thresholding algo-
et al. proposed LPG-PCA for image denoising in order to mitigate the rithm which was statistically optimum. Despite the widespread acclaim
issue with fixed basis function in DCT transform. The PCA introduced of the statistical methods, they are observed to introduce ringing arte-
the concept of adaptive basis functions. facts in the images [162].
In this method, a two-step approach is followed where each of The methods based on random fields is another very popular do-
noisy pixel and its typical neighbourhood is grouped together into a main which is readily exploited in the field of image denoising as well
vector variable. These vectors so obtained are transformed into well- as other low level processing such as image classification and segmen-
known PCA domain or KL transform using eigen values and vectors. tation. In these methods, the value of intensity of a pixel depends on
The shrinkage is carried out in PCA domain. The same process is fol- neighbouring pixels. These methods are based on an observation that a
lowed in the second step. The local fine grain-like edges are prone to global representation of an image can be obtained from its local physical
have incorrect representation in BM3D due to fixed basis function. LPG- structures which is accomplished with the help of a conditional prob-
PCA is able to overcome this issue with help of local adaptive basis ability distribution function popularly called as Markov random Field
functions. However, input to second stage is filtered patches from the (MRF). Further under a Bayesian framework, a maximum a priori (MAP)
first stage. The second stage is exact replica of first stage except for probability is defined which states that a particular clean and sharp im-
different noise levels. This factor due to persistence of errors at first age could have been made noisy in order to obtain a specific degraded
stage degrades the denoising performance at the second stage as well image. These MRF models establish relationship between the neighbour-
[153]. ing pixels and between observed and estimated values at each pixel
In the field of transform domain image denoising the primitive [163].
efforts were centred on thresholding of orthogonal wavelet coeffi- In [164], a three layered hierarchical MRF model has been proposed
cients of noisy pixels. Then further the idea of translational invari- for image denoising where each of three layers depict the basic textures
ant schemes came into picture with the use of undecimated wavelet and smooth regions, the noisy image and observed noisy data respec-
transform. More recently several experiments have been carried out tively. In a work given by Chen et al., another MAP estimation based
with tree-based methods and modifications of various thresholding denoising solution has been employed using MRF’s for SAR despeckling
schemes like Bayesian thresholding, level-dependent thresholding and [165]. Zhong and Wang exploited both spectral and spatial dependen-
so on. Then focus substantially shifted towards highly directional cies and proposed denoising algorithms in the multiple-spectral-band-

230
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

conditional random fields (MSB-CRF) for hyperspectral images [166]. coming limitations of each of them. For instance, various spatial domain
In an interesting work given by Ho and Hwang, separate Bayesian net- filtering methods are able to preserve the edges but they tend to over-
works were obtained for different input images with the help of a single blur low contrast details in the homogenous regions. On the other hand,
Bayesian network estimation of wavelet coefficients [167]. transform domain methods are able to represent the textures and low
In another efficient denoising algorithm, Bayesian and MAP criterion contrast information but they tend to introduce ringing artefacts around
were put in use in order to induce preservation of edges while removing the edges.
AWGN and multiplicative noise having Rayleigh and Poisson distribu- Based on this approach Knaus and Zwicker combined two popular
tion respectively from CT images [168]. The significant contribution of filtering techniques namely bilateral filtering from spatial domain and
this work was development of a denoising algorithm using Sylvester Short Term Fourier transform (STFT) with wavelet shrinkage from trans-
Lyapunov equation in Bayesian Framework. In a work given by Wang form domain. With the help of this hybrid approach authors were able
et al. in [169], a robust adaptive directional lifting (RADL) algorithm to preserve low contrast details along with preservation of edges. This
was performed at each pixel level by estimating the noise idea of pixel method is popularly called Dual-Domain Image denoising (DDID) and
pattern classifier. was given in 2013 [172]. This work was further extended as Dual Do-
While understanding the concept of inter-scale and intra-scale de- main Filtering (DDF) by employing wavelet shrinkage and bilateral fil-
pendencies of the wavelet coefficients, Romberg et al. [170], established tering as robust noise estimators in two different domains. The extension
that where MRF models are more efficient at capturing intra-scale corre- of DDID was employed by introducing guided filtering using a second
lations, Hidden Markov models (HMM) work well for inter-scale correla- guiding image. This method outperformed the various others competi-
tions amongst edges and features. Their method employed the attributes tive methods such BM3D, NLM and PLOW with much lower computa-
of HMM to capture the higher order statistics, coefficient correlation tional complexity [173].
at same scale and local structures. In order to acquire the interaction PID (Progressive Image Denoising) is rather a quite recent and widely
amongst coefficients across scales the model was constructed using Hid- acclaimed work of Knaus and Zwicker, in which the connection between
den Markov Trees. Further the MAP estimation maximisation technique wavelet shrinkage and robust estimation of differential noise was ex-
is employed in order to restore the denoised image. The method has been ploited. However this method required higher number of iterations via
shown to work except for higher computational and training burden. In guided filtering as compared to its prior version named DDF. Contrary
order to address this issue, Malfait proposed a GSM model for neigh- to DDID it does not require to differentiate between noisy image and
bourhood estimation of wavelet coefficients with the help of a random guided image. This method extends the vision of rejecting bias by intro-
scalar multiplier [171]. ducing signals as outliers in the frequency domain employing, wavelet
Exploiting a model based on intrascale and interscale dependencies shrinkage as robust noise estimation. And finally these methods em-
of coefficients a method has been proposed in [252], for denoising for ploy DA (Deterministic annealing) and SA (Stimulated Annealing). This
grey scale images. It has already been explained in concept of Gaus- method unifies the spatial domain and frequency domain resulting in a
sian scale mixtures that image can be estimated with the help of parent much simpler but effective denoising algorithm competitive with most
coefficients. The method comprises of two stages in which first stage state-of-the-art techniques in case of synthetic images [3]. Further a
comprises of pre-filtering with Kalman filter [KL] and second stage con- much faster version of DDID (approximately three times) was designed
stitutes denoising of noisy wavelet coefficients with KF only. In fact by Pierazzo et al. [174]. This method was shown to mitigate the arte-
employment of KF facilitates to estimate coefficients of high-frequency facts created by DDID in the frequency domain.
subbands from coefficients of coarser scales and noisy observations of A progressive dual-domain filter was proposed in order to enhance
neighbouring coefficients. and denoise low quality optical remote sensing images. Initially high
For the first time, Hidden Markov tree with mixture of one-sided contrast images, obtained with the help of bilateral filter were enhanced
exponential density has been exploited to remove noise in images in a using histogram modification method. Further the low contrast image
work given in [253]. Further this step was combined with Dual con- approximation structures were captured using STFT and enhanced by
tourlet transform which induced shift invariance and mitigated Gibbs employing a corrective parameter which was adaptive [175]. These con-
phenomena. tinuous variations in the dual domain image denoising strongly witness
The methods based on MRF field have achieved considerable success the influence of hybrid domain in image denoising.
in image denoising however it has been found that, in real time applica- The isotropic diffusion (ID) and PM (Perona–Malik Diffusion) has
tions, the energy of the random fields should be computationally feasible been widely employed independent methods in denoising and have
in lesser time and it is always extremely exigent to derive near optimal shown to give good performance. However ID causes edge blurring,
energy functions which could provide desired solutions [4]. whereas PM introduces staircase effect leading to loss of the local details
in image. Therefore in [176], the authors proposed a hybrid image de-
5. Hybrid denoising techniques noising model by improving the ID and PM models by employing second
order directional derivative (directional Laplacian). The integrated ap-
The field of denoising has witnessed the design of large number of proach of modified PM and ID model resulted in suppression of noise in
efficient and widely applicable denoising algorithms. Each of them has homogenous regions along with preservation of fine details and textures
its own share contributions, applications and limitations. This abundant besides mitigating the stair case effects.
amount of denoising algorithms has exploited various domains of digital In an extremely interesting work, a recent method is given by Shah-
signal processing and relevant theory. Initially various types of spatial doosti and Hazavei in which compressed sensing, block-matching and
filtering techniques with exploitation of local and non-local similari- sparse representation based learned dictionaries are coupled together
ties were designed. Besides this, various wavelet based adaptive shrink- in bandlet domain. In this method first noisy image is transformed in
age methods were employed. Other methods included directionally bandlet domain and noise coefficients are threshold using hard thresh-
sensitive transforms and statistical methods exploiting the inter-scale olding. This method assumes that image is sparsely represented over a
and intra-scale dependencies. Several algorithms have been deployed dictionary rather than basis elements. Also block matching technique
based on sparse representation, random fields and dictionary learn- is employed to represent dictionary elements such that identical sparse
ing. These methods have shown to be progressively improved in one vectors are used for all patches in a group. The given method was tested
manner or the other in terms of quantitative measurements and visual on eight well known grey scale images and it outperformed several state-
quality. of-the-art denoising techniques [251].
However the current state-of-the-art techniques combine more than For removal of Gaussian Noise from images, Yu et al. proposed an
one domain in order to exploit advantages of multi-domain while over- algorithm by employing a trivariate wavelet shrinkage in the spatial do-

231
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

main bilateral filtering. This method was motivated from the fact that representations. In this category of image denoising methods, [189] is
wavelet based methods are prone to low frequency noise and ringing another work given by Durand and Forment, in which after the decom-
artefacts and spatial domain methods are computationally expensive. position of the noisy image into the sub-bands, the total variation min-
The wavelet coefficients were modelled as trivariate Gaussian shrink- imisation is employed in order to threshold the noisy coefficients.
age based on maximum a priori estimator. The overall method resulted In a rather recent piece of work, in order to address the issue of
in cost cutting and joint bilateral filtering by using wavelet transform Gibbs-like artefacts while improving PSNR, Shahdoosti et al. has pro-
[177]. posed an image denoising algorithm based on ripplet formulation of the
In another work given in [178], learning based algorithm is embed- total variation minimisation [190]. In their work smooth and textured
ded with curvelet transform tiling. The optimal tiling for best represen- regions were obtained using twin support vector machines. Instead of
tation of a given class of images was decided on the basis of maximiz- minimisation of variation of noisy images, this method tried to min-
ing PSNR and minimizing MSE by variation of parameters like angular imise variation in the image obtained from non-textured regions of the
and scale decompositions. The method was shown to have robust per- ripplet sub-bands. Besides this another work proposes to combine the
formance as compared to curvelet transform and NLM filter. Another non-linear anisotropic diffusion with curvelet shrinkage [191]. In this
method under hybrid domain methods is a combination of PDE and gen- method of post shrinkage via ripplet transform, the reconstructed im-
eralised cross validation (GCV) theory in tetrolet transform. The image age was processed with total variation diffusion by which only high
was decomposed using tetrolet transform and an adaptive threshold was frequency insignificant coefficients were suppressed with the help of a
constructed using the GCV theory. The method has shown to give out- constrained projection. By combining shift invariance of DTCWT and
standing performance in terms of PSNR values [179]. directional sensitivity of ridgelet transform a new hybrid approach was
The well-known BM3D algorithm for denoising is no less example presented in order to suppress the AWGN [192].
of hybridisation of spatial and transform domain filtering. Inspired by In 2013 Shreyamsha proposed a denoising scheme based on Gaus-
this idea of clustering 3D groups of similar patches, brushlet threshold- sian/Bilateral filtering and noise thresholding. In this scheme the idea
ing was employed to remove the noise in the frequency domain. This of method noise has been presented. The method noise here refers to
method depicted improved visual quality of the ultrasound images as the difference between original image and image denoised by certain
compared to BM3D alone [180]. algorithm. The difference in itself is noise and is called method noise. In
An intriguing approach of integrating non-local total variation GBFMT (Gaussian/Bilateral Filter Method Noise Thresholding), a noisy
(NLTV) and BM3D is given in [181] for image denoising. The method ob- image is assumed to be corrupted with Gaussian white additive noise
tained a pre-denoised image with the help of BM3D. Further while con- with zero mean and a known variance which can be varied for testing at
structing the fidelity term of NLTV model, the pre-processed image was high and low noise values. The corrupted image is filtered using bilateral
put to use for formulating energy functions and weighing functions in filtering and the residual image so obtained is thresholded in the wavelet
the regularisation term. The energy compaction equations were solved domain using hard thresholding [193]. In a similar work bilateral filter
using split Bregman algorithm. The proposed method demonstrated ex- was replaced by non-local filter to exploit the concept of method noise
ceptionally high visual quality with removal of artefacts and outper- thresholding (NLMNT) [194]. Since then many different combinations
formed NLTV and BM3D. have been tried in context of method noise thresholding [195]. Similarly
It has been discussed earlier that TV models are known to have given in [196] the concept of denoising has been extended to hybridisation of
good image smoothing properties; however they tend to propel staircase Weiner filter method and Shear let transform. The multi-scale and multi-
effect, by overly smoothing the flat areas. Also NLM filter, shows good resolution representations have led to the improvement in the PSNR and
performance in reconstructing low frequency smooth regions but fail at MSE values. Besides this a lot many hybridisation image denoising tech-
edges as there are not enough similar patches which leads prevalence of niques have been employed using sparse representation and dictionary
noise around edges (rare patch effect). This method employs a non-local learning methods. In order to abridge higher level comprehension of
data fidelity term instead of non-local regularisation term. The proposed these methods, some of the hybrid methods are briefed under sparse
method was able to perform efficient denoising free form staircase and representation based methods and some are listed under recent trends
rare patch effect [182]. in image denoising.
Further ahead, in literature, combining wavelet or curvelet domain Through extensive literature survey it has been observed that proper-
with TV and diffusion methods was an extremely common approach ties like multi-resolution, sparsity and edge detection have made trans-
to address the issue of Gibbs phenomena. In a study given in [183], form methods very popular methods for denoising. However, these
authors established that replacing these representations with Shear let methods cannot represent smooth transitions very well and tend to lose
transform is much better approach when combined with total variation on intricate image details and features. On the other hand, edge pre-
minimisation for removal of artefacts and noise with preservation of serving spatial domain filters tends to retain noise features and arte-
edges. This method is one other exemplary instance of hybrid domain facts. To overcome such issues various improvisations are being made
methods. In one other method, local Weiner filtering with the help of el- by exploiting the properties of both spatial and transform domain filters.
liptic directional windows is employed in differently oriented sub-bands These combinational methods with intuitive adaptive thresholding hold
of wavelet representation [184]. A hybrid denoising method based on a lot of scope in context of efficient image denoising in terms of both
replacing of 2-D separable wavelet transform with Bandlet transform in objective and subjective evaluation.
NSCT domain is given in [185]. The proposed method was employed to
remove Gaussian and Poisson noise [185]. Another space-transform do- 6. Sparse representation and dictionary learning methods
main image denoising includes combining of non-local mean algorithm
with hard thresholding in the wavelet domain. The resultant images are The redundant and sparse representations have long been the ba-
processed with guided filter in order to get rid of the ringing artefacts sic dynamic foundation for researchers in image denoising. One major
introduced at the first stage [186]. Besides this a band-specific shearlet reason for requirement of redundancy arises from its ability to induce
based image denoising technique is given in [187] which employ dif- shift invariance. The concepts of sparsity and redundancy have been
ferent thresholding techniques in different sub-bands containing varied widely made applicable by use of various multi-scale and directional
amount of noise. In a significant piece of work, Candes et al. proposed transforms like curvelets, ridgelets, brushlets, and directionlet. On the
to combine the total variation minimisation model with curvelet and very same approach of introducing sparsity and redundancy in the im-
ridgelet expansions in which sub-bands coefficients are approximately age representation methods, the sparse representation based dictionary
threshold [188]. A lot of authors have rather employed the total varia- learning methods were designed. This approach varied in the manner
tion minimisation schemes in combination with other transform domain that it possesses an example based restoration. In this category, the

232
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

common successful approach is BLS-GSM (Bayes Least Square Gaussian late weights of filter by employing offline training using abundance of
Scale Mixture) which formulates as image denoising problems as inverse images. These methods ensure an adaptive image restoration process as
problems using Bayesian minimisation and hence need of an image-prior fixed kernel coefficients are replaced by trained coefficients. For each
based on entropy, smoothness or sparsity in some transform domain be- noisy patch, same classification is employed and filtered coefficients are
comes indispensible. While these methods tend to derive their image obtained from the look up table (LUT) constructed in previous train-
estimate from a mathematical expression representative of the image ing process. These methods are computationally fast as LUT has to be
prior, these example based methods tend to derive their image prior trained only once offline. Some of the major examples of these methods
from pre-existing images. While incorporating the idea of prior learn- are M-SKR (metric- steering kernel regression) [203] and K-SPR (semi-
ing in sparse and redundant representations, most important thing is parametric regularisation) [204].
dictionary which provides a learned set of parameters. These methods In image denoising approaches the major issues lies in extraction of
eliminate need for fixed basis functions as is the case in curvelet or con- appropriate features capable of determining the relation between ad-
tourlets transforms [197]. jacent coefficients. Most of denoising algorithms are not able to effec-
The introduction of orthogonal matching pursuit (OMP) laid the tively suffice this purpose. DNN is class of machine learning algorithms
foundation of commencement of image denoising problem as a direct which proposes to learn flexible models devoid of any pre-requisite as-
sparse decomposition technique with redundant dictionaries [198–200]. sumptions and data distribution. In this class of image denoising algo-
In this category of methods quality of denoised image largely depends rithms Shahdoosti and Remza proposed a novel image denoising algo-
on type of dictionary which can either be learned from patches of noise- rithm based on CNN model and shearlet transform which can perform
free images or adaptive learning of noisy image itself. much better in terms feature classification then various existing state-of-
In other words it can be stated that these methods adapt a dictionary. the-art algorithms for denoising [244]. Support vector machines (SVMs)
These methods are usually employed as patch based denoising methods is an another important dictionary learning algorithm is introduced by
and averaging is carried out in case of overlapping patches. The dictio- Vapnik [245] in 1998 and have been extensively applied in function
nary constructed from learned examples contains atoms which can be estimation and classification problems. They are developed in context
understood as approximating the patches using a sparse linear combi- of statistical learning theory and aimed at structural risk minimisation.
nation of atoms. Least square support vector machine is a computationally effective and
In this section we review only some of the major dictionary learning faster version of SVM owing to obtaining of solutions to linear equations
methods which are K-Singular Value decomposition (K-SVD), Expected instead of quadratic programming.
Patch Log Likelihood (EPLL), K-Locally Learned Dictionaries (K-LLD), In a work proposed by Wang et al., in 2010 LS-SVM were combined
Learned Simultaneous Sparse Coding (LSSC), Patch based Locally Opti- with undecimated wavelet transform in order to denoise images while
mal Wiener (PLOW), Trained Filters, Metric Steering Kernel Regression preserving the edges [246]. Then again in 2013, he employed LS-SVM to
(SKR) filters, Convolution Sparse Representation (CSR), Locally adap- classify NSCT coefficients into noisy coefficients and edge coefficients.
tive Regression Kernels (LARKs) and various others [4]. With the help of LS-SVM, NSCT coefficients at different scales are clas-
Elad et al. with an aim to embed local over-complete dictionaries sified into smooth regions and areas containing texture regions. For the
in Bayesian global estimator proposed K-SVD as an iterative denoising maximisation of texture information, edge coefficients are left unmodi-
method which adaptively learns from the entire range of pixels of noisy fied, whereas noisy coefficients are threshold using Bayesian threshold.
image. Each of the iteration comprises of employing OMP to estimate The combination of NSCT (multi-scale, multi-direction, shift-invariant)
coefficients for each patch (an initial dictionary is used for computing and LS-SVM, this method was able to outperform methods like Prob-
sparse approximations of all patches) and updating the dictionary using shrink and BLS-GSM objectively [247].
singular value decomposition for one column at a time. Each patch can The major problem with the dictionary learning methods is their
be estimated from a series of patches from the dictionary. K-SVD simu- computational complexity as they require several iterations and have
lated the era of denoising with learned dictionaries which restores im- unstructured dictionaries which makes them unsuitable for real time
age information using a more adaptive model. However this method still applications [4].
needs improvement in case of large patches due to its computational bur-
den and limited size of the dictionary [197]. In another method known 7. Recent trends in image denoising
as K-LLD, in order to avoid selection of patches with different struc-
tures to be mistaken as similar matches, SKR classification is exploited From local filtering to non-local self-similarity exploration and from
[201]. adaptive thresholding to learned dictionaries, the field of denoising has
The LSSC method is quite similar to K-SVD where similar patches evolved in an inexplicable manner. The study of these large numbers of
are denoised using same sparse decomposition. This method exploits methods of reported in literature is overwhelming, it is nearly impossible
self-similarity criteria of images in a non-local manner [202]. The mo- to discuss each of these pioneering methods. In this arriving section of
tivation behind this method was that, in case of K-SVD slight change in literature we try to embark upon recent trends and main areas of plan
input can lead to significant change in dictionary atoms which is unde- of action witnessed in the field of denoising.
sirable. Therefore LSSC was proposed as an improvisation to K-SVD by In 2010, in [205] Chatterjee and Milanfar eluded the inherent ques-
exploiting similar patches in a non-local manner. This method certainly tion of final bounds of denoising. With so many methods pouring in,
sped up the process of search of atoms in an unstructured dictionary. one ponders is there a final limit to denoising. In this work the au-
This method has found to be giving absolute state-of-the-art denoising thors estimated lower bound on image denoising in terms of MSE. It
performance but it has a high computational complexity. Although LSSC was depicted that despite the phenomenal growth in recent techniques
has achieved satisfactory denoising results by employing clustering in there is still a room for improvement and the field of denoising is not
sparse decompositions however its performance largely depends on ini- dead yet.
tial dictionary which has to be trained offline with high quality images. Further in 2012, these authors proposed the PLOW filter, moti-
The CSR denoising method was able to achieve much higher perfor- vated by the statistical observations listed in the above mentioned work.
mance than LSSC along with lower complexity [4]. They employed a patch based redundant Weiner filter where parametric
Another class of methods which exploit dictionaries and are simi- model was designed considering both geometric and photometric sim-
lar to spatial domain filtering are trained filters. These filters are either ilarity. The method has shown to have given an extremely high visual
nonparametric or semi-parametric and have given excellent denoising quality of images, outperforming the famous BM3D [206].
performance. These filters are similar to spatial domain filters except for Following the lead of intense focus on hybrid domain image de-
the fact that they carry out a non-parametric process in order to calcu- noising Zhang et al. in 2013, proposed an image restoration method

233
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

based on joint statistical modelling (JSM) in the space-transform do- age. Then SUnSAL classifier is employed in order to distinguish noisy
main. This method derives a novel method for exploiting both local and edge related coefficients. The prime idea is to minimise energy of
smoothness and nonlocal self-similarity to generate high fidelity im- noise coefficients using diffusion equations. The noise free image ob-
age reconstruction. JSM is hybrid platform which generates a powerful tained from anisotropic diffusion filtered coefficients and textured co-
amalgam of regularisation based (total variation models) and non-local efficients (which remained unchanged) is evidently artefact free, con-
methods. The effectiveness of proposed algorithm is demonstrated with stitutes edge information outperforming methods like BLS-GSM, Prob-
the help of rigorous experiments for Gaussian as well as impulse noise shrink and work given in [4].
[207]. Over wide range of seamless efforts for image denoising the major
The performance of patch based methods for instance BM3D is in- underlying approach is edge recovery. This approach has been worked
herently limited due to insufficient number of similar patches and due upon in a recent work in [216]. This method consists of an edge detec-
to increasing complexity in case of larger images. Therefore, Talabi and tion and fusion model based on anisotropic diffusion model. A denoised
Milanfar in 2014, contributed significantly to this line of research by image is obtained and edges of denoised images are replaced by edges
designing global image denoising filter (GLIDE) which takes into ac- extracted by various edge detection and regularisation based methods.
count all similar or informative parts of an image. In this method, af- Redefining the boundaries of well-known bilateral filter, a new
ter the pre-filtering of the input image, a small number of pixels are boundary preserving filter named co-occurrence filter has been recently
sampled and are fed as inputs to Nystrom equation method. The fi- designed by weighing pixels which occur frequently rather than con-
nal image is estimated by thresholding the eigen values of filter matrix sidering the Gaussian range. The low-occurring pixels on other hand
[208]. would have lower weights in the co-occurrence matrix. This approach
Another work based on hybridisation of interscale and intrascale de- resulted in boundary preservation besides recovery of edges. This work
pendencies in Direction let domain was proposed in [209] in 2015. The introduced the definition of co-occurrence matrix and its application to
image is first transformed into the Direction let coefficients and then the denoising and pioneers a direction for designing various new methods
coefficients are modelled using bivariate PDF to exploit the in-scale and which learns co-occurrence matrix [217].
cross-scale dependence. This method gives far better performance than In case of stochastic approach towards image denoising, an alterna-
bivariate shrinkage in the wavelet domain. tive approach based on Markov-Chain Monte Carlo sampling has been
Another motivating approach towards image denoising is found in given in literature. This method dynamically adapts itself to image char-
[210]. This work on addressed a natural question to adaptively replace acteristics and noise statistics providing high performance in terms of
images using prior of an image scene. It exploits the external and inter- image denoising while exhibiting lower complexity. This method for-
nal correlations thereby clustering the similar patches of noisy and true mulates the problem on the basis of AWGN, computing prior from the
images available on web. This method suggested a two-stage image de- noisy image [218]. This additive image denoising problem is formu-
noising using graph based optimisation and transform domain filtering. lated as a Bayesian least squares problem, where the goal is to esti-
The method was able to outperform BM3D over wide range of noise val- mate the denoised image given the noisy image as the measurement and
ues. Recently in 2018, some authors studied the image denoising prob- an estimated posterior. The posterior is estimated using a nonparamet-
lem as deduction of a non-linear diffusion model into a cross-diffusion ric importance-weighted Markov-Chain Monte Carlo sampling approach
problem [211]. based on an adaptive Geman–McClure objective function [219]. Some
Patch based processing has emerged as the core idea in image approaches improvisations of this method have been given in [220–
restoration problems. This process consists of decomposing the image 223].
into similar overlapping patches and restoring them by plain averag- Out of many approaches towards denoising the primary and current
ing. This method, besides image denoising, has been widely employed focus of researchers is exploiting non-local self-similarity, combining
in areas of image inpainting and deblurring. However, major flaw in them with other structure preserving techniques or improvising the al-
these patch based methods is imposition of prior on intermediate re- ready existing self-similarity based approaches like BM3D and NLM.
sults, rather than processing the final outcome. The EPLL was designed A new patch matching method employing nearest neighbour collab-
in order to address this issue. In a recent work in 2016, a method to ex- orative filtering (NN-CF) is demonstrated in [254]. This work presents
tend and improve EPLL has been proposed [212]. This method not only two variations of CF based patch making criterion. By employing new
imposes prior on final image but constructs a multi-scale prior for differ- patch making approach to various patch matching based denoising
ent scale patches extracted from the input image. This method achieved methods like BM3D and NLM, authors depicted robustness to patch mak-
considerable improvement over the major state-of-the-art image restora- ing and new formulation to integrate external and internal denoising. An
tion algorithms including BM3D. image denoising approach using generalised Cauchy filter is illustrated
Another optimised pixel based weighted averaging approach has in [255] by using the particle swarm optimisation.
been designed for patch based image denoising in [213]. The various A detail preserving cluster wise denoising method providing a sta-
image denoising algorithms restores clean image while leaving behind ble noise level estimation is given in [256]. This method finely clus-
the residual. The residual image left behind should contain uncorrelated ters non local patches with diverse features using adaptive clustering
contaminating noise, but it contains some remnants from the clean im- method. The noise in this technique is removed by employing three
age as well. In a patch based image denoising algorithm, a regularisation steps namely progressive PCA thresholding; by SVD thresholding based
approach was proposed to render the residual patches as uncorrelated on Marchenko–Pastur law and coefficient wise PCA domain LMMSE fil-
as possible. With this the authors generated an analytical solution for tering.
sparse coding and a new online dictionary [214]. Out of numerous improvisations of NLM denoising a work that effec-
It has been discussed that NSST has depicted exceptional perfor- tively exploits Affine Invariant Self Similarity (AV-SS) prevalent in natu-
mance in terms of multi-resolution analysis and shift invariance. This ral images is given in [257]. This method enables selection of more sim-
approach was further exemplified in a work, which constructs a fea- ilar patches by adapting to different size and shapes of patches by per-
ture vector based spatial regularity in NSST domain using SUnSAL. The forming appropriate patch comparison using AV-SS. It has been shown
SUnSAL classifier distinguishes between edge based and noise based co- that this method achieves top-tier performance in terms of quantitative
efficients which are threshold using Bayes Shrink method [215]. analysis.
In 2016, pioneers, Shahdoosti and Khayat presented an innova- In contrast to assuming fixed noise problem, targeting real time de-
tive and effective method for denoising of images by the hybridisa- noising problems authors in [258], proposed a blind denoising method
tion of NSST, SUnSAL and ansitropic diffusion [249]. To obtain sev- which permits estimation of noise from a single image which is sig-
eral scale and directional components, NSST is applied to noisy im- nal and frequency dependent. The efficient performance of the method

234
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

is demonstrated on real JPEG images where formation model was un- ploits the approach of residual image containing the remnants of clean
known and it depicted state-of the-art performance. image.
Besides BM3D, PLOW, LPG-PCA and Shape adaptive SVD, a method
employing non-local self-similarity and low rank approximation is given
8. Applications of image denoising
in [259] and is proposed as a modification of patch based methods.
Initially similar patches are classified using block matching technique.
Image denoising is fundamental image processing problem which
Then SVD is employed to factorise each group of similar patches and
finds applications in various fields of technology. With voluptuous
only singular vector and largest singular values are selected. The final
growth of science and technology, scope of human action is expand-
image is generated by aggregating processed patches. This method be-
ing, creating higher demand of information which results in high reso-
sides effectively removing noise provided optimal energy compaction
lution images. The noise disturbs this information and image denoising
and is shown to be computationally efficient.
becomes increasingly important for subsequent image analysis. Image
Having attained a tremendous progress image representation, limit-
denoising not only generates visually pleasant images, but also “uncov-
ing the computational burden of algorithm became the principal niche
ers” the corrupted information pixel for further image analysis and in-
in denoising. Non-local approach has been evidently depicted as most
formation extraction. The process of image denoising has left no area
efficient criterion for denoising. However it was found that some of the
of application unaffected which includes medical imaging, biometrics,
nearest neighbour approaches introduced bias in results. Therefore Fro-
remote sensing, HVS (Human Visual System), military surveillance and
sio and Kautz proposed a statistical neighbour approach in 2019 which
infrared image denoising.
depicted that only fewer statistical nearest neighbours are required to
generate images of better quality both in case of grey scale and coloured
images [260]. 8.1. Medical image denoising
In another work outlying limiting performance of BM3D on seismic
data, authors designed an improvisation to BM3D by including a fac- Magnetic resonance imaging is used for the visualisation of soft tis-
tor, local similarity indicator, between noisy signal and true signal. This sues of various organs of the body. MRI’s are invariably corrupted with
method prevented drainage of signal energy by recovering it and adding various noise factors like variable field strength, radio frequency pulses
back to the denoised result. This intriguing approach has been proven or due to receiver bandwidth. In [223], motivated by radon transform,
to ensure fidelity of structure representation in final results [261]. a wavelet based multi-scale MRI noise suppression method has been
Palmprint recognition is a reliable identification method in biomet- given. The noise present was modelled as the Gaussian Noise and trans-
rics and is an important image processing application. However pres- lational invariant wavelet transform was used to decompose noisy MR
ence of noise in low resolution palmprints significantly degrades bio- image into multiple scales. The noise was estimated using Rician noise
metrics methods based on them. In [262] a generative adversial network distribution and final denoised image was obtained by using inverse
method has been proposed to mitigate presence of noise which reduced Radon transform. The method was evaluated using Phantom, simulation
error rate from 10.841% to 1.532% after denoising [262]. In order to and human brain MR images. The results for simulation brain images are
preserve complementary contextual information and to reconstruct fine depicted in [223] have been given Fig. 4.
patterns and textures, Li et al. proposed multi-scale gated fusion network
(MGFN) as an improvisation of deep convolutional network for image
denoising. This method was able to put forth a simple but effective loss
function to train the network [263].
With a means to remove noise from different spectral bands of hyper-
spectral images a denoising problem is presented as band fusion prob-
lem. In [264], first a target noisy band is selected and in order to denoise
it, superior quality bands related to it are fused together. The method
has been depicted to shown as superior to various existing HSI denoising
methods in terms of visual and objective analysis.
With an objective to harness the attributes of various techniques
a hybridised denoised method is proposed in [265], which integrates
TV regularisation and group sparsity related to self-similarity of image
blocks. This concatenated technique consists of four steps i.e. match-
ing of blocks, updating of basis vectors, sparsity regularisation and TV
smoothing. In contrast to various other which employ fixed basis func-
tions this method used updated local basis vectors derived from SVD
to induce sparsity. This method has been shown to be outperforming
methods such BM3D itself [265]. Another recently published denoising
method combines the nuclear norm, TV norm, and L1 norm, which fa-
cilitates to exploit the low- rank property of natural images, enhancing
smoothness and removing sparse noise [266]. Fig. 4. Noisy data set for simulation brain images (first column); denoised im-
While exploiting the concept of method noise in deep learning, a ages for Weiner filter (second column); wavelet (third column); (fourth column)
deep feed forwarding denoising convolutional neural network based [223].
method has been proposed in [267] to denoise medical images. Un- The Computed Tomography is yet another form of imaging technolo-
like other methods, this method carries outs learning of residual noise gies which suffers from image degradation due to noise owing to the in-
from noisy image where denoised images are obtained by subtract- herent problem of hardware restriction. Therefore a lot of methods have
ing learned residual from noisy image. Experimental results reveal been designed in order denoise the CT images [269]. In [224], authors
that this approach has better performance than some other methods. propose to preserve edges, corners, textures and sharp edges with the
Another residual correlation based regularisation for image denois- help of Tetrolet transform based locally shrinkage rule to denoise high
ing method is given [268]. In this method analytical and a novel on- frequency coefficients. The proposed method in this work revealed good
line dictionary update is performed generating solution for sparse cod- performance in terms of noise removal and structural preservation. The
ing i.e. atom selection and coefficient calculation. This method ex- method has been evaluated in terms of MSE and PSNR and the visual

235
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Fig. 5. Original CT image; noisy image at 𝜎 = 20; de-

noised image with biorthogonal wavelet thresholding;
tetrolet shrinkage rule (Left to Right) [224].

results for one of the data sets are reproduced from 224 and are given
in Fig. 5.

8.2. Underwater image denoising

Underwater laser imaging has been rapidly employed in order to

acquire information regarding underwater regions. However the image
resolution degrades due to the corruption of acquisitions with noise due
to watery medium. The major reasons of occurrence of noise in underwa- Fig. 7. Kochi LANDSAT image; direction estimation; denoised image with HDL
ter imaging is attenuation and back scattering of laser waves by water (left-right) [226].
resulting in poor image quality. A work in [225], analyses the appli-
cation of MATLAB encoded image denoising algorithms for underwater
images. The underwater source images have been restored using median 8.4. SAR image denoising
filter, average filter and a hybrid method based on wavelet transform.
The results for grey scale images have been given in Fig. 6. To despeckle the SAR (synthetic aperture radar) images, is one of the
important research problems in the field of remote sensing. This imaging
technology finds wide range of applications in military surveillance. To
overcome the limitations of spare-representation based image denoising
methods, a novel image denoising algorithm based on sparse representa-
tion in shearlet domain and continuous cycle spinning has been given in
[227]. After the decomposition of the noisy image using shearlet trans-
form, an optimal denoising model was constructed using cycle spinning
theory in the sparse representational domain. The method not only im-
proved PSNR but also considerably improved the visual results which
have been shown in Fig. 8 for reader comprehension.

Fig. 6. Underwater noisy image; denoised images using averaging filter (first
row L-R); median filter; wavelet method (second row L-R) [225].

Fig. 8. Noisy SAR image; SAR denoised image [227].

8.3. Remote sensing-image denoising

Remotely sensed satellite images are invariably corrupted with noise 8.5. Infrared image denoising
either during signal transmission or acquisition. The removal of high fre-
quency noisy components becomes inevitable in order to improve the The infrared imaging technology is employed in order to capture the
visual appearance, extract and analyse the satellite imaging informa- thermal radiation emitted by the metal objects and it finds extensive
tion. The exploitation of directional correlation in spatial and transform application in the night mode imaging and military surveillance. The
domain has been a common idea for image denoising. However the suc- infrared images are corrupted with Poisson noise due to low-light con-
cess of these methods largely depends on the amount of orientation cor- ditions and atmospheric perturbations. In [228] Anscombe transform
relation measurement. Therefore in [226], author’s proposed a hybrid has been used in order generalise the Poisson distribution into Gaus-
directional lifting scheme for denoising of satellite images. This method sian distribution and to denoise the infrared images. The method was
employed pixel classification and orientation estimation both. To deal comprised of total variation regularisation based improvisation in the
with denoising of the land remote-sensing satellite images (LANDSAT). wavelet domain in order to remove heavy Poisson noise from low-light
The results for Kochi dataset have been depicted in Fig. 7. infrared images and same is being depicted in Fig. 9.

236
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Fig. 9. Noisy INFRARED image; denoised re-

sults by NLM; BM3D; method given in [228].

9. Experimental setup One needs to be aware of the fact that PSNR improvement or
any other objective metric’s comparative increment is not the ultimate
In order to present a performance comparison various existing pop- objective of image denoising. The primary goal of image denoising is to
ular denoising techniques have been analysed on the basis of (i) Visual generate denoised images with high visual quality, better edge preser-
Analysis; (ii) Objective Analysis. The source image dataset is chosen as vation and pattern identification which form an integral part of the Hu-
House Image; MR (Magnetic Resonance Image) Image; PAN (Panchro- man Visual System (HVS). Studies related to HVS reveals that human
matic Image) [229]; so as to reflect adequate amount of diversity in eye can hardly comprehend slight differences in orientations, scales and
the complexity of image content. These images include the natural im- texture surface. However it is a lot more sensitive to sharp contours,
age, medical image and remote sensing image. The experiments are deformed faces, words and symbols, artefacts in smooth regions. Also
carried out on Intel® Core(TM) i5-72000U- CPU @2.50 GHz processor visual analysis is very important for comparison of different image de-
and 8 GB memory using MATLABTM . The resolution of these images is noising methods, because human is usually the end user Therefore the
256 × 256. visual analysis is weighed more for discussion of any type of image de-
noising technique.
9.1. Image contamination model In this section, the objective evaluation metric Peak Signal-Noise Ra-
tio has been used for performance comparison amongst various meth-
The AWGN noise of standard deviation 𝜎 = 10, 20, 30, 40 and 50 has ods. The experimental analysis section is vital to comprehensive strength
been added to source images, so as to check consistency in performance of the review. We have selected methods from various classes of de-
of denoising methods. noising. All the methods chosen for experimentation are tested on our
own versatile dataset (medical image, natural image and remote sens-
9.2. Objective evaluation metrics ing image) for additive white Gaussian Noise. In this section we anal-
yse the actual performance of various methods on varying noise lev-
In order to quantify numerically the performance of a denoising al- els and different types of images. A lot of methods have been ex-
gorithm in terms of its extent of noise removal and ability of edge preser- perimentally demonstrated in literature on variety of dataset. In this
vation various objective metrics have been devised and are being read- manuscript we have tried to choose some representative methods from
ily employed in literature. The major performance metrics employed each class of denoising. Some of these methods have been reviewed in
in field of denoising are (in the order of their popularity): PSNR (Peak the text above and some are cited for reader reference as discussing
Signal-to-Noise Ratio), MSE (Mean Square Error) SSIM (structural sim- all the methods is nearly impossible. Moreover these methods will be
ilarity), VIF (Visual Image Fidelity), Entropy, AD (Average difference), covered at a stretch in terms of objective and subjective evaluation.
MD (Maximum Difference), Normalised Absolute Error (NAE) and Nor- Under the category of spatial domain filters methods employed are
malised Correlation [4,230–242] The mathematical formulas of these SBF [48], RBF [54], WBF [54], NLM [2], bitonic filter [75], Diffusion
metrics along with their significance is summarised in Table 3. filter assortments (DF(A-B)) [242], Gaussian filter [8], Median Filter

Table 3
Quantitative performance metrics.

Metric Mathematical formula Significance

∑
𝑀 ,𝑁 [𝐼1 (𝑚,𝑛)−𝐼2 (𝑚,𝑛)]2
MSE 𝑀×𝑁
Lower MSE depicts better image restoration
𝑅2
PSNR 10𝑙𝑜𝑔10 ( 𝑀𝑆𝐸 ), 𝑅 = 255 Higher PSNR means better quality of the denoised image; frequently
applied image quality metric in literature
(2𝜇𝑥 𝜇𝑦 +𝐶1 )(2𝜎𝑥𝑦 +𝐶2
SSIM (x, y) (𝜇𝑥2 +𝜇𝑦2 +𝐶1 ) (𝜎𝑥2 +𝜎𝑦2 +𝐶2 )
C1 and C2 are constants, 𝜇 x and 𝜇 y are the mean Metric correlated to human perception; high SSIM indicates more
intensities, 𝜎 is the standard deviation restoration of original information
∑
𝐼(𝐷𝑚,𝑘 ;𝑇 𝑚, 𝑘 |𝑠𝑚,𝑘
VIF ∑𝑘𝜖𝑠𝑢𝑏𝑏𝑎𝑛𝑑𝑠 𝑚,𝑘 ;𝑅𝑚, 𝑘 |𝑠𝑚,𝑘
where Dm, k represents m elements of the random Information fidelity criterion; quantifies the Shannon information which is
𝑘𝜖𝑠𝑢𝑏𝑏𝑎𝑛𝑑𝑠 𝐼(𝐷

field that describes the coefficients from subband k, Tm, k and Rm, k shared between the distorted and the reference image, fidelity aspect
depicts the visual signal at the output of the HVS model for correlating with visual quality
reference and test images
∑
Entropy SE = 𝐴2𝑖 log(𝐴2𝑖 ); EN is a statistical computation of randomness which conveys the amount of
𝑖
(EN) texture information;
EN= SE(A)-SE(B), where SE stands for Shannon entropy
∑𝑀 ∑𝑁
𝑥=1 𝑦=1 (𝐴(𝑥,𝑦)−𝐵(𝑥,𝑦)
AD 𝑀×𝑁
Gives the average difference between the clean and denoised image
MD 𝑀 𝐷 = 𝑀 𝐴𝑋|𝐴(𝑥, 𝑦) − 𝐵(𝑥, 𝑦)| Lower value of MD means better image quality
∑𝑀 ∑𝑁
𝑦=1 (𝐴(𝑥,𝑦)−𝐵(𝑥,𝑦))
NAE 𝑁𝐴𝐸 = 𝑥=1
∑𝑀 ∑𝑁 Gives the normalized rate of error; lower NAE gives better denoised images
𝑥=1 𝑦=1 (𝐴(𝑥,𝑦)
∑𝑀 ∑𝑁
𝑥=1 𝑦=1 (𝐴(𝑥,𝑦)×(𝐵(𝑥,𝑦)
NCC ∑𝑀 ∑𝑁 Depicts the amount of correlation between the clean and denoised image
𝑦=1 (𝐴(𝑥,𝑦)
2
𝑥=1

Assumption: A and B represent the clean and denoised images respectively and M × N, is the size of the image.

237
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Table 4
Performance comparison of various state-of-the-art denoising techniques.

PSNR
Denoising
MRI HOUSE PAN
methods
𝜎 = 10 20 30 40 50 10 20 30 40 50 10 20 30 40 50

SBF 32.18 25.78 20.79 17.44 15.13 33.23 26.11 20.95 17.60 15.14 28.62 23.44 19.60 16.91 14.74
RBF 30.20 29.59 28.54 27.13 25.63 31.75 30.98 29.69 28.19 26.37 19.54 19.37 19.14 18.84 18.40
WBF 32.55 30.18 28.64 27.15 25.64 33.64 31.30 29.74 28.19 26.19 28.66 24.03 21.56 20.02 19.22
INLM 35.01 31.18 29.37 27.79 26.49 35.63 32.88 30.93 29.44 28.04 28.32 24.10 21.61 19.87 18.73
Bitonic 29.39 27.27 25.25 23.25 21.70 31.89 29.59 27.29 25.51 24.00 20.16 19.88 19.48 16.91 14.74
DF(A) 25.33 24.57 23.65 22.51 21.46 24.66 24.47 24.13 23.73 23.28 17.12 17.02 16.91 16.76 16.60
DF(B) 23.08 22.64 22.07 21.33 20.56 22.17 22.12 22.00 21.82 21.63 15.67 15.62 15.55 15.46 15.36
DF(C) 28.40 26.98 25.48 23.93 22.62 22.21 21.99 21.64 21.12 20.63 17.84 17.69 17.45 17.22 16.90
DF(D) 28.45 26.84 25.16 23.44 22.01 21.75 21.59 21.32 20.98 20.61 18.38 18.18 17.92 17.59 17.26
vi-hard 29.11 25.26 22.86 21.20 19.92 30.69 27.58 25.96 24.53 23.45 23.14 19.12 17.15 16.02 15.32
vi-soft 26.14 22.76 20.82 19.48 21.53 28.05 25.23 23.67 22.45 21.64 19.79 16.85 15.45 14.60 14.03
GF 26.38 25.38 24.21 22.84 21.64 29.67 28.88 28.75 26.72 25.65 18.13 18.01 17.84 17.63 17.39
MF 24.35 23.74 22.96 21.98 21.04 27.80 27.48 26.93 26.39 25.70 16.92 16.84 16.74 16.61 16.46
SH3D 29.51 23.53 20.05 17.63 15.81 28.10 22.10 18.65 16.29 14.55 28.31 22.40 19.01 16.67 14.93
BM3D 35.84 32.25 30.28 28.63 27.77 36.71 33.77 32.08 30.64 29.69 29.32 24.62 22.25 20.70 19.51
SAPCA 35.82 32.30 30.29 28.89 27.78 37.01 33.89 32.12 30.74 29.52 29.62 24.95 22.25 20.70 19.51
SURELET 34.55 31.14 29.05 27.35 26.25 35.29 32.06 30.03 28.47 27.19 28.91 24.14 21.85 20.37 19.30
GBFMT 29.56 23.58 20.10 17.69 15.87 28.14 22.14 18.68 16.33 14.60 28.37 22.45 19.07 16.72 14.97
NLFMT 29.52 23.55 20.08 17.66 15.84 28.13 22.12 18.67 16.31 14.58 28.33 22.41 19.02 16.68 14.94
NSST 34.37 30.83 29.05 27.55 26.74 34.86 31.98 30.30 28.97 27.89 26.89 22.40 20.52 19.34 18.47
MRF 30.64 26.61 20.80 17.99 16.32 32.80 26.39 19.15 16.51 14.93 26.88 23.27 19.24 16.77 15.18
NLGRTV 24.12 23.71 20.14 22.54 22.27 26.87 26.62 25.48 25.24 25.67 17.43 17.26 19.00 16.65 16.73
LPGPCA 35.56 31.77 29.63 28.10 26.94 36.12 33.07 31.20 29.72 28.50 28.88 24.08 21.68 20.16 19.08
NCSR 31.48 30.43 27.73 25.00 18.26 36.78 33.86 32.11 30.52 29.05 24.85 24.49 21.99 201.7 18.83
Ker. Reg. 28.45 27.63 26.01 22.86 18.91 29.94 30.89 30.57 24.44 17.52 23.90 23.16 21.50 18.61 15.76
GFOE 33.77 28.91 25.90 23.29 21.46 34.01 30.10 27.19 24.83 22.57 23.09 22.09 20.92 20.05 19.27
PGPCA 34.43 30.14 26.43 23.97 22.30 35.18 32.00 30.13 28.88 27.56 28.95 23.88 21.27 19.64 18.50
PLPCA 34.54 30.37 26.65 24.27 22.60 35.81 32.39 30.46 28.88 27.76 29.20 24.05 21.44 19.82 18.65
PHPCA 34.47 30.33 26.65 24.27 22.54 35.67 32.38 30.39 29.08 27.84 29.11 23.96 21.36 19.67 18.55
TV1 28.71 27.04 25.96 24.99 24.21 31.78 29.93 28.49 27.39 26.47 19.45 18.94 18.39 17.85 17.41
Deep NN 34.19 30.63 29.43 27.68 26.45 36.02 32.96 31.78 30.11 29.61 29.30 23.67 21.56 19.63 19.37

[8,15], Gaussian Field of Experts (GFOE) [270], Total variation min- MRI, BM3D and its Shape adaptive principal component analysis version
imisation [276], non-local version of general relative total variation gives highest value of PSNR. Same is the case for House image as PSNR
(NLGRTV) [271]; the different transform domain methods include Vi- for BM3D, SAPCA and NCSR is the highest. However in case of PAN im-
hard [94], Vi-soft [95], SURELET [97], NSST [59, 144]. Under the cat- age only SAPCA is able to give highest amount of PSNR while NCSR and
egory of statistical methods we have included Markov Random Fields BM3D give a comparatively lower value. It can be seen that the type of
(MRF) [163]. To illustrate an idea of different sparse representation and data set plays a significant role in determining the performance of the
dictionary learning methods we have included Nonlocally Centralised specific algorithm; as SAPCA gives PSNR 32 for house image and 22 for
Sparse Representation (NCSR) [272], Locally Adaptive Kernel Regres- PAN at standard deviation 30.
sion (LARK) [273], deep neural networks (NN) [274] in experimental Further the second level for PSNR is begged by methods like DNN,
analysis. Besides this some of the other representative methods include LPGPCA, NSST, SURELET, and INLM for MRI; NCSR LPGPCA NSST,
LPGPCA [153], BM3D-SH3D (joint image sharpening and denoising) SURELET, and INLM for House image; and WBF, DNN, PLPCA, PGPCA,
[243], BM3D [151], BM3D-SAPCA [157], PGPCA (Patch based Global NCSR and LPGPCA for PAN image at standard deviation 30.
PCA), PLPCA (Patch based local PCA), PHPCA (Patch Based Hierarchical Other methods like Bitonic filter, SBF and DF (A-D) gives lowest val-
PCA) [275]. The methods employed under hybrid domain are: GBFMT ues of PSNR for all three types of data sets.
[193], NLFMT [194]. The objective results for these methods in terms
of Peak Signal to Noise ratio at low to high noise levels have been given 9.4. Visual analysis
in the Table 4. In Fig. 10, the visual results at standard deviation of
30 are being shown for House image only due to space constraint. The For the sake of minimalism, visual results for denoised images at
experimental results given in Table 4 and Fig. 10 above summarise standard deviation 30 have been shown in the manuscript. Referring
the subjective and objective evaluation of various image denoising to Fig. 10, it is quite apparent that non-local similarity and patch based
techniques. methods such as BM3D, SAPCA, INLM, LPGPCA and NCSR demonstrates
top tier performance as in the resultant image: edges are well preserved;
9.3. Objective analysis visual quality is high, homogenous regions are free from noise, low-
contrast objects are visible, textures are also well preserved and there is
In order to quantify and analyse the performance of different do- no emergence of artefacts. This conforms the fact, that the idea of non-
mains of image denoising, we have simulated MR image, House image local grouping forms the major reform in image denoising. These visual
and PAN image with AWGN at standard deviation 10, 20, 30, 40 and results resonate with the objective evaluation for these best performing
50. The noise from these images is removed using 31 image denoising techniques.
methods representative of their domains and PSNR is calculated. The Further methods such as NSST, DNN, PLPCA, PHPCA and TV1 gives
parameter settings for these methods have been selected as given by au- considerable amount of denoising and corner and edges are also pre-
thors in reference papers. It is clearly evident in Table 4 that in case of served. However these methods induce some ringing artefacts around

238
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

Fig. 10. House image denoised using various

methods at standard deviation 30.

edges. Some methods like PGPCA, Ker. Reg., NLFMT and SURELET tend to induce artefact around edges and methods like WBF, RBF and
causes over-smoothing and generates artificial textures. Rest of the Bitonic filter are able to preserve edges but they tend to give poor de-
methods are rather performing poorly both in case of denoising abil- noising performance in homogenous regions.
ity and edge preservation. Besides the varied level of performance of different methods it can be
The SURELET transform gives higher PSNR value than shearlet trans- seen that almost all methods depicts degradation in performance with
form, however, it can be seen that in case of visual results the amount increasing levels of noise.
of denoising performed by SURELET is lower which may be due to
poor preservation of edge feature and the noise still persists. The WBF 10. Conclusion
and RBF show improved results than SBF in terms of their denoising
performances at higher noise level. This can be attributed to the fact In this article, an earnest effort has been made in order to compare,
SBF simply disables denoising at higher noise levels, whereas WBF ade- classify and evaluate various image denoising methods. Extensive efforts
quately removes the high levels of noise along with the preservation by a huge number of researchers have generated a structural literature
of edges. Some of the hybrid methods like GBFMT and NLFMT are which exhibits substantial progressive growth attained by a chain of
also able to achieve fair amount of denoising comparable to vi-hard sequential incremental improvements. While it is nearly impossible to
and vi-soft. Besides, these hybrid methods give much better visual re- cover all of them, we have covered each domain of image denoising with
sults than the conventional wavelet thresholding methods. The details several methods representative of each category. These methods have
and edges are drastically lost in the visual results given by vi-hard and been divided into five categories namely spatial domain methods, trans-
vi-soft. form domain methods, methods in statistical domains, hybrid methods
Besides this there is a striking observation that although NCSR and sparse representation and dictionary learning methods. The basic
gives high quality visual results, its quantitative analysis shows a lower idea of image denoising is edge preservation while getting rid of noisy
level of PSNR as compared to top tier assessments. On the other hand pixels. The era of image denoising started from basic averaging filter
SURELET method is depicting considerable level of PSNR, it gives low and later on transformed into edge preserving filters. In spatial domain,
visual quality. Methods like MRF model are performing moderately both improved non-local means filter perform well. Through various analyt-
objectively and subjectively. ical procedures we observed that exploitation of non-local image simi-
Therefore it can be concluded that methods like BM3D and SAPCA larity or non-local grouping stems the major reform in improvement of
preserves both edges and gives smooth homogeneous regions. Methods results for image denoising for instance NLM filters and BM3D. In trans-
like NSST and DNN are able to preserve low contrast details but they form domain methods, transforms with overcomplete basis functions,

