- Email: [email protected]
Water cycle algorithm based multi-objective contrast enhancement approach
Accepted Manuscript Title: Water Cycle Algorithm based Multi-Objective Contrast Enhancement Approach Author: Manisha Kaushal Baljit Singh Khehra Akashdeep Sharma PII: DOI: Reference:
S0030-4026(17)30436-9 http://dx.doi.org/doi:10.1016/j.ijleo.2017.04.041 IJLEO 59083
To appear in: Received date: Accepted date:
19-1-2017 11-4-2017
Please cite this article as: M. Kaushal, