International Journal of Pure and Applied Mathematics

Volume 118 No. 20 2018, 923-933

ISSN: 1314-3395 (on-line version)
url: http://www.ijpam.eu
Special Issue
ijpam.eu

RAIN OR SNOW REMOVING IN A SINGLE COLOR IMAGE- A

NOVEL ALGORITHM
1
Archana H., 2 Dr.M. Priya 3 Dr.A. Muthukumaravel
1
M.Phil-CS Research Scholar, Department of MCA, BIHER, Chennai, Tamil Nadu, India
2
Associate professor,Department of MCA, BIHER, Chennai, Tamil Nadu, India
3
Dean-Faculty of Arts & Science, & HOD-Department of MCA, BIHER, Chennai, Tamil Nadu, India

ABSTRACT: the current end; we tend to style a 3-layer

hierarchical scheme. In the initial layer, an over
We propose a novel efficient
complete lexicon is trained and 3
methodology to get rid of rain or snow from a
classifications are distributed to classify the
single color image. In image processing, there
high-frequency part into rain/snow and non-
are two widespread techniques used
rain/snow parts within which some common
specifically, Image decomposition and
characteristics of rain/snow are used. The
dictionary learning that is employed to
second layer is the next important layer for the
decompose the image and extraction of the
combination of rain/snow detection and guided
image. At first, a mix of rain or snow is
filtering which is worked out on the rain/snow
detected and so when a guided filter is
part obtained within the initial layer. In the
employed to decompose the input image to a
third layer, the sensitivity of variance across
complementary pair: 1) the low-frequency part
color channels is manipulated to reinforce the
that's freed from rain or snow almost fully and
visual quality of rain/snow-removed image.
2) the high-frequency part that contains not
The effectiveness of our methodology is
solely the rain/snow part however additionally
verified through each subjective (the visual
some or perhaps several details of the image.
quality) and objective (through rendering
Then, we tend to target the extraction of
rain/snow on some ground-truth images)
image’s details from the high-frequency part to

923
International Journal of Pure and Applied Mathematics Special Issue

approaches, that shows a superiority over many of this, the objects will not be occluded
progressive works. completely but some blurring may appear.
Second, pixels with different intensities will be
INTRODUCTION:
affected by rain differently. When the pixels
IT is well known that a bad weather, e.g., haze, primary intensity is relatively low, rain will
rain, or snow, affects severely the quality of the enhance its intensity. When a high intensity
captured images or videos, which consequently pixel is affected by rain, its intensity will
degrades the performance of many image become lower.
processing and computer vision algorithms
LITERATURE SURVEY:
such as object detection, tracking, recognition,
and surveillance. A study by Garg et al. reveals Detection and removal of rain from videos
that rain and snow belong to the dynamic
Rain removal algorithms require certain
weather they contain constituent particles of
number of consecutive frames for reducing the
relatively large sizes so that they can be
visibility of rain from videos. These algorithms
captured easily by cameras. On the other hand,
are designed for the rain videos captured by the
haze belongs to the steady weather the particles
fixed camera. In this study, it is demonstrated
are much smaller in size and can hardly be
that using global motion compensation all the
filmed. As a result, rain or snow leads to
rain removal algorithms developed for fixed
complex pixel variations and obscures the
camera can be used for moving camera too. In
information that is conveyed in the image or
this study a novel, efficient and simple
video. Especially, the degradation on the
algorithm for detection and removal of rain
involved algorithms performance would be
from video using meteorological properties is
severe if the algorithm is based on some
proposed. Here meteorological properties are
features in the image or video. As compared to
used to separate the rain pixels from the non
the de haze problem where some excellent
rain pixels. The proposed algorithm achieves
solutions have been achieved (e.g., removing
good accuracy in spite of less number of
of rain or snow is much more challenging.
consecutive frames, which reduces the buffer
Though belonging to the dynamic weather
size and delay. It works only on the intensity
category, rain and snow still have some
plane that further reduces the complexity and
differences when appearing in the image or
execution time significantly. Quantitative and
video. First, rain is semi transparent. Because
qualitative analyses show that the proposed

924
International Journal of Pure and Applied Mathematics Special Issue

algorithm removes rain effectively under of the first Coupling Processes in the
constrained buffer size and delay in Equatorial Atmosphere (CPEA-I) project
comparison with most of the competing rain reveal a difference between clouds
removal algorithms. precipitating in the early afternoon and clouds
precipitating in the nighttime. In the early
Single image haze removal using dark
afternoon, the precipitating clouds were
channel prior
dominated by shallow convective types with
The quality of images taken outsides will be high rainfall rate at the surface. In the
impaired severely in bad weather conditions, nighttime, precipitating clouds were dominated
such as haze, mist, fog or rain. In this paper, to by stratiform types with small rainfall rate at
improve the visibility with only single hazy the surface. A diurnal variation of horizontal
image, a kind of haze removal algorithm is wind was observed over this area. The westerly
proposed. Firstly, the raw atmospheric in the lower troposphere and the easterly in the
transmission map is estimated by use of the middle troposphere began to be enhanced in
dark channel prior. And then, Fields of Experts the afternoon (1400–1700 LT). DSD
model is adopted to amend the raw atmosphere parameters were retrieved from VHF band
transmission map. Finally, the scene albedo is Doppler radar data. A modified gamma
restored based on the atmosphere scattering distribution was used to model DSD
model. Experiments on a variety of outdoor parameters. The shape parameter (μ) was larger
hazy images verify the feasibility and validity during stratiform precipitation than during
of our algorithm. shallow convective precipitation events, as
shown by previous studies. During stratiform
The distribution of raindrops with size
rain events on 5 May 2004, the median volume
The diurnal variability of raindrop size diameter (D0) was dominantly greater than 1
distribution (DSD) in precipitating clouds over mm, which is larger than D0 during shallow
Kototabang, West Sumatra, Indonesia (0.20°S, convective rain events. Results presented in
100.32°E), is studied using three types of this paper indicate that DSD has a diurnal cycle
Doppler radars, operated at VHF (47 MHz), over the mountainous region of Sumatra.
UHF (1.3 GHz), and X band (9.4 GHz)
Rain removal in video by combining
frequencies. Two precipitation events from 5 to
temporal and chromatic properties
6 May 2004 in the first observation campaign

925
International Journal of Pure and Applied Mathematics Special Issue

Removal of rain streaks in video is a to decompose the input image into a

challenging problem due to the random spatial complementary pair: 1) the low-frequency part
distribution and fast motion of rain. This paper that is free of rain or snow almost completely
presents a new rain removal algorithm that and 2) the high-frequency part that contains not
incorporates both temporal and chromatic only the rain/snow component but also some or
properties of rain in video. The temporal even many details of the image. Then, we focus
property states that an image pixel is never on the extraction of image's details from the
always covered by rain throughout the entire high-frequency part. To this end, we design a
video. The chromatic property states that the 3-layer hierarchical scheme. In the first layer,
changes of R, G, and B values of rain-affected an overcomplete dictionary is trained and three
pixels are approximately the same. By using classifications are carried out to classify the
both properties, the algorithm can detect and high-frequency part into rain/snow and non-
remove rain streaks in both stationary and rain/snow components in which some common
dynamic scenes taken by stationary cameras. characteristics of rain/snow have been utilized.
To handle videos taken by moving cameras, the In the second layer, another combination of
video can be stabilized for rain removal, and rain/snow detection and guided filtering is
destabilized to restore camera motion after rain performed on the rain/snow component
removal. It can handle both light rain and obtained in the first layer. In the third layer, the
heavy rain conditions. Experimental results sensitivity of variance across color channels is
show that the algorithm performs better than computed to enhance the visual quality of
existing algorithms rain/snow-removed image. The effectiveness of
our algorithm is verified through both
Spatio temporal frequency analysis for
subjective (the visual quality) and objective
removing rain and snow from videos
(through rendering rain/snow on some ground-
In this paper, we propose an efficient algorithm truth images) approaches, which shows a
to remove rain or snow from a single color superiority over several state-of-the-art works.
image. Our algorithm takes advantage of two
EXISTING SYSTEM:
popular techniques employed in image
processing, namely, image decomposition and As compared to the existing 2-step methods the
dictionary learning. At first, a combination of novelty of our proposed approach is two-fold.
rain/snow detection and a guided filter is used In the first step, instead of applying a low-pass

926
International Journal of Pure and Applied Mathematics Special Issue

filtering simply, we combine rain/snow The pipeline of our proposed rain/snow

detection together with a guided filter. removal algorithm consists of two steps. In the
first step, the input image is decomposed into
By doing this, we can achieve a much
the low frequency part and high-frequency
improved balance between removing rain/snow
part. Note that low frequency part is free of
components and preserving image’s details the
rain or snow almost completely but usually
resulted low-frequency part becomes free of
blurred, while high-frequency part contains
rain or snow almost completely and at the same
rain/snow components and some or even many
time contains the image’s details to a
details of the image. In the second step, we
reasonable extent.
design a 3-layer hierarchy of extracting non-
DISADVANTAGE: dynamic components from high-frequency part
First, rain/snow detection is performed
This method can produce good snow-removal
to produce a binary location map. Because the
results and keep acceptable image quality.
location map is binary, holes appear at the
However, when snow is heavy and has larger
rain/snow locations. Then, we fill each hole
size (e.g., the second image), it cannot
with the mean value of its neighboring non-
recognize snow.
rain/snow pixels. At last, a guided filter is
Another defect of this work is that it will utilized to generate the low-frequency and
mistreat small image details as snow, like in high-frequency part.
the fourth image. The last column presents our Finally to say simply, we detect
results, showing a much better snow-removal raindrops on the windshield in a single image,
performance. and then we modeled the geometric shape of
raindrops. After that we utilize the photometric
It can be seen that we have removed
property to construct a relationship between
successfully the majority of snow in images
raindrop and the environment. Later on, some
and maintained most image details at the same
learning-based image decomposition methods
time, leading to a better visual quality in the
were proposed to remove rain in a single
snow-removed images.
image. In the meantime, a guided filter based