239
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

increased number of scales, directions and orientations generated bet- [16] R. Yang, L. Yin, M. Gabbouj, J. Astola, Y. Neuvo, Optimal weighted median filters
ter results. The properties like multiresolution, sparsity, shift invariance under structural constraints, in: ISCAS, 1993, pp. 942–945.
[17] R.C. Hardie, K.E. Barner, Rank conditioned rank selection filters for signal restora-
and anisotropy are highly needed for an image transform to work well tion, IEEE Trans. Image Process. 3 (2) (1994).
for image denoising. The exploration of overcomplete basis functions in [18] P. Perona, J. Malik, Scale-space and edge detection using anisotropic diffusion,
BLS-GSM resulted in improved performance than wavelets with fixed ba- IEEE Trans. Pattern Anal. Mach. Intell. 12 (7) (1990) 629–639.
[19] P. Perona, J. Malik, Scale space and edge detection using anisotropic diffusion, in:
sis functions. Multi-resolution based algorithms perform far better than Proc. IEEE Computer Society Workshop on Computer Vision, Dec., IEEE Computer
single resolution ones. Society Press, Washington, 1987, pp. 16–22.
The spatial domain filters is capable of eliminating high frequency [20] P. Perona, T. Shiota, J. Malik, Anisotropic diffusion, in: Geometry-Driven Diffusion
in Computer Vision, Springer, 1994, pp. 73–92.
noise but at the cost of blurring of low contrast information, whereas
[21] C. Tsiotsios, M. Petrou, On the choice of the parameters for anisotropic diffusion
transform domain methods were able to preserve details in homoge- in image processing, Pattern Recognit. 46 (5) (2013) 1369–1381.
neous regions but tend to introduce ringing artefacts around the edges. [22] A. Dogra, B. Goyal, S. Agrawal, Osseous and digital subtraction angiography image
fusion via various enhancement schemes and Laplacian pyramid transformations,
The dictionary learning and sparse representation methods possesses
Future Gener. Comput. Syst. 82 (2018) 149–157.
redundant and overcomplete dictionaries, but they often lead to com- [23] Y.-L. You, W. Xu, A. Tannenbaum, M. Kaveh, Behavioral analysis of anisotropic dif-
putational burden. fusion in image processing, IEEE Trans. Image Process. 5 (11) (1996) 1539–1553.
It is observed that an ideal denoising algorithm should possess over- [24] O. Scherzer, J. Weickert, Relations between regularization and diffusion filtering,
J. Math. Imaging Vis. 12 (1) (2000) 43–63.
complete representation, non-local grouping, adaptive basis functions [25] L. Alvarez, P.-L. Lions, J.-M. Morel, Image selective smoothing and edge detection
and be computationally efficient. A single denoising algorithm is not by nonlinear diffusion. II, SIAM J. Numer. Anal. 29 (3) (1992) 845–866.
able to exhibit desired level of applicability which is true even for [26] S. Osher, L.I. Rudin, Feature-oriented image enhancement using shock filters, SIAM
J. Numer. Anal. 27 (4) (1990) 919–940.
BM3D (non-local grouping leads to search complexity). Therefore it was [27] C.R. Vogel, M.E. Oman, Iterative methods for total variation denoising, SIAM J.
suggested to combine various methods in different domains to harness Sci. Comput. 17 (1) (1996) 227–238.
the attributes of various domains while overcoming limitations of each [28] L. Alvarez, L. Mazorra, Signal and image restoration using shock filters and
anisotropic diffusion, SIAM J. Numer. Anal. 31 (2) (1994) 590–605.
other. The hybrid domain methods have attained brilliant levels of vi- [29] G. Gilboa, N. Sochen, Y.Y. Zeevi, Forward-and-backward diffusion processes for
sual quality of denoised images besides substantial increment in PSNR. adaptive image enhancement and denoising, IEEE Trans. Image Process. 11 (7)
Besides this, it can be stated that HVS is primary evaluation factor for (2002) 689–703.
[30] G. Gilboa, N. Sochen, Y.Y. Zeevi, Image enhancement and denoising by complex
performance evaluation of these methods. It is clear from detailed sur-
diffusion processes, IEEE Trans. Pattern Anal. Mach. Intell. 26 (8) (2004) 1020–
vey that each category of methods is important and has its own specific 1036.
set of applications. For instance, time sensitive applications require lo- [31] Z. Qiu, L. Yang, W. Lu, A new feature-preserving nonlinear anisotropic diffusion
for denoising images containing blobs and ridges, Pattern Recognit. Lett. 33 (3)
cal filtering or transform domain methods as non-local grouping leads
(2012) 319–330.
to computational burden. Though dictionary learning methods signifi- [32] F. Catté, P.-L. Lions, J.-M. Morel, T. Coll, Image selective smoothing and edge de-
cantly boost the performance but they are not fit where memory and tection by nonlinear diffusion, SIAM J. Numer. Anal. 29 (1) (1992) 182–193.
complexity are major concerns and also online training is not viable in [33] M.R. Hajiaboli, An anisotropic fourth-order diffusion filter for image noise removal,
Int. J. Comput. Vis. 92 (2) (2011) 177–191.
practical applications. [34] M. Lysaker, A. Lundervold, X.-C. Tai, Noise removal using fourth-order partial dif-
Moreover, it has been observed that most of the methods tend to ferential equation with applications to medical magnetic resonance images in space
leave behind residual noise and are not able to sustain their performance and time, IEEE Trans. Image Process. 12 (12) (2003) 1579–1590.
[35] M.R. Hajiaboli, A self-governing hybrid model for noise removal, in: Pacific-Rim
at higher noise levels. The era of denoising is seemingly approaching its Symposium on Image and Video Technology, 2009, pp. 295–305.
bounds. However it still holds margin of improvement in terms of perfor- [36] W. Zeng, X. Lu, X. Tan, Non-linear fourth-order telegraph-diffusion equation for
mance consistency, high visual quality, and preservation of fine feature noise removal, IET Image Process. 7 (4) (2013) 335–342.
[37] Y.-L. You, M. Kaveh, Fourth-order partial differential equations for noise removal,
details, artefact mitigation, more structured dictionaries and adaptive IEEE Trans. Image Process. 9 (10) (2000) 1723–1730.
basis functions. [38] S.-M. Chao, D.-M. Tsai, An improved anisotropic diffusion model for detail-and
edge-preserving smoothing, Pattern Recognit. Lett. 31 (13) (2010) 2012–2023.
[39] D. Chen, S. MacLachlan, M. Kilmer, Iterative parameter-choice and multigrid
References methods for anisotropic diffusion denoising, SIAM J. Sci. Comput. 33 (5) (2011)
2972–2994.
[1] M. Lebrun, M. Colom, A. Buades, J.-M. Morel, Secrets of image denoising cuisine, [40] J. Xu, Y. Jia, Z. Shi, K. Pang, An improved anisotropic diffusion filter with
Acta Numer. 21 (2012) 475–576. semi-adaptive threshold for edge preservation, Signal Process. 119 (2016) 80–91.
[2] A. Buades, B. Coll, J.-M. Morel, A review of image denoising algorithms, with a [41] O. Ghita, P.F. Whelan, A new GVF-based image enhancement formulation for
new one, Multiscale Model. Simul. 4 (2) (2005) 490–530. use in the presence of mixed noise, Pattern Recognit. 43 (8) (2010) 2646–
[3] C. Knaus, M. Zwicker, Progressive image denoising, IEEE Trans. Image Process. 23 2658.
(7) (2014) 3114–3125. [42] Y.Q. Wang, J. Guo, W. Chen, W. Zhang, Image denoising using modified Per-
[4] H.-Y. Yang, X.-Y. Wang, P.-P. Niu, Y.-C. Liu, Image denoising using nonsubsam- ona–Malik model based on directional Laplacian, Signal Process. 93 (9) (2013)
pled shearlet transform and twin support vector machines, Neural Netw. 57 (2014) 2548–2558.
152–165. [43] L.I. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal al-
[5] M. Elad, M. Aharon, Image denoising via sparse and redundant representations over gorithms, Physica D 60 (1–4) (1992) 259–268.
learned dictionaries, IEEE Trans. Image Process. 15 (Dec (12)) (2006) 3736–3745. [44] A. Chambolle, An algorithm for total variation minimization and applications, J.
[6] L. Shao, R. Yan, X. Li, Y. Liu, From heuristic optimization to dictionary learning: a Math. Imaging Vis. 20 (1–2) (2004) 89–97.
review and comprehensive comparison of image denoising algorithms, IEEE Trans. [45] P.L. Combettes, J.-C. Pesquet, Proximal splitting methods in signal processing,
Cybern. 44 (7) (2014) 1001–1013. in: Fixed-Point Algorithms for Inverse Problems in Science and Engineering, 49,
[7] T. Buades, Y. Lou, J.-M. Morel, Z. Tang, A note on multi-image denoising, in: Local Springer, 2011, pp. 185–212.
and Non-Local Approximation in Image Processing, 2009. LNLA 2009. Interna- [46] L. Yaroslavsky, M. Eden, Fundamentals of Digital Optics: Digital Signal Processing
tional Workshop on, 2009, pp. 1–15. in Optics and Holography, Birkhäuser, Boston, 1996.
[8] R.C. Gonzalez, R.E. Woods, Image processing, Digit. Image Process. 2 (2007). [47] L.P. Yaroslavsky, Digital Picture Processing: An Introduction, 9, Springer Science
[9] K. Irie, A.E. McKinnon, K. Unsworth, I.M. Woodhead, A model for measurement of & Business Media, 2012.
noise in CCD digital-video cameras, Meas. Sci. Technol. 19 (4) (2008) 045207. [48] C. Tomasi, R. Manduchi, Bilateral filtering for gray and color images, in: Computer
[10] E. Jakeman, On the statistics of K-distributed noise, J. Phys. A 13 (1) (1980) 31. Vision, 1998. Sixth International Conference on, 1998, pp. 839–846.
[11] R. Yan, L. Shao, L. Liu, Y. Liu, Natural image denoising using evolved local adaptive [49] S.M. Smith, J.M. Brady, SUSAN—a new approach to low level image processing,
filters, Signal Process. 103 (2014) 36–44. Int. J. Comput. Vis. 23 (1) (1997) 45–78.
[12] L. Shapiro, G. Linda, Stockman, Computer Vision, Prentice-Hall, 2001. [50] M. Zhang, B.K. Gunturk, Multiresolution bilateral filtering for image denoising,
[13] J. Benesty, J. Chen, Y. Huang, Study of the widely linear Wiener filter for noise IEEE Trans. Image Process. 17 (12) (2008) 2324–2333.
reduction, in: Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE Inter- [51] M. Elad, On the origin of the bilateral filter and ways to improve it, IEEE Trans.
national Conference on, 2010, pp. 205–208. Image Process. 11 (10) (2002) 1141–1151.
[14] A.K. Jain, Fundamentals of Digital Image Processing, Prentice-Hall, Upper Saddle [52] K.N. Chaudhury, D. Sage, M. Unser, Fast O(1) bilateral filtering using trigonometric
River, 1989. range kernels, IEEE Trans. Image Process. 20 (12) (2011) 3376–3382.
[15] I. Pitas, A. Venetsanooulos, Nonlinear Digital Filters: Principles and Applications, [53] K.N. Chaudhury, Acceleration of the shiftable O(1) algorithm for bilateral filtering
Kluwer, Boston, 1990. and nonlocal means, IEEE Trans. Image Process. 22 (4) (2013) 1291–1300.

