IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 10, No. 2, June 2021, pp. 389~397

ISSN: 2252-8938, DOI: 10.11591/ijai.v10.i2.pp389-397 389

Multi-scale fusion for underwater image enhancement using

multi-layer perceptron

M. Sudhakara, M. Janaki Meena

Department of SCOPE, VIT University, Chennai, India

Article Info ABSTRACT

Article history: Underwater image enhancement (UIE) is an imperative computer vision
activity with many applications and different strategies proposed in recent
Received Oct 27, 2020 years. Underwater images are firmly low in quality by a mixture of noise,
Revised Jan 12, 2021 wavelength dependency, and light attenuation. This paper depicts an
Accepted Apr 3, 2021 effective strategy to improve the quality of degraded underwater images.
Existing methods for dehazing in the literature considering dark channel
prior utilize two separate phases for evaluating the transmission map (i.e.,
Keywords: transmission estimation and transmission refinement). Accurate restoration is
not possible with these methods and takes more computational time. A
Gamma correction proposed three-step method is an imaging approach that does not need
Image dehazing particular hardware or underwater conditions. First, we utilize the multi-layer
Image fusion perceptron (MLP) to comprehensively evaluate transmission maps by base
Multi-layer perceptron channel, followed by contrast enhancement. Furthermore, a gamma-adjusted
Multi-scale fusion version of the MLP recovered image is derived. Finally, the multi-scale
fusion method was applied to two attained images. The standardized weight
is computed for the two images with three different weights in the fusion
process. The quantitative results show that significantly our approach gives
the better result with the difference of 0.536, 2.185, and 1.272 for PCQI,
UCIQE, and UIQM metrics, respectively, on a single underwater image
benchmark dataset. The qualitative results also give better results compared
with the state-of-the-art techniques.
This is an open access article under the CC BY-SA license.

Corresponding Author:
M. Sudhakara
Department of SCOPE
VIT University, Chennai Campus, India
Email: [email protected]

1. INTRODUCTION
Underwater image enhancement in the recent decade attained great popularity in processing images
and underwater vision [1]. It is hard to enhance an underwater image because of the typical underwater
conditions and lighting effects [2], [3]. The color distortion influences underwater situations, and contrast
degrades because of absorption [4]. These artifacts produce underwater images with less quality. Many
researchers have investigated underwater images to attain the best quality. The methods for the improvement
of underwater images are classified into three categories. The first category applies the image degradation
model [5] to enhance an underwater image by exploring the physical process of natural light underwater to
build an underwater image degradation pattern. The second category improves image contrast by utilizing
thresholding and histogram equalization to map the grey values into a uniform distribution for adjusting the bright
or dark portions. There exist various enhanced methods based on histogram equalization [6]. Ancuti et al. [7], [8]
focused fusion mechanism on underwater image and video, which comes under the third category.

Journal homepage: http://ijai.iaescore.com

390 ISSN: 2252-8938

Dehazing strategies, dark channel priors (DCP) [9], [10] are applied to improve quality. In [11], the
authors presented a model to increase underwater image quality, and DCP helps estimate the transmission. A red
channel model [12] recovers minimum wavelength colors during the restoration of skipped contrast. In [13],
underwater DCP was proposed for image restoration. In [14], contrast limited adaptive histogram
equalization is utilized to improve the picture quality at some rate, and the histogram of the preferred portion
relatively matches the particular histogram. For multiple image development modes, a technique requires two
polarization images with varying degrees [15], and the other approach uses many images to increase the
underwater image clarity. A fusion-based mechanism [16] derived from adding multiple filters over input,
and a fusion mechanism [17] applies the techniques such as white balancing and global contrast stretching for
UIE. In [18], the authors proposed another fusion-based method to increase contrast and color, including the
inclusion of two images from the color compensated and white-adjusted adoption of the original image. A
two-phase method [19] proposes a colour correction mechanism of piece-wise transformation to restore the
color. Afterwards, an ideal technique is utilized for contrast improvement. Drews et al. [20] developed upon
DCP and considered it a significant source of visual data underwater in a blue color channel and green color
channel. The underwater DCP of those two channels estimates better transmission of underwater images than
the traditional DCP. Emberton et al. [21] developed a hierarchical rank-based technique by considering some
features to detect image portions with the right quality image. The further fusion of these two HR images
took place to get the final result. In [22], the authors proposed a Water-Net network, training on the UIEB
data with convolution neural networks (CNNs). The fusion-based approaches give a better output than the
existing dehazing techniques by utilizing a few weight maps to the low-quality images [23].
The primary objective of this work is to increase the quality of the underwater images. A wide
variety of strategies are available for the mentioned requirement, each with its inherent advantages and
disadvantages. Current state-of-the-art techniques utilize the dark channel prior based methods and many
other methods using multiple images for processing. These techniques require a lot of computational time
with less accurate restoration. In this regard, the authors have perceived a strong need to develop a suitable
imaging methodology underwater. This paper introduces a novel procedure to eliminate the haze present in
the underwater images captured by conventional cameras. The blueprint of the proposed work has appeared
in Figure 1. The crucial steps of our proposed methodology are described below.
− A refined transmission map is estimated directly from the minimum channel to recover the image, and
the multi-layer perceptron was used for training.
− Two variants of the recovered image with improving quality are produced by applying the contrast
stretching and gamma correction algorithms.
− Multi-scale image fusion is applied to these two images alongside their weight maps to promote the
transfer of the edges and colour contrast to the output image.
− The dominant colours are balanced by adapting a colour correction algorithm to the fused image.

Figure 1. Proposed outline for UIE

Terminology
The background concepts related to the underwater image, such as scattering and multi-layer
perceptron, are discussed in this section.
− Scattering model
An atmospheric scattering model is a mathematical model developed around the crude dynamical
conditions that oversee underwater climatic movements-the atmospheric scattering model by [24] for image
formation given in (1).

𝐼 (𝑝, 𝑞) = 𝐽(𝑝, 𝑞) ∗ 𝑡(𝑝, 𝑞) + 𝐴[1 − 𝑡 (𝑝, 𝑞)] (1)

Int J Artif Intell, Vol. 10, No. 2, June 2021: 389 – 397
Int J Artif Intell ISSN: 2252-8938 391

Here, 'I' is the original image, ‘J’, ‘A’,‘t’ are the real scene without impact, environmental light and
transmission map separately. The transmission map is assessed as (2).

𝑡 (𝑝 , 𝑞) = 𝑒 −𝛽∗𝑑(𝑝,𝑞) (2)

In (2), the 'β' is the scattering coefficient, and d (p, q) is scene depth. As shown in (1), modified as
(3) to acquire a haze-free image.

𝐼(𝑝,𝑞)−𝐴
J (p, q) = +𝐴
𝑡(𝑝,𝑞)

It is recognizable that in (3), the original image contains the real scene, transmission map, and
atmospheric light. The job here is recovering J (p, q) from the original image by eliminating the transmission
and atmospheric light. The intensity of the dark channel proposed by DCP for a given image I (p, q) is (4).

min
𝐼𝑑𝑎𝑟𝑘 (𝑝, 𝑞)= 𝐶∈𝑅,𝐺,𝐵
𝑚𝑖𝑛
[ 𝑧∈Ω(p,q) 𝐼𝐶 (𝑧) (4)

Here Ω (p, q) is a part of m X m, centered over (p, q), IC the colour of the channel of the image, and ‘z’
denotes pixel index in Ω(p, q) domain. 𝑡̂ (𝑝, 𝑞) can be computed as (5).

𝐼 𝑑𝑎𝑟𝑘(𝑝,𝑞)
𝑡̂ (𝑝, 𝑞)=1-ω [ 𝐴
] (5)

In (5), 'ω' defines the estimated restoration levels, while 'I' is the maximum possible value. Here, ‘A’ is
computed as (6).
𝑎𝑟𝑔 𝑚𝑎𝑥 𝑑𝑎𝑟𝑘
A=max ∑3𝐶=1 𝐼𝐶 { (𝑝,𝑞)∈(0.1%∗ℎ∗𝑤) [𝐼 (𝑝, 𝑞)]} (6)