PROPOSED SYSTEM: method was used by to remove rain or snow

from a single image by designing a rain/snow-
free guidance image.

927
International Journal of Pure and Applied Mathematics Special Issue

ADVANTAGE: take the more mundane view that it is a

programming language.
We have outlined several common
Matlab is a program that was originally
characteristics of rain and snow, from which
designed to simplify the implementation of
two metrics are defined, namely, the sensitivity
numerical linear algebra routines. It has since
of variance across color channels (SVCC) and
grown into something much bigger, and it is
the principal direction of an image patch
used to implement numerical algorithms for a
(PDIP).
wide range of applications. The basic language
A low-frequency part that is free of rain or used is very similar to standard linear algebra
snow almost completely has been generated, notation, but there are a few extensions that
thanks to the use of a combination of rain/snow will likely cause you some problems at first.
detection and a guided filter (as the low-pass
IMPLEMENTATION:
filter), while the corresponding high-frequency
Rain and snow removal
part is made complementary to the low
Algorithms mentioned above deal with the rain
frequency part.
and snow removing problem in videos.
A 3-layer hierarchy of extracting image’s Nevertheless, rain and snow removal from a
details from the high-frequency part has been single image seems more useful in practice, but
designed. Specifically, the first layer is a 3- also more challenging.
times classification that is based on a trained To the best of our knowledge, for the first time,
dictionary the second layer applies another detected raindrops on the windshield in a single
combination of rain/snow detection and a image, modeled the geometric shape of
guided filter and the third layer utilizes the raindrops, and utilized the photometric
SVCC to enhance the visual quality of the property to construct a relationship between
rain/snow-removed image. raindrop and the environment.

SOFTWARE ENVIRONMENT: Later on, some learning-based image

MATLAB decomposition methods were proposed to
MATLAB stands for Matrix Laboratory. remove rain in a single image.
According to The Mathworks, its producer, it is
a "technical computing environment". We will

928
International Journal of Pure and Applied Mathematics Special Issue

Image decomposition Dictionary learning

Because the location of rain/snow in the image The on-line dictionary learning method to learn
is random it is difficult to accurately separate an over-complete dictionary for factorizing IH.
rain/snow with other non rain snow This method addresses the factorization
components by normal detection methods. problem with a new on-line optimization
Dictionary learning is an excellent image algorithm that is based on stochastic
decomposition method, which can decompose approximation.
an image into many components.
We choose this method because it suits for both
Some are rain/snow components and the other small and large data set, and can train adaptive
is non-rain/snow components. In this over-complete dictionaries iteratively.
subsection, we try to represent IH by a sparse
On the other hand, K-SVD can do a similar job
coding that is based on learning an over-
but is slower in speed; whereas MCA utilizes a
complete we choose the on-line dictionary
different transform to construct an over-
learning method to learn an over-complete
complete dictionary that is not adaptive. In the
dictionary for factorizing.
following, we introduce this on-line dictionary
This method addresses the factorization learning method briefly.
problem with a new on-line optimization
Guided filtering
algorithm that is based on stochastic
approximation. We choose this method because This method addresses the factorization
it suits for both small and large data set, and problem with a new on-line optimization
can train adaptive over-complete dictionaries algorithm that is based on stochastic
iteratively. approximation.