240
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

[54] K.N. Chaudhury, K. Rithwik, Image denoising using optimally weighted bilateral [91] I.K. Fodor, C. Kamath, Denoising through wavelet shrinkage: an empirical study,
filters: a sure and fast approach, in: Image Processing (ICIP), 2015 IEEE Interna- J. Electron. Imaging 12 (1) (2003) 151–161.
tional Conference on, 2015, pp. 108–112. [92] S. Sulochana, R. Vidhya, Image denoising using adaptive thresholding in framelet
[55] F. Durand, J. Dorsey, Fast bilateral filtering for the display of high-dynamic-range transform domain, Int. J. Adv. Comput. Sci. Appl. 3 (9) (2012).
images, in: ACM Transactions on Graphics (TOG), 21, 2002, pp. 257–266. [93] A. Dixit, P. Sharma, A comparative study of wavelet thresholding for image denois-
[56] F. Porikli, Constant Time O(1) Bilateral Filtering, 2008. ing, Int. J. Image Graph. Signal Process. 12 (2014) 39–46.
[57] Q. Yang, K.-H. Tan, N. Ahuja, Real-time O(1) bilateral filtering, in: Computer [94] D.L. Donoho, I.M. Johnstone, Adapting to unknown smoothness via wavelet shrink-
Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, 2009, age, J. Am. Stat. Assoc. 90 (432) (1995) 1200–1224.
pp. 557–564. [95] D.L. Donoho, I.M. Johnstone, Threshold selection for wavelet shrinkage of noisy
[58] Z. Farbman, R. Fattal, D. Lischinski, R. Szeliski, Edge-preserving decompositions data, in: Proceedings of 16th Annual International Conference of the IEEE Engi-
for multi-scale tone and detail manipulation, in: ACM Transactions on Graphics neering in Medicine and Biology Society, 1, IEEE, 1994, pp. A24–A25.
(TOG), 27, 2008, p. 67. [96] A. Piˇzurica, W. Philips, Estimating the probability of the presence of a signal of in-
[59] B. Goyal, A. Dogra, S. Agrawal, B.S. Sohi, A three stage integrated denoising ap- terest in multiresolution single and multiband image denoising, IEEE Trans. Image
proach for grey scale images, J. Ambient Intell. Humaniz. Comput. (2018) 1–16. Process. 15 (3) (2006) 654–665.
[60] L. Jin, C. Xiong, H. Liu, Improved bilateral filter for suppressing mixed noise in [97] F. Luisier, T. Blu, M. Unser, A new sure approach to image denoising: inter-
color images, Digit. Signal Process. 22 (6) (2012) 903–912. scale orthonormal wavelet thresholding, IEEE Trans. Image Process. 16 (3) (2007)
[61] M. Zhang, B.K. Gunturk, Multiresolution bilateral filtering for image denoising, 593–606.
IEEE Trans. Image Process. 17 (12) (2008) 2324–2333. [98] F. Luisier, T. Blu, SURE-LET multichannel image denoising: interscale orthonormal
[62] W. Shi, J. Li, M. Wu, An image denoising method based on multiscale wavelet wavelet thresholding, IEEE Trans. Image Process. 17 (4) (2008) 482–492.
thresholding and bilateral filtering, Wuhan Univ. J. Nat. Sci. 15 (2) (2010) [99] G.Y. Chen, T.D. Bui, A. Krzyzak, Image denoising using neighbouring wavelet co-
148–152. efficients, Integr. Comput. Aided Eng. 12 (1) (2005) 99–107.
[63] H. Peng, R. Rao, S.A. Dianat, Multispectral image denoising with optimized vector [100] D. Zhou, X. Shen, Image denoising using block thresholding, in: Image and Signal
bilateral filter, IEEE Trans. Image Process. 23 (1) (2014) 264–273. Processing, 2008. CISP’08. Congress on, 3, 2008, pp. 335–338.
[64] H. Yu, L. Zhao, H. Wang, Image denoising using trivariate shrinkage filter in the [101] J.-L. Starck, F. Murtagh, E.J. Candès, D.L. Donoho, Gray and color image contrast
wavelet domain and joint bilateral filter in the spatial domain, IEEE Trans. Image enhancement by the curvelet transform, IEEE Trans. Image Process. 12 (6) (2003)
Process. 18 (10) (2009) 2364–2369. 706–717.
[65] K. He, J. Sun, X. Tang, Guided image filtering, in: European Conference on Com- [102] M. Lang, H. Guo, J.E. Odegard, C.S. Burrus, R.O. Wells, Noise reduction using an
puter Vision, 2010, pp. 1–14. undecimated discrete wavelet transform, IEEE Signal Process. Lett. 3 (1) (1996)
[66] X. Zhang, X. Feng, W. Wang, Two-direction nonlocal model for image denoising, 10–12.
IEEE Trans. Image Process. 22 (1) (2013) 408–412. [103] N. Kingsbury, The dual-tree complex wavelet transform: a new efficient tool for
[67] M. Mahmoudi, G. Sapiro, Fast image and video denoising via nonlocal means of image restoration and enhancement, in: Signal Processing Conference (EUSIPCO
similar neighborhoods, IEEE Signal Process. Lett. 12 (12) (2005) 839–842. 1998), 9th European, 1998, pp. 1–4.
[68] S. Wang, Y. Xia, Q. Liu, J. Luo, Y. Zhu, D.D. Feng, Gabor feature based nonlocal [104] W. Ding, F. Wu, X. Wu, S. Li, H. Li, Adaptive directional lifting-based wavelet
means filter for textured image denoising, J. Vis. Commun. Image Represent. 23 transform for image coding, IEEE Trans. Image Process. 16 (2) (2007) 416–427.
(7) (2012) 1008–1018. [105] R.R. Coifman, D.L. Donoho, Translation-invariant de-noising, in: Wavelets and
[69] B. Goossens, H. Luong, A. Pizurica, W. Philips, An improved non-local denoising Statistics, 103, Springer, 1995, pp. 125–150.
algorithm, in: 2008 International Workshop on Local and Non-Local Approximation [106] E.P. Simoncelli, W.T. Freeman, The steerable pyramid: a flexible architecture for
in Image Processing (LNLA 2008), 2008, pp. 143–156. multi-scale derivative computation, in: Image Processing Proceedings., Interna-
[70] S. Xu, Y. Zhou, H. Xiang, S. Li, Remote sensing image denoising using patch tional Conference, 3, 1995, pp. 444–447.
grouping-based nonlocal means algorithm, IEEE Geosci. Remote Sens. Lett. 14 (12) [107] R.D. Da Silva, R. Minetto, W.R. Schwartz, H. Pedrini, Adaptive edge-preserving im-
(2017) 2275–2279. age denoising using wavelet transforms, Pattern Anal. Appl. 16 (4) (2013) 567–580.
[71] P. Coupé, P. Yger, S. Prima, P. Hellier, C. Kervrann, C. Barillot, An optimized [108] T. Qiu, A. Wang, N. Yu, A. Song, LLSURE: local linear SURE-based edge-preserving
blockwise nonlocal means denoising filter for 3-D magnetic resonance images, IEEE image filtering, IEEE Trans. Image Process. 22 (1) (2013) 80–90.
Trans. Med. Imaging 27 (4) (2008) 425–441. [109] E.J. Candes, Ridgelets: Theory and Applications, Stanford University, Stanford,
[72] D. Tschumperlé, L. Brun, Non-local image smoothing by applying anisotropic dif- 1998.
fusion PDE’s in the space of patches, in: Image Processing (ICIP), 2009 16th IEEE [110] E.J. Candès, D.L. Donoho, Ridgelets: a key to higher-dimensional intermittency?
International Conference on, 2009, pp. 2957–2960. Philos. Trans. R. Soc. Lond. Ser. A 357 (1999) 2495–2509.
[73] S. Grewenig, S. Zimmer, J. Weickert, Rotationally invariant similarity measures [111] S.R. Deans, The Radon Transform and Some of Its Applications, Wiley, New York,
for nonlocal image denoising, J. Vis. Commun. Image Represent. 22 (2) (2011) 1983.
117–130. [112] M.N. Do, M. Vetterli, The finite ridgelet transform for image representation, IEEE
[74] Y. Teng, Y. Zhang, Y. Chen, C. Ti, Adaptive morphological filtering method for Trans. Image Process. 12 (1) (2003) 16–28.
structural fusion restoration of hyperspectral images, IEEE J. Sel. Top. Appl. Earth [113] E.D. Bolker, The finite Radon transform, Contemp. Math. 63 (1987) 27–50.
Obs. Remote Sens. 9 (2) (2016) 655–667. [114] X. Wang, Wrap-around effect removal finite ridgelet transform for multiscale image
[75] G. Treece, The bitonic filter: linear filtering in an edge-preserving morphological denoising, Pattern Recognit. 43 (11) (2010) 3693–3698.
framework, IEEE Trans. Image Process. 25 (11) (2016) 5199–5211. [115] Q. Huang, B. Hao, S. Chang, Adaptive digital ridgelet transform and its application
[76] Z. Xiong, K. Ramchandran, M.T. Orchard, Y.-Q. Zhang, A comparative study of in image denoising, Digit. Signal Process. 52 (2016) 45–54.
DCT-and wavelet-based image coding, IEEE Trans. Circuits Syst. Video Technol. 9 [116] M.N. Do, M. Vetterli, Image denoising using orthonormal finite ridgelet trans-
(5) (1999) 692–695. form, in: Wavelet Applications in Signal and Image Processing VIII, 4119, 2000,
[77] S. Mallat, A Wavelet Tour of Signal Processing, Elsevier, 1999. pp. 831–843.
[78] M. Misiti, Y. Misiti, G. Oppenheim, J.P. Michel, Wavelet Toolbox: For Use With [117] J.-L. Starck, E.J. Candès, D.L. Donoho, The curvelet transform for image denoising,
Matlab, 1996. IEEE Trans. Image Process. 11 (6) (2002) 670–684.
[79] S.G. Mallat, A theory for multiresolution signal decomposition: the wavelet repre- [118] Z. Chen, S. Wang, G. Fang, J. Wang, Ionograms denoising via curvelet transform,
sentation, IEEE Trans. Pattern Anal. Mach. Intell. 11 (7) (1989) 674–693. Adv. Space Res. 52 (7) (2013) 1289–1296.
[80] C.K. Chui, Wavelets: a tutorial in theory and applications, First and Second Volume [119] B. Zhang, J.M. Fadili, J.-L. Starck, Wavelets, ridgelets, and curvelets for Poisson
of Wavelet Analysis and Its Applications, 1992. noise removal, IEEE Trans. Image Process. 17 (7) (2008) 1093–1108.
[81] J. Kovacevic, W. Sweldens, Wavelet families of increasing order in arbitrary di- [120] F.G. Meyer, R.R. Coifman, Brushlets: a tool for directional image analysis and image
mensions, IEEE Trans. Image Process. 9 (3) (2000) 480–496. compression, Appl. Comput. Harmon. Anal. 4 (2) (1997) 147–187.
[82] G. Pajares, J.M. De La Cruz, A wavelet-based image fusion tutorial, Pattern Recog- [121] D.L. Donoho, Wedgelets: nearly minimax estimation of edges, Ann. Stat. 27 (3)
nit. 37 (9) (2004) 1855–1872. (1999) 859–897.
[83] A. Dogra, Performance comparison of different wavelet families based on bone [122] D.L. Donoho, X. Huo, Beamlets and multiscale image analysis, in: Multiscale and
vessel fusion, Asian J. Pharm. 10 (04) (2017). Multiresolution Methods, Springer, 2002, pp. 149–196.
[84] A. Dogra, B. Goyal, S. Agrawal, C.K. Ahuja, Efficient fusion of osseous and vascular [123] M.N. Do, M. Vetterli, The contourlet transform: an efficient directional multiresolu-
details in wavelet domain, Pattern Recognit. Lett. 94 (2017) 189–193. tion image representation, IEEE Trans. Image Process. 14 (12) (2005) 2091–2106.
[85] V. Bruni, B. Piccoli, D. Vitulano, A fast computation method for time scale signal [124] M.N. Do, M. Vetterli, Contourlets: a directional multiresolution image representa-
denoising, Signal Image Video Process. 3 (1) (2009) 63–83. tion, in: International Conference on Image Processing Proceedings 2002, 1, 2002
[86] D.L. Donoho, J.M. Johnstone, Ideal spatial adaptation by wavelet shrinkage, I–I.
Biometrika 81 (3) (1994) 425–455. [125] R. Eslami, H. Radha, The contourlet transform for image denoising using cycle spin-
[87] D.L. Donoho, De-noising by soft-thresholding, IEEE Trans. Inf. Theory 41 (3) (1995) ning, in: Signals, Systems and Computers, 2004. Conference Record of the Thirty–
613–627. Seventh Asilomar Conference on, 2, 2003, pp. 1982–1986.
[88] S.G. Chang, B. Yu, M. Vetterli, Adaptive wavelet thresholding for image denoising [126] R. Eslami, H. Radha, Translation-invariant contourlet transform and its application
and compression, IEEE Trans. Image Process. 9 (9) (2000) 1532–1546. to image denoising, IEEE Trans. Image Process. 15 (11) (2006) 3362–3374.
[89] B. Vidakovic, Nonlinear wavelet shrinkage with Bayes rules and Bayes factors, J. [127] A.L. Da Cunha, J. Zhou, M.N. Do, The nonsubsampled contourlet transform: theory,
Am. Stat. Assoc. 93 (441) (1998) 173–179. design, and applications, IEEE Trans. Image Process. 15 (10) (2006) 3089–3101.
[90] A. Fathi, A.R. Naghsh-Nilchi, Efficient image denoising method based on a new [128] L. Huang, H. Wang, B. Zhu, Adaptive thresholds algorithm of image denoising based
adaptive wavelet packet thresholding function, IEEE Trans. Image Process. 21 (9) on nonsubsampled contourlet transform, in: Computer Science and Software Engi-
(2012) 3981–3990. neering, 2008 International Conference on, 6, 2008, pp. 209–212.