In the above equation, 'w' is the width, 'h' is the height of 𝐼𝑑𝑎𝑟𝑘 (p, q) and 'C' represents every RGB
channel of an input image I (p, q). As shown in (3) helps in estimating an image with no affection J (p, q).
The DCP yields precise estimations of ‘A’ and t(p, q). The utilization of windows of m×m in the computation
of 𝑡̂ (p, q) creates artefacts over edges of the recovered image J(p, q).
− Multi-layer perceptron
Artificial neural network (ANN) is a machine learning approach, a profoundly simplified scientific
model from biological neural networks. The connection between the processing elements (neurons) is called
a network. The inputs Xi multiplied by a weight Wij, and the bias bj added to develop the perceptron. This
function illustrated as (7).

Yj = f [∑𝑖(𝑊𝑖𝑗 ∗ 𝑋𝑖 + 𝑏𝑗 )] (7)

To activate the perceptron, the activation function is utilized. In many networks, the tangent function
is used for activation, and it is given as in (8).

Tanh(x) = 2*σ (2x)-1 (8)

where

𝑒𝑥
σ(x) =1+𝑒 𝑥 (9)

As shown in (9) is the sigmoid activation function [25]. The preparation of ANN changes the weight
and bias for getting the ideal outcome dependent on the different input combinations. Different perceptrons in
one or many hidden layer topologies are connected; this system is known as MLP. Backpropagation is a well-
known strategy for MLP, and it is a generalization of the least-squares component where the loads are altered
dependent on mean square error (MSE). The idea is finding the smallest error in relationship with the
associated weights and is given in (10).
1 2
E=2 ∑𝑗∈𝑀(𝑦̂
𝑚 − 𝑦𝑚 ) (10)

Multi-scale fusion for underwater image enhancement using multi-layer perceptron (M. Sudhakara)
392 ISSN: 2252-8938

2. PROPOSED METHODOLOGY
The proposed work is mainly a three-step process. First, the input image is acquired from the
underwater. Second, an MLP is utilized to recover the marine image. For the MLP retrieved image, we used
two contrast enhancement techniques to improve the contrast (i.e., contrast stretching, gamma correction).
Before applying the contrast stretching, the RGB version of the MLP recovered image is altered to the
L*A*B model, and an algorithm is used on the luminance channel of the L*A*B image. Later multi-scale
fusion is applied to the derived images to acquire the result. Further, simple colour balancing is used to
balance the colour before reaching the final output image.
The process of MLP on the given input image is represented in Figure 2. For a given input image,
air-light is estimated. It is nothing but applying the morphological erosion operation on the input image and
obtaining the dark channel map. The methods belong under the DCP mechanism yielded good results when
compared with existing models. As per DCP, the transmission map is estimated under two phases. Initially, a
coarse transmission map is estimated through square-patches with the user-defined size. The transmission
map is enhanced with a different filter and soft matting, and median filter in the next step. According to [26],
the transmission map is updated, as shown in (11), (12).

t min (p, q) = 1- ω Imin (p, q) (11)

min 𝑐
𝐼 (𝑝,𝑞)
Imin (p, q) = (12)
𝑐∈(𝑅,𝐺,𝐵) 𝐴𝑐

The tmin is equal to the dark channel only when Ω = 1. tmin (p, q) has a decent resolution. MLP for
estimating a transmission map t'(p, q) is (13).

𝑡′(𝑝, 𝑞) = 𝑀𝐿𝑃 [𝑡𝑚𝑖𝑛 (𝑝, 𝑞)] (13)

Figure 2. Dehazing process using MLP

The primary thought of this strategy is estimating the refined transmission map. The recovered
image is converted from RGB to L*a*b format, and the contrast of the luminance is increased. Once air-light
is estimated, we estimated the transmission map using the input image, air-light. Further, we applied a multi-
layer perceptron on the transmission estimated image to get the MLP recovered image. Then two images are
derived from the MLP improved model. The first input image is a gamma-corrected version of the MLP
recovered image. Gamma correction expects to improve the global contrast in the picture, and it is
appropriate to our work since many marine photos look too bright. Hence, the correction will increase the
contrast between the darker regions by reducing the more elegant areas. As mentioned earlier, contrast
stretching is applied to the MLP recovered image on the Luminance channel to get the second image for the
fusion process. Then we utilized the multi-scale fusion to get a solution for the single underwater image. The
weight maps are calculated based on the saliency metrics in such a way that the highest weights are
represented in the image (i.e., Laplacian weight (W L), saliency weight (WS), and saturation weight (Wsat)).
Subsequently, Laplacian Filter is applied to each luminance channel, and W L estimates the global contrast by
computing the total value. This weight allocates high costs to edges, and textures. This weight is not
satisfactory because it cannot differentiate among the different regions (i.e., ramp, flat). W L highlights the
objects that miss their importance in the input image and is derived from [27]. W sat adopts the chromatic
information from the image, which is advantageous to the fusion process. Saturation weight is nothing but the
deviation between each colour channel in the RGB image and is computed, as shown in (14).

1
Wsat =√3 [ (𝑅𝑘 − 𝐿𝑘 )2 + (𝐺𝑘 − 𝐿𝑘 )2 + (𝐵𝑘 − 𝐿𝑘 )2 ] (14)

Int J Artif Intell, Vol. 10, No. 2, June 2021: 389 – 397
Int J Artif Intell ISSN: 2252-8938 393

For every input, these three weight maps are combined into a single weight map, and the weights are
normalized, and finally, a normalized weight map is obtained. With these standardized weight maps, we
utilized multi-scale fusion to fuse the derived images. According to this, the Laplacian and Gaussian
pyramids are constructed and combining with Laplacian and Gaussian normalized weights achieved
independently at every level l, as shown in (15).

Rl (x) = ∑𝐾 ̅̅̅̅
𝑘=1 𝐺𝑙 {𝑊𝑘 (𝑥) } 𝐿𝑙 { 𝐼𝑘 (𝑥)} (15)

Here, 'l' signifies the pyramid levels, and 'k' symbolizes input images. The pyramid levels may be varied
depending upon the size of the picture. By combining all the pyramid level images with the proper
upsampling, we get the final fused image.

3. RESULTS AND DISCUSSION

The proposed method was evaluated on real-time underwater images used as a benchmark dataset
for the single UIE methods. For a multi-scale fusion, the levels depend on the size of the input image. The six
underwater images with different backgrounds are taken and applied our multi-layer perceptron based multi
fusion strategy. The input image, recovered image 1, recovered image 2, and the fusion of these dual images
are shown in Figure 3. The experimental outcomes are shown using the human visual system (HVS) and
automatic evaluation measures. As per the best of our knowledge, generating ground-truth images for
underwater images is so difficult. Due to the lack of this ground-truth information, HVS by specialists helps
evaluate the quality of the image by incorporating the perceptual metrics. We compared our method with four
approaches used for dehazing underwater and outdoor images. We have tested our result with many images,
but only four image results are shown in Figure 4. In this work, three quantitative measures are considered
(PCQI, UCIQE, and UIQM). Among these metrics, PCQI is a universal metric, and the other two parameters
UCIQE, UIQM [28], are the metrics for evaluating the underwater images. PCQI depends on the patch-based
model depends on global statistics. It considers three quantities of the image which are not dependent (mean,
signal strength, and structure). Mathematically PCQI can be computing using (16).

𝑃𝐶𝑄𝐼 = 𝑄𝑖(𝑝, 𝑞) ∗ 𝑄𝑐(𝑝, 𝑞) ∗ 𝑄𝑠(𝑝, 𝑞) (16)

In (16), 𝑄𝑖(𝑝, 𝑞), 𝑄𝑐(𝑥, 𝑦), 𝑄𝑠(𝑥, 𝑦) are the contrast mean intensity, structural distortion, and
contrast change. This metric is computationally less costly compared with other evaluation metrics. UCIQE
on three attributes such as Chroma, contrast, and saturation of CIELab and evaluated by using (17).

𝑈𝐶𝐼𝑄𝐸 = 𝐾1 × 𝜎𝑐 + 𝐾2 × 𝑐𝑜𝑛𝑡𝑟𝑎𝑠𝑡 + 𝐾3 × µ𝑠 (17)

‘σc’ is the standard deviation of Chroma, and ‘µs’ is the mean of the saturation. The acronym of
UIQM is an underwater image quality measure and is dissimilar from predefined computation measures.
UIQM computed by considering three attributes colorfulness, contrast, and sharpness. It can be evaluated by
(18).

𝑈𝐼𝑄𝑀 = 𝑘1 × 𝑈𝐼𝐶𝑀 + 𝑘2 × 𝑈𝐼𝑆𝑀 + 𝑙3 × 𝑈𝐼𝐶𝑜𝑛𝑀 (18)

In (18), k1, k2, and k3 are the application dependent parameters; for example, extra weight must be
put on k1 for submerged colour correct and k2 to increase the perceptibility. The comparison metrics are
shown in Table 1, and the parameters are based on (16), (17), and (18). The larger values correspond to the
better result.

Table 1. Comparison results with the proposed method (continue)

Author\Image Shipwreck Fish Reef1 Reef2 Reef3 Galdran1 Ancuti1 Ancuti2 Ancuti3 Average
He, et al. PCQI 1.012 1.023 1.000 0.774 1.022 1.056 0.860 0.649 1.071 0.940
[32] UCIQE 0.565 0.602 0.612 0.702 0.606 0.593 0.485 0.456 0.577 0.577
UIQM 0.565 0.509 0.592 0.749 0.578 0.578 0.353 0.437 0.538 0.480
Drew-Jr PCQI 0.649 0.863 1.046 0.483 0.793 0.749 0.909 0.475 0.973 0.771
[13] UCIQE 0.550 0.623 0.649 0.659 0.620 0.544 0.499 0.492 0.535 0.574
UIQM 0.492 0.571 0.657 0.653 0.584 0.519 0.383 0.344 0.492 0.521
Galdren, PCQI 0.920 0.835 0.794 0.769 0.883 0.507 0.962 0.591 1.021 0.809
et.al. [12] UCIQE 0.646 0.527 0.576 0.633 0.533 0.529 0.641 0.529 0.614 0.580
UIQM 0.605 0.528 0.565 0.671 0.524 0.569 0.458 0.525 0.646 0.565

Multi-scale fusion for underwater image enhancement using multi-layer perceptron (M. Sudhakara)
394 ISSN: 2252-8938

Table 1. Comparison results with the proposed method

Author\Image Shipwreck Fish Reef1 Reef2 Reef3 Galdran1 Ancuti1 Ancuti2 Ancuti3 Average
Emberten, PCQI 0.945 1.156 1.078 0.607 0.943 1.147 1.036 0.603 1.129 0.960
et al. [21] UCIQE 0.632 0.705 0.660 0.718 0.678 0.652 0.499 0.529 0.555 0.625
UIQM 0.588 0.759 0.690 0.757 0.677 0.664 0.407 0.425 0.563 0.614
Ancuti, et PCQI 1.131 1.089 0.978 0.893 1.191 1.125 1.074 1.015 1.171 1.195
al. [33] UCIQE 0.634 0.669 0.655 0.718 0.705 0.643 0.588 0.590 0.652 0.650
UIQM 0.629 0.598 0.674 0.733 0.737 0.669 0.547 0.683 0.693 0.662
Ancuti, et PCQI 1.172 1.117 1.083 1.075 1.276 1.152 1.022 0.914 1.207 1.113
al. [18] UCIQE 0.632 0.667 0.658 0.711 0.697 0.659 0.594 0.592 0.664 0.652
UIQM 0.668 0.624 0.687 0.781 0.766 0.680 0.507 0.687 0.651 0.672
Our result PCQI 1.301 1.280 2.573 2.358 1.849 1.900 1.214 1.121 1.246 1.649
UCIQE 4.859 2.891 3.703 4.415 3.466 1.860 1.315 2.158 0.870 2.837
UIQM 2.210 2.403 0.995 2.387 2.345 3.289 0.086 1.606 2.176 1.944

Input Recovered Image 1 Recovered Image 2 Fused Image

Figure 3, The results of multi-layer perceptron based multi fusion strategy. Recovery image 1 is a gamma-
corrected version of MLP improved image. Recovery image 2 is a contrast stretching of MLP improved
image. The fused image is the balanced colour image of multi-scale fusion approach.