929
International Journal of Pure and Applied Mathematics Special Issue

We choose this method because it suits for both combination of rain snow detection and a
small and large data set, and can train adaptive guided filtering. Finally, we have presented a
over-complete dictionaries iteratively. large set of results to show that our method can
remove rain or snow from images effectively,
On the other hand, K-SVD can do a similar job
leading to an enhanced visual quality in the
but is slower in speed; whereas MCA utilizes a
rain snow removed images.
different transform to construct an over-
complete dictionary that is not adaptive. REFERENCES:
CONCLUSION:
[1] K. Garg and S. K. Nayar, “Detection and
This paper has attempted to solve the removal of rain from videos,” IEEE
rain/snow removing problem from a single Conference on Computer Vision and Pattern
color image by utilizing the common Recognition (CVPR- 2004), pp. 528-535,
characteristics of rain and snow. To this end, Washington DC, USA, June 27-July 2, 2004.
we defined the principal direction of an image [2] K. He, J. Sun and X. Tang, “Single image
patch (PDIP) and the sensitivity of variance of haze removal using dark channel prior,” IEEE
color channel (SVCC) to describe the Transactions on Pattern Analysis and Machine
difference of rain or snow from other image Intelligence, vol. 33, no. 12, pp. 2341-2353,
components. We acquired the low and high Dec. 2011.
frequency parts by implementing rain snow [3] J. S. Marshall and W. Mc K.Palmer, “The
detection and applying a guided filter. For the distribution of raindrops with size,” Journal of
high frequency part, dictionary learning and the Atmospheric Sciences, vol. 5, no. 4, pp.
three classifications of dictionary atoms are 165-166, 1948.
implemented to decompose it into non- [4] S. K. Nayar and S. G. Narasimhan, “Vision
dynamic components and dynamic (rain or in bad weather,” IEEE International
snow) components, where some common Conference on Computer Vision (ICCV-1999),
characteristics of rain snow defined earlier in vol. 2, pp. 820-827, Kerkyra, Greece, Sep. 20-
our work are utilized. 27, 1999.
[5] K. Garg and S. K. Nayar, “Photorealistic
Moreover, we have designed two
rendering of rain streaks,” ACM Transactions
additional layers of extracting image details
on Graphics, vol. 25, no. 3, pp. 996-1002, July
from the high frequency part, which are based
2006.
on, respectively, the SVCC map and another

930
International Journal of Pure and Applied Mathematics Special Issue

[6] K.S.Priyanka, G.Ravikumar, “Fake Workshop on Structural, Syntactic, and

Biometric Detection Applied to Iris, Statistical Pattern Recognition, vol. 5342, pp.
Fingerprint, and Face Recognition by Using 451-458, Olando, USA, Dec. 2008.
Image Quality Assessment”, International [12] M. Roser and A. Geiger, “Video-based
Journal of Innovations in Scientific and raindrop detection for mproved image
Engineering Research (IJISER), Vol.2, No.3, registration,” IEEE International Conference
pp.57-72, 2015. on Computer Vision Workshops (ICCV
[7] X. Zhang, H. Li, Y. Qi, W. K. Leow, and T. Workshops 2009), pp. 570-577, Kyoto, Japan,
K. Ng, “Rain removal in video by combining Sept. 29-Oct. 2, 2009.
temporal and chromatic properties,” IEEE [13] J. Bossu, N. Hautiere, and J. P. Tarel,
International Conference on Multimedia and “Rain or snow detection in image sequences
Expo (ICME-2006), pp. 461-464, Toronto, through use of a histogram of orientation of
Ontario, Canada, July 9-12, 2006. streaks,” International Journal of Computer
[8] K. Garg and S. K. Nayar, “Vision and rain,” Vision, vol. 93, no. 3, pp. 348-367, July 2011.
International Journal of Computer Vision, vol. [14] J. C. Halimeh and M. Roser, “Raindrop
75, no. 1, pp. 3-27, 2007. detection on car windshields using geometric-
[9] P. Barnum, T. Kanade, and S. Narasimhan, photometric environment construction and
“Spatio-temporal frequency analysis for intensitybased correlation,” Intelligent Vehicles
removing rain and snow from videos,” Symposium, 2009, Xi’an, China, June 3-5,
International Workshop on Photometric 2009.
Analysis For Computer Vision (PACV-2007), [15] Y. H. Fu, L. W. Kang, C. W. Lin, and C.
Rio de Janeiro, Brazil, Oct. 2007. T. Hsu, “Single-frame-based rain removal via
[10] P. C. Barnum, S. Narasimhan, and T. image decomposition,” IEEE International
Kanade, “Analysis of rain and snow in Conference on Acoustics, Speech and Signal
frequency space,” International Journal of Processing (ICASSP-2011), pp. 1453- 1456,
Computer Vision, vol. 86, no. 2, pp. 256-274, Prague, Czech Republic, May 22-27, 2011.
2010. [16] L. W. Kang, C. W. Lin, and Y. H. Fu,
[11] N. Brewer and N. Liu, “Using the shape “Automatic single-image-based rain streaks
characteristics of rain to identify and remove removal via image decomposition,” IEEE
rain from video,” Joint IAPR International Transactions on Image Processing, vol. 21, no.
4, pp. 1742-1755, April 2012.

931
International Journal of Pure and Applied Mathematics Special Issue

932
933
934