241
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

[129] M.F. Hossain, M.R. Alsharif, K. Yamashita, LMMSE-based image denoising in non- [162] A. Fathi, A.R. Naghsh-Nilchi, Efficient image denoising method based on a new
subsampled contourlet transform domain, in: International Conference on Image adaptive wavelet packet thresholding function, IEEE Trans. Image Process. 21 (9)
and Signal Processing, 2010, pp. 36–43. (2012) 3981–3990.
[130] S. Satheesh and K. Prasad, “Medical image denoising using adaptive threshold [163] A. Rangarajan, R. Chellappa, Markov random field models in image processing,
based on contourlet transform,” arXiv:1103.4907, 2011. in: M.A. Arbib (Ed.), The Handbook of Brain Theory and Neural Networks, 1995,
[131] J. Xu, L. Yang, D. Wu, Ripplet: a new transform for image processing, J. Vis. Com- pp. 564–567.
mun. Image Represent. 21 (7) (2010) 627–639. [164] Y. Cao, Y. Luo, S. Yang, Image denoising based on hierarchical Markov random
[132] D. Gupta, R.S. Anand, B. Tyagi, Ripplet domain non-linear filtering for speckle field, Pattern Recognit. Lett. 32 (2) (2011) 368–374.
reduction in ultrasound medical images, Biomed. Signal Process. Control 10 (2014) [165] S. Chen, M. Liu, W. Zhang, J. Liu, Edge preserving image denoising with a closed
79–91. form solution, Pattern Recognit. 46 (3) (2013) 976–988.
[133] E. Le Pennec, S. Mallat, Non-Linear Image Approximation with Bandelets, [166] P. Zhong, R. Wang, Multiple-spectral-band CRFs for denoising junk bands of hy-
CMAP/École Polytechnique, 2003 Tech. Rep. perspectral imagery, IEEE Trans. Geosci. Remote Sens. 51 (4) (2013) 2260–2275.
[134] H.B. Kekre, S.D. Thepade, J. Jain, N. Agrawal, Iris recognition using texture features [167] J. Ho, W.-L. Hwang, Image denoising using wavelet Bayesian network models, in:
extracted from haarlet pyramid, Int. J. Comput. Appl. 11 (12) (2010) 1–5. Acoustics, Speech and Signal Processing (ICASSP), 2012 IEEE International Con-
[135] F. Smeraldi, Ranklets: orientation selective non-parametric features applied to face ference on, 2012, pp. 1105–1108.
detection, in: Object recognition supported by user interaction for service robots, [168] J.M. Sanches, J.C. Nascimento, J.S. Marques, Medical image noise reduction us-
3, IEEE, 2002, pp. 379–382. ing the Sylvester–Lyapunov equation, IEEE Trans. Image Process. 17 (9) (2008)
[136] J.R. Kaplan, D.L. Donoho, The morphlet transform: a multiscale representation for 1522–1539.
diffeomorphisms, in: DEFORM’06-Proceedings of the Workshop on Image Registra- [169] X.T. Wang, G.M. Shi, Y. Niu, L. Zhang, Robust adaptive directional lifting wavelet
tion in Deformable Environments, 2006, p. 21. transform for image denoising, IET Image Process. 5 (3) (2011) 249–260.
[137] J. Krommweh, Tetrolet transform: a new adaptive haar wavelet algorithm for [170] J.K. Romberg, H. Choi, R.G. Baraniuk, Bayesian wavelet-domain image modeling
sparse image representation, J. Vis. Commun. Image Represent. 21 (4) (2010) using hidden Markov trees, in: Image Processing, 1999. ICIP 99. Proceedings. 1999
364–374. International Conference on, 1, 1999, pp. 158–162.
[138] http://www.laurent-duval.eu/siva-wits-where-is-the-starlet.html (Accessed on 23. [171] M. Malfait, D. Roose, Wavelet-based image denoising using a Markov random field
11November .2016). a priori model, IEEE Trans. Image Process. 6 (4) (1997) 549–565.
[139] F. Luisier, T. Blu, M. Unser, Undecimated Haar thresholding for Poisson intensity [172] C. Knaus, M. Zwicker, Dual-domain image denoising, in: 20th IEEE International
estimation, in: Image Processing (ICIP), 2010 17th IEEE International Conference Conference on Image Processing (ICIP), 2013, 2013, pp. 440–444.
on, 2010, pp. 1697–1700. [173] C. Knaus, M. Zwicker, Dual-domain filtering, SIAM J. Imaging Sci. 8 (3) (2015)
[140] V. Velisavljevic, B. Beferull-Lozano, M. Vetterli, P.L. Dragotti, Directionlets: 1396–1420.
anisotropic multidirectional representation with separable filtering, IEEE Trans. [174] M. Rais, Fast and Accurate Image Registration. Applications to On-Board Satellite
Image Process. 15 (7) (2006) 1916–1933. Imaging, Université Paris-Saclay, 2016.
[141] R. Sethunadh, T. Thomas, Spatially adaptive image denoising using inter-scale de- [175] Z. Huang, Y. Zhang, Q. Li, T. Zhang, N. Sang, H. Hong, Progressive dual-domain
pendence in directionlet domain, IET Image Process. 9 (2) (2014) 142–152. filter for enhancing and denoising optical remote-sensing images, IEEE Geosci. Re-
[142] J. Liu, Y. Wang, K. Su, W. He, Image denoising with multidirectional shrinkage in mote Sens. Lett. 15 (5) (2018) 759–763.
directionlet domain, Signal Process. 125 (2016) 64–78. [176] N. Wang, et al., A hybrid model for image denoising combining modified isotropic
[143] Q. Gao, Y. Lu, D. Sun, Z.-L. Sun, D. Zhang, Directionlet-based denoising of SAR diffusion model and modified Perona-Malik model, IEEE Access (2018).
images using a Cauchy model, Signal Process. 93 (5) (2013) 1056–1063. [177] H. Yu, L. Zhao, H. Wang, Image denoising using trivariate shrinkage filter in the
[144] G. Easley, D. Labate, W.-Q. Lim, Sparse directional image representations using the wavelet domain and joint bilateral filter in the spatial domain, IEEE Trans. Image
discrete shearlet transform, Appl. Comput. Harmon. Anal. 25 (1) (2008) 25–46. Process. 18 (10) (2009) 2364–2369.
[145] W.-Q. Lim, The discrete shearlet transform: a new directional transform and [178] H. Al-Marzouqi, G. AlRegib, Curvelet transform with learning-based tiling, Signal
compactly supported shearlet frames, IEEE Trans. Image Process. 19 (5) (2010) Process. 53 (2017) 24–39.
1166–1180. [179] C. Zhang, Y. Chen, C. Duanmu, Y. Yang, Image denoising by using PDE and GCV
[146] K. Guo, D. Labate, W.-Q. Lim, Edge analysis and identification using the continuous in tetrolet transform domain, Eng. Appl. Artif. Intell. 48 (2016) 204–229.
shearlet transform, Appl. Comput. Harmon. Anal. 27 (1) (2009) 24–46. [180] Y. Gan, E. Angelini, A. Laine, C. Hendon, BM3D-based ultrasound image denoising
[147] P. Jain, V. Tyagi, A survey of edge-preserving image denoising methods, Inf. Syst. via brushlet thresholding, in: Biomedical Imaging (ISBI), 2015 IEEE 12th Interna-
Front. 18 (1) (2016) 159–170. tional Symposium on, 2015, pp. 667–670.
[148] B. Goyal, A. Dogra, S. Agrawal, B.S. Sohi, Two-dimensional gray scale image de- [181] Y. Bai, Y. Liu, Q. Zhang, L. Jia, Z. Gui, Image denoising via an improved non-local
noising via morphological operations in NSST domain & bitonic filtering, Future total variation model, J. Eng. 2018 (8) (2018) 745–752.
Gener. Comput. Syst. 82 (2018) 158–175. [182] C. Sutour, C.-A. Deledalle, J.-F. Aujol, Adaptive regularization of the NL-means:
[149] B. Hou, X. Zhang, X. Bu, H. Feng, SAR image despeckling based on nonsubsampled application to image and video denoising, IEEE Trans. Image Process. 23 (8) (2014)
shearlet transform, IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 5 (3) (2012) 3506–3521.
809–823. [183] G.R. Easley, D. Labate, F. Colonna, Shearlet-based total variation diffusion for de-
[150] Y. Shi, X. Yang, Y. Guo, Translation invariant directional framelet transform com- noising, IEEE Trans. Image Process. 18 (2) (2009) 260–268.
bined with Gabor filters for image denoising, IEEE Trans. Image Process. 23 (1) [184] P.-L. Shui, Image denoising algorithm via doubly local wiener filtering with di-
(2014) 44–55. rectional windows in wavelet domain, IEEE Signal Process. Lett. 12 (10) (2005)
[151] K. Dabov, A. Foi, V. Katkovnik, K. Egiazarian, Image denoising by sparse 3-D 681–684.
transform-domain collaborative filtering, IEEE Trans. Image Process. 16 (8) (2007) [185] X. Wang, J. Gao, Image denoising method based on nonsubsampled contourlet
2080–2095. transform and bandelet transform, in: Information Science and Engineering (ICISE),
[152] J. Portilla, V. Strela, M.J. Wainwright, E.P. Simoncelli, Image denoising using scale 2009 1st International Conference on, 2009, pp. 1278–1281.
mixtures of Gaussians in the wavelet domain, IEEE Trans. Image Process. 12 (11) [186] S.K. Panigrahi, S. Gupta, P.K. Sahu, Curvelet-based multiscale denoising using
(2003) 1338–1351. non-local means & guided image filter, IET Image Process. 12 (6) (2018) 909–918.
[153] L. Zhang, W. Dong, D. Zhang, G. Shi, Two-stage image denoising by principal [187] A. Karami, R. Heylen, P. Scheunders, Band-specific shearlet-based hyperspectral
component analysis with local pixel grouping, Pattern Recognit. 43 (4) (2010) image noise reduction, IEEE Trans. Geosci. Remote Sens. 53 (9) (2015) 5054–5066.
1531–1549. [188] E.J. Candès, F. Guo, New multiscale transforms, minimum total variation synthe-
[154] L. Sendur, I.W. Selesnick, Bivariate shrinkage functions for wavelet-based denois- sis: applications to edge-preserving image reconstruction, Signal Process. 82 (11)
ing exploiting interscale dependency, IEEE Trans. Signal Process. 50 (11) (2002) (2002) 1519–1543.
2744–2756. [189] S. Durand, J. Froment, Reconstruction of wavelet coefficients using total variation
[155] L. Sendur, I.W. Selesnick, Bivariate shrinkage with local variance estimation, IEEE minimization, SIAM J. Sci. Comput. 24 (5) (2003) 1754–1767.
Signal Process. Lett. 9 (12) (2002) 438–441. [190] H.R. Shahdoosti, S.M. Hazavei, Combined ripplet and total variation image denois-
[156] P. Moulin, J. Liu, Analysis of multiresolution image denoising schemes using gen- ing methods using twin support vector machines, Multimedia Tools Appl. 77 (6)
eralized Gaussian and complexity priors, IEEE Trans. Inf. Theory 45 (3) (1999) (2018) 7013–7031.
909–919. [191] J. Ma, G. Plonka, Combined curvelet shrinkage and nonlinear anisotropic diffusion,
[157] K. Dabov, A. Foi, V. Katkovnik, K. Egiazarian, BM3D image denoising with IEEE Trans. Image Process. 16 (9) (2007) 2198–2206.
shape-adaptive principal component analysis, SPARS’09-Signal Processing with [192] G.Y. Chen, B. Kégl, Image denoising with complex ridgelets, Pattern Recognit. 40
Adaptive Sparse Structured Representations, 2009. (2) (2007) 578–585.
[158] Y. Hou, C. Zhao, D. Yang, Y. Cheng, Comments on image denoising by sparse 3-D [193] B.S. Kumar, Image denoising based on gaussian/bilateral filter and its method noise
transform-domain collaborative filtering, IEEE Trans. Image Process. 20 (1) (2011) thresholding, Signal Image Video Process. 6 (7) (2013) 1159–1172.
268–270. [194] B.S. Kumar, Image denoising based on non-local means filter and its method noise
[159] S. Yin, L. Cao, Y. Ling, G. Jin, Image denoising with anisotropic bivariate shrinkage, thresholding, Signal Image Video Process. 7 (6) (2013) 1211–1227.
Signal Process. 91 (8) (2011) 2078–2090. [195] G. Gao, Image denoising by non-subsampled shearlet domain multivariate model
[160] H. Rabbani, R. Nezafat, S. Gazor, Wavelet-domain medical image denoising us- and its method noise thresholding, Optik 124 (22) (2013) 5756–5760.
ing bivariate laplacian mixture model, IEEE Trans. Biomed. Eng. 56 (12) (2009) [196] P. Xu, Q. Miao, X. Tang, J. Zhang, A denoising algorithm via wiener filtering in the
2826–2837. shearlet domain, Multimedia Tools Appl. 71 (3) (2014) 1529–1558.
[161] G. Yu, G. Sapiro, S. Mallat, Solving inverse problems with piecewise linear esti- [197] M. Elad, M. Aharon, Image denoising via sparse and redundant representations
mators: from Gaussian mixture models to structured sparsity, IEEE Trans. Image over learned dictionaries, IEEE Trans. Image Process. 15 (12) (2006) 3736–
Process. 21 (5) (2012) 2481–2499. 3745.