Int J Artif Intell, Vol. 10, No. 2, June 2021: 389 – 397
Int J Artif Intell ISSN: 2252-8938 395

Ref/ Ancuti1 Shipwreck Fish Dehaze3

Image

Input
Image

[29]

[30]

[31]

[18]

Our
Method

Figure 4. Evaluation of various underwater dehazing techniques. Chiang et al. [29], Fu et al. [30],
Salzara et al, [31], Ancuti et al. [18].

4. CONCLUSION
Image enhancement is the best approach to increase the quality of the submerged images distorted
by limited light range, blurs, and reduced contrast. This paper introduces a multi-scale fusion strategy by
utilizing the multi-layer perceptron to improve the underwater image quality. A single image dehazing
technique works based on the fusion and needn't bother about extra data. Firstly, the multi-layer perceptron is
used to dehaze an image from the originally captured image. Then fusion technique is applied to the two
images derived from contrast stretching and gamma corrected version of the images. The two images are
mixed with a multi-scale fusion procedure with the standardized weight, followed by a color balancing
strategy to adjust the color. The upgraded submerged images progressively more visible, interpretable with
decent clarity. Though this method uses synthetic images using MLP, it took less computation time, and
qualitative observations of this approach show better outcomes compared with the current UIE techniques.
The quantitative results show that significantly our approach gives the better result with the difference of
0.536, 2.185, and 1.272 for PCQI, UCIQE, and UIQM metrics, respectively, on the ancuti dataset. However,
the method is to be improved to develop a single UIE technique that can make applicable to all sorts of
distorted images underwater.

Multi-scale fusion for underwater image enhancement using multi-layer perceptron (M. Sudhakara)
396 ISSN: 2252-8938

REFERENCES
[1] Islam, Md Jahidul, Youya Xia, and Junaed Sattar, “Fast underwater image enhancement for improved visual
perception,” IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 3227-3234, 2020. doi:
10.1109/LRA.2020.2974710.
[2] Sudhakar, M., and M. J. Meena, “Underwater image enhancement using conventional techniques with quality
metrics,” International Journal of Innovative Technology and Exploring Engineering (IJITEE), vol. 8, no. 7S,
pp. 241-249, 2019. (https://www.ijitee.org/wp-content/uploads/papers/v8i7s/G10450587S19.pdf).
[3] Sudhakar, M., and M. Janaki Meena, “An efficient interactive segmentation algorithm using colour correction for
underwater images,” Wireless Networks, pp. 1-12, 2019. doi: https://doi.org/10.1007/s11276-019-02044-0.
[4] Sudhakar, M., and M. Janaki Meena, "A novel method to detect foreground region using morphological operations
with a block-based enhancement for underwater images," International Journal of Engineering & Technology,
vol. 7, no. 3, pp. 1751-1756, 2018. (https://www.sciencepubco.com/index.php/ijet/article/view/13159).
[5] Roser M., Dunbabin M., Geiger A., “Simultaneous underwater visibility assessment, enhancement and improved
stereo,” 2014 IEEE International Conference on Robotics and Automation (ICRA), Hong Kong, China, 2014,
pp. 3840-3847. doi: 10.1109/ICRA.2014.6907416.
[6] Sudhakar Malla and Meena M., “An edge detection mechanism using L*A*B colour-based contrast enhancement
for underwater images,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 18,
no. 1, pp. 41-48, 2020. doi: http://doi.org/10.11591/ijeecs.v18.i1.pp41-48.
[7] Ancuti C. O., Ancuti C., Haber, et al., “Fusion-based restoration of the underwater images,” 2011 18th IEEE
International Conference on Image Processing, Brussels, Belgium, 2011, pp. 1557-1560. doi:
10.1109/ICIP.2011.6115744.
[8] Ancuti C. O., Ancuti C., Bekaert P., “Effective single image dehazing by fusion,” 2010 IEEE International
Conference on Image Processing, Hong Kong, China, 2010, pp. 3541-3544. doi: 10.1109/ICIP.2010.5651263.
[9] Li C., Guo J., Cong R., Pang Y., Wang B., “Underwater image enhancement by dehazing with minimum
information loss and histogram distribution prior,” IEEE Transactions on Image Processing, vol. 25, no. 12,
pp. 5664-5677, 2016. doi: 10.1109/TIP.2016.2612882.
[10] Peng Y. T., Cao K., Cosman P. C., “Generalization of the dark channel prior for single image restoration,” IEEE
Transactions on Image Processing, vol. 27, no. 6, pp. 2856-2868, 2018. doi: 10.1109/TIP.2018.2813092.
[11] Wen H., Tian Y., Huang T., Gao W., "Single underwater image enhancement with a new optical model," 2013
IEEE International Symposium on Circuits and Systems (ISCAS), Beijing, China, 2013, pp. 753-756. doi:
10.1109/ISCAS.2013.6571956.
[12] Galdran A., Pardo D., Picn A., Alvarez-Gila A., “Automatic red-channel under water image restoration,” Journal of
Visual Communication and Image Representation, vol. 26, pp. 132-145, 2015. doi:
https://doi.org/10.1016/j.jvcir.2014.11.006.
[13] Drews P. L., Nascimento E. R., Botelho S. S., Campos M. F. M., “Underwater depth estimation and image
restoration based on single images,” IEEE Computer Graphics and Applications, vol. 36, no. 2, pp. 24-35, 2016.
doi: 10.1109/MCG.2016.26.
[14] Hitam M. S., Awalludin E. A., Yussof W. N., Bachok Z., "Mixture contrast limited adaptive histogram equalization
for underwater image enhancement," 2013 International Conference on Computer Applications Technology
(ICCAT), Sousse, Tunisia, 2013, pp. 1-5. doi: 10.1109/ICCAT.2013.6522017.
[15] Schechner Y. Y., Averbuch Y., “Regularized image recovery in scattering media,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 29, no. 9, pp. 1655-1660, 2007. doi: 10.1109/TPAMI.2007.1141.
[16] Ancuti C., Ancuti C. O., Haber T., Bekaert P., "Enhancing underwater images and videos by fusion," 2012 IEEE
Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, 2012, pp. 81-88. doi:
10.1109/CVPR.2012.6247661.
[17] Fang S., Deng R., Cao Y., Fang C., “Effective single underwater image enhancement by fusion,” Journal of
Computers, vol. 8, no. 4, pp. 904-911, 2013. doi:10.4304/jcp.8.4.904-911.
[18] Ancuti C. O., Ancuti C., Vleeschouwer C. D., Bekaert P., “Color balance and fusion for underwater image
enhancement,” IEEE Transactions on Image Processing, vol. 27, no. 1, pp. 379-393, 2018. doi:
10.1109/TIP.2017.2759252.
[19] Fu X., Fan Z., Ling M., Huang Y., Ding X., “Two-step approach for single underwater image enhancement,” 2017
International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS), Xiamen, China,
2017, pp. 789-794. doi: 10.1109/ISPACS.2017.8266583.
[20] P. Drews Jr., E. Nascimento, F. Moraes, S. Botelho, M. Campos, and R. Grande-Brazil, “Transmission estimation
in underwater single images,” 2013 IEEE International Conference on Computer Vision Workshops, Sydney,
NSW, Australia, 2013, pp. 825-830. doi: 10.1109/ICCVW.2013.113.
[21] S. Emberton, L. Chittka, and A. Cavallaro, “Hierarchical rank-based veiling light estimation for underwater
dehazing,” Processing BMVC, 2015, pp. 121-125. doi: https://dx.doi.org/10.5244/C.29.125.
[22] Li, Chongyi, et al., “An underwater image enhancement benchmark dataset and beyond,” IEEE Transactions on
Image Processing, vol. 29, pp. 4376-4389, 2020. doi: 10.1109/TIP.2019.2955241.
[23] Honnutagi, Pooja, V. D. Mytri, and Y. S. Lalitha, “Fusion-Based Underwater Image Enhancement by Weight Map
Techniques,” Recent Developments in Machine Learning and Data Analytics, Singapore, vol. 740, 2019. 327-339.
doi: https://doi.org/10.1007/978-981-13-1280-9_31.
[24] E. J. McCartney and F. F. Hall, “Optics of the atmosphere: scattering by molecules and particles,” Physics Today,
vol. 30, no. 5, pp. 76-77, 1977. doi: https://doi.org/10.1063/1.3037551.

Int J Artif Intell, Vol. 10, No. 2, June 2021: 389 – 397
Int J Artif Intell ISSN: 2252-8938 397

[25] Mourgias-Alexandris, George, et al., “An all-optical neuron with a sigmoid activation function,” Optics Express,
vol. 27, no. 7, pp. 9620-9630, 2019. doi:10.1364/OE.27.009620.
[26] Y. Cheng, W. Niu, and Z. Zhai, “Video dehazing for surveillance unmanned aerial vehicle,” 2016 IEEE/AIAA 35th
Digital Avionics Systems Conference (DASC), Sacramento, CA, USA, 2016, pp. 1-5. doi:
10.1109/DASC.2016.7777960.
[27] R. Achantay, S. Hemamiz, F. Estraday, and S. Susstrunk, “Frequencytuned salient region detection,” 2009 IEEE
Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 2009, pp. 1597-1604. doi:
10.1109/CVPR.2009.5206596.
[28] M. Yang and A. Sowmya, “An underwater colour image quality evaluation metric,” IEEE Transactions on Image
Processing, vol. 24, no. 12, pp. 6062-6071, 2015. doi: 10.1109/TIP.2015.2491020.
[29] Chiang, John Y., and Ying-Ching Chen, “Underwater image enhancement by wavelength compensation and
dehazing,” IEEE Transactions on Image Processing, vol. 21, no. 4, pp. 1756-1769, 2011. doi:
10.1109/TIP.2011.2179666.
[30] Fu, Xueyang, et al., “Two-step approach for single underwater image enhancement,” 2017 International Symposium
on Intelligent Signal Processing and Communication Systems (ISPACS), Xiamen, China, 2017, pp. 789-794. doi:
10.1109/ISPACS.2017.8266583.
[31] Salazar-Colores, Sebastián, Ivan Cruz-Aceves, and Juan-Manuel Ramos-Arreguin, “Single image dehazing using a
multilayer perceptron,” Journal of Electronic Imaging, vol. 27, no. 4, p. 043022, 2018.
https://doi.org/10.1117/1.JEI.27.4.043022.
[32] K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 33, no. 12, pp. 2341-2353, 2012. doi: 10.1109/TPAMI.2010.168.
[33] C. Ancuti, C. O. Ancuti, C. De Vleeschouwer, and A. Bovik, “Night-time dehazing by fusion,” 2016 IEEE
International Conference on Image Processing (ICIP), Phoenix, AZ, USA, 2016, pp. 2256-2260. doi:
10.1109/ICIP.2016.7532760.

BIOGRAPHIES OF AUTHORS

Mr. M Sudhakar, Currently working as Research Associate at Vellore Institute of Technology,

Chennai Campus. He finished his master‟s degree in Computer Science and Engineering in 2012
from JNTU University, Anantapur. He finished graduation from the same university in the year
2010. He has four years of teaching experience as an Assistant Professor in reputed engineering
colleges in Andhra Pradesh, India. He started his research career in 2015 in Computing Science
and Engineering at VIT Chennai Campus. He published several UGC and Scopus Indexed
journals. His research interests include Image Analytics, Image Processing, and Deep Learning
in computer vision.

M. Janaki Meena is currently working as a professor in VIT Chennai. She has competed by her
graduation from Manonmanium Sundarnar University, post-graduation from Madurai Kamaraj
University, and a doctorate from Anna University. She has got total teaching experience for 19
years from top engineering institutions in Tamilnadu. Her research interests include algorithms,
text mining, big data analytics, image processing, etc. She has authored more than twenty
research papers and one book on the theory of computation.

Multi-scale fusion for underwater image enhancement using multi-layer perceptron (M. Sudhakara)