242
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

[198] S. Mallat, Z. Zhang, Matching Pursuit With Time-Frequency Dictionaries, Courant [234] B. Goyal, S. Agrawal, B.S. Sohi, A. Dogra, Noise reduction in MR brain image via
Institute of Mathematical Sciences, New York, United States, 1993. various transform domain schemes, Asian J. Res. Chem. 9 (7) (2016) 919.
[199] Y.C. Pati, R. Rezaiifar, P.S. Krishnaprasad, Orthogonal matching pursuit: recursive [235] A. Dogra, B. Goyal, S. Agrawal, From multi-scale decomposition to non-multi-scale
function approximation with applications to wavelet decomposition, in: Confer- decomposition methods: a comprehensive survey of image fusion techniques and
ence Record of The Twenty-Seventh Asilomar Conference on Signals, Systems and its applications, IEEE Access 5 (2017) 16040–16067.
Computers, 1993, 1993, pp. 40–44. [236] J. Yadav, A. Dogra, B. Goyal, S. Agrawal, A review on image fusion methodologies
[200] S.S. Chen, D.L. Donoho, M.A. Saunders, Atomic decomposition by basis pursuit, and applications, Res. J. Pharm. Technol. 10 (4) (2017) 1239.
SIAM Rev. 43 (1) (2001) 129–159. [237] B. Goyal, A. Dogra, S. Agrawal, B.S. Sohi, Dual way residue noise thresholding
[201] P. Chatterjee, P. Milanfar, Clustering-based denoising with locally learned dictio- along with feature preservation, Pattern Recognit. Lett. 94 (2017) 194–201.
naries, IEEE Trans. Image Process. 18 (7) (2009) 1438–1451. [238] A. Dogra, S. Kadry, B. Goyal, S. Agrawal, An efficient image integration algorithm
[202] J. Mairal, F. Bach, J. Ponce, G. Sapiro, A. Zisserman, Non-local sparse models for for night mode vision applications, MultimedIS Tools Appl. (2018) 1–18.
image restoration, in: 12th International Conference on Computer Vision, 2009 [239] A. Dogra, B. Goyal, S. Agrawal, Current and future orientation of anatomical
IEEE, 2009, pp. 2272–2279. and functional imaging modality fusion, Biomed. Pharmacol. J. 10 (4) (2017)
[203] X. Zhu, P. Milanfar, Automatic parameter selection for denoising algorithms using a 1661–1663.
no-reference measure of image content, IEEE Trans. Image Process. 19 (12) (2010) [240] A. Dogra, M.S. Patterh, CT and MRI brain images registration for clinical applica-
3116–3132. tions, J. Cancer Sci. Ther. 6 (2014) 018–026.
[204] P. Bouboulis, K. Slavakis, S. Theodoridis, Adaptive kernel-based image denois- [241] A. Dogra, P. Bhalla, CT and mri brain images matching using ridgeness correlation,
ing employing semi-parametric regularization, IEEE Trans. Image Process. 19 (6) Biomed. Pharmacol. J. 7 (2) (2014) 20.
(2010) 1465–1479. [242] R. Kumar, 2010 (https://in.mathworks.com/matlabcentral/fileexchange/28112-
[205] P. Chatterjee, P. Milanfar, Is denoising dead? IEEE Trans. Image Process. 19 (4) diffusion-filtering-for-image-denoising).
(2010) 895–911. [243] K. Dabov, A. Foi, V. Katkovnik, K. Egiazarian, Joint image sharpening and denoising
[206] P. Chatterjee, P. Milanfar, Patch-based near-optimal image denoising, IEEE Trans. by 3D transform-domain collaborative filtering, in: Proc. 2007 Int. TICSP Workshop
Image Process. 21 (4) (2012) 1635. Spectral Meth. Multirate Signal Process., SMMSP, 2007, Citeseer, 2007.
[207] J. Zhang, D. Zhao, R. Xiong, S. Ma, W. Gao, Image restoration using joint statistical [244] H.R. Shahdoosti, Z. Rahemi, Edge-preserving image denoising using a deep convo-
modeling in a space-transform domain, IEEE Trans. Circuits Syst. Video Technol. lutional neural network, Signal Process. 159 (2019) 20–32.
24 (6) (2014) 915–928. [245] V. Vapnik, The support vector method of function estimation, in: Nonlinear Mod-
[208] H. Talebi, P. Milanfar, Global image denoising, IEEE Trans. Image Process. 23 (2) eling, Springer, Boston, MA, 1998, pp. 55–85.
(2014) 755–768. [246] X.Y. Wang, H.Y. Yang, Z.K. Fu, A new wavelet-based image denoising using un-
[209] R. Sethunadh, T. Thomas, Spatially adaptive image denoising using inter-scale de- decimated discrete wavelet transform and least squares support vector machine,
pendence in directionlet domain, IET Image Process. 9 (2) (2014) 142–152. Expert Syst. Appl. 37 (10) (2010) 7040–7049.
[210] H. Yue, X. Sun, J. Yang, F. Wu, Image denoising by exploring external and internal [247] X.Y. Wang, H.Y. Yang, Y. Zhang, Z.K. Fu, Image denoising using SVM classification
correlations, IEEE Trans. Image Process. 24 (6) (2015) 1967–1982. in nonsubsampled contourlet transform domain, Inf. Sci. 246 (2013) 155–176.
[211] G. Galiano, J. Velasco, On a cross-diffusion system arising in image denoising, Com- [248] E.J. Balster, Y.F. Zheng, R.L. Ewing, Feature-based wavelet shrinkage algorithm for
put. Math. Appl. 26 (5) (2018) 984–996. image denoising, IEEE Trans. Image Process. 14 (12) (2005) 2024–2039.
[212] V. Papyan, M. Elad, Multi-scale patch-based image restoration, IEEE Trans. Image [249] H.R. Shahdoosti, O. Khayat, Combination of anisotropic diffusion and non-subsam-
Process. 25 (1) (2016) 249–261. pled shearlet transform for image denoising, J. Intell. Fuzzy Syst. 30 (6) (2016)
[213] J. Feng, L. Song, X. Huo, X. Yang, W. Zhang, An optimized pixel-wise weighting 3087–3098.
approach for patch-based image denoising, IEEE Signal Process. Lett. 22 (1) (2015) [250] H.R. Shahdoosti, Z. Rahemi, A maximum likelihood filter using non-local informa-
115–119. tion for despeckling of ultrasound images, Mach. Vis. Appl. 29 (4) (2018) 689–702.
[214] G. Baloch, H. Ozkaramanli, R. Yu, Residual correlation regularization based image [251] H.R. Shahdoosti, S.M. Hazavei, A new compressive sensing based image denoising
denoising, IEEE Signal Process. Lett. 25 (2) (2018) 298–302. method using block-matching and sparse representations over learned dictionaries,
[215] H.R. Shahdoosti, O. Khayat, Image denoising using sparse representation classifi- Multimedia Tools Appl. 78 (9) (2019) 12561–12582.
cation and non-subsampled shearlet transform, Signal Image Video Process. 10 (6) [252] H.R. Shahdoosti, Two-stage image denoising considering interscale and intrascale
(2016) 1081–1087. dependencies, J. Electron. Imaging 26 (6) (2017) 063029.
[216] H. Ma, Y. Nie, An edge fusion scheme for image denoising based on anisotropic [253] H.R. Shahdoosti, S.M. Hazavei, Image denoising in dual contourlet domain using
diffusion models, J. Vis. Commun. Image Represent. 40 (2016) 406–417. hidden Markov tree models, Digit. Signal Process. 67 (2017) 17–29.
[217] R.J. Jevnisek, S. Avidan, Co-occurrence filter, in: CVPR, 2017, pp. 3816–3824. [254] S. Parameswaran, E. Luo, T.Q. Nguyen, Patch matching for image denoising using
[218] A. Wong, A. Mishra, W. Zhang, P. Fieguth, D.A. Clausi, Stochastic image denois- neighborhood-based collaborative filtering, IEEE Trans. Circuits Syst. Video Tech-
ing based on Markov-chain Monte Carlo sampling, Signal Process. 91 (8) (2011) nol. 28 (2) (2016) 392–401.
2112–2120. [255] A. Karami, L. Tafakori, Image denoising using generalised Cauchy filter, IET Image
[219] L. Xu, M. Shaifee, A. Wong, D. Clausi, Fully-connected continuous conditional ran- Process. 11 (9) (2017) 767–776.
dom field with stochastic cliques for dark-spot detection in SAR imagery, IEEE J. [256] W. Zhao, Y. Lv, Q. Liu, B. Qin, Detail-preserving image denoising via adaptive
Sel. Top. Appl. Earth Obs. Remote Sens. 9 (7) (2016) 2882–2890. clustering and progressive PCA thresholding, IEEE Access 6 (2017) 6303–6315.
[220] A. Wong, J. Scharcanski, Monte Carlo despeckling of transrectal ultrasound images [257] V. Fedorov, C. Ballester, Affine non-local means image denoising, IEEE Trans. Im-
of the prostate, Digit. Signal Process. 22 (5) (2012) 768–775. age Process. 26 (5) (2017) 2137–2148.
[221] F. Li, L. Xu, A. Wong, D.A. Clausi, QMCTLS: quasi Monte Carlo texture likelihood [258] M. Lebrun, M. Colom, J.-M. Morel, Multiscale image blind denoising, IEEE Trans.
sampling for despeckling of complex polarimetric SAR images, IEEE Geosci. Remote Image Process. 24 (10) (2015) 3149–3161.
Sens. Lett. 12 (7) (2015) 1566–1570. [259] Q. Guo, C. Zhang, Y. Zhang, H. Liu, An efficient SVD-based method for image de-
[222] L. Xu, F. Li, A. Wong, D.A. Clausi, Hyperspectral image denoising using a spa- noising, IEEE Trans. Circuits Syst. Video Technol. 26 (5) (2015) 868–880.
tial–spectral Monte Carlo sampling approach, IEEE J. Sel. Top. Appl. Earth Obs. [260] I. Frosio, J. Kautz, Statistical nearest neighbors for image denoising, IEEE Trans.
Remote Sens. 8 (6) (2015) 3025–3038. Image Process. 28 (2) (2018) 723–738.
[223] X. Yang, B. Fei, A wavelet multiscale denoising algorithm for magnetic resonance [261] H. Wang, S. Cao, K. Jiang, H. Wang, Q. Zhang, Seismic data denoising for complex
(MR) images, Meas. Sci. Technol. 22 (2) (2011) 025803. structure using BM3D and local similarity, J. Appl. Geophys. (2019).
[224] M. Kumar, M. Diwakar, CT image denoising using locally adaptive shrinkage rule [262] S. Chen, S. Chen, Z. Guo, Y. Zuo, Low-resolution palmprint image denoising by
in tetrolet domain, J. King Saud Univ.-Comput. Inf. Sci. 30 (1) (2018) 41–50. generative adversarial networks, Neurocomputing 358 (2019) 275–284.
[225] S. Jian, W. Wen, in: Study on underwater image denoising algorithm based on [263] S. Li, Y. Chen, R. Jiang, X. Tian, Image denoising via multi-scale gated fusion net-
wavelet transform, 806, 2017, p. 01. 2006. work, IEEE Access 7 (2019) 49392–49402.
[226] T.S. Sharmila, K. Ramar, Efficient analysis of hybrid directional lifting technique [264] X. Zheng, Y. Yuan, X. Lu, Hyperspectral image denoising by fusing the selected
for satellite image denoising, Signal Image Video Process. 8 (7) (2014) 1399–1404. related bands, IEEE Trans. Geosci. Remote Sens. 57 (5) (2018) 2596–2609.
[227] S. Liu, M. Liu, P. Li, J. Zhao, Z. Zhu, X. Wang, SAR image denoising via sparse [265] J. Liu, S. Osher, Block matching local SVD operator based sparsity and TV regular-
representation in shearlet domain based on continuous cycle spinning, IEEE Trans. ization for image denoising, J. Sci. Comput. 78 (1) (2019) 607–624.
Geosci. Remote Sens. 55 (5) (2017) 2985–2992. [266] H. Wang, Y. Li, Y. Cen, Z. He, L. Zhang, Multi-matrices low-rank decomposition
[228] Y. Shen, et al., Improved anscombe transformation and total variation for denoising with structural smoothness for image denoising, IEEE Trans. Circuits Syst. Video
of lowlight infrared images, Infrared Phys Technol 93 (2018) 192–198. Technol. (2019).
[229] https://www.mathworks.com/matlabcentral/fileexchange/67703-image- [267] W. Jifara, F. Jiang, S. Rho, M. Cheng, S. Liu, Medical image denoising using con-
processing-dataset-for-color-grey-image-fusion–image-blending–image- volutional neural network: a residual learning approach, J. Supercomput. 75 (2)
denoising–enhancement (accessed Accessed on 9. 10October. 2017) (2019) 704–718.
[230] A. Dogra, S. Agrawal, B. Goyal, Efficient representation of texture details in medical [268] G. Baloch, H. Ozkaramanli, R. Yu, Residual correlation regularization based image
images by fusion of Ripplet and DDCT transformed images, Trop. J. Pharm. Res. denoising, IEEE Signal Process. Lett. 25 (2) (2018) 298–302.
15 (9) (2016) 1983–1993. [269] M. Diwakar, M. Kumar, A review on CT image noise and its denoising, Biomed.
[231] A. Dogra, B. Goyal, S. Agrawal, Bone vessel image fusion via generalized Reisz Signal Process. Control 42 (2018) 73–88.
wavelet transform using averaging fusion rule, J. Comput. Sci. 21 (2017) 371–378. [270] Y. Weiss, W.T. Freeman, What makes a good model of natural images? in: 2007
[232] A. Dogra, S. Agrawal, B. Goyal, N. Khandelwal, C.K. Ahuja, Color and grey scale IEEE Conference on Computer Vision and Pattern Recognition, June, IEEE, 2007,
fusion of osseous and vascular information, J. Comput. Sci. 17 (2016) 103–114. pp. 1–8.
[233] A. Dogra, S. Agrawal, N. Khandelwal, C. Ahuja, Osseous and vascular information [271] Q. Liu, B. Xiong, M. Zhang, Adaptive sparse norm and nonlocal total variation
fusion using various spatial domain filters, Asian J. Res. Chem. 9 (7) (2016) 937. methods for image smoothing, Math. Probl. Eng. (2014) (2014).

243
B. Goyal, A. Dogra and S. Agrawal et al. Information Fusion 55 (2020) 220–244

[272] W. Dong, L. Zhang, G. Shi, X. Li, Nonlocally centralized sparse representation for [275] C.-A. Deledalle, J. Salmon, A.S. Dalalyan, Image denoising with patch based PCA:
image restoration, IEEE Trans. Image Process. 22 (4) (2012) 1620–1630. local versus global, in: BMVC, 81, 2011, pp. 425–455.
[273] H. Takeda, S. Farsiu, P. Milanfar, Deblurring using regularized locally adaptive [276] A. Chambolle, V. Caselles, D. Cremers, M. Novaga, T. Pock, An introduction to
kernel regression, IEEE Trans. Image Process. 17 (4) (2008) 550–563. total variation for image analysis, Theor. Found. Numer. Methods Sparse Recov. 9
[274] H.C. Burger, C.J. Schuler, S. Harmeling, Image denoising: can plain neural networks (263–340) (2010) 227.
compete with BM3D? in: 2012 IEEE Conference on Computer Vision and Pattern
Recognition, IEEE, 2012, pp. 2392–2399.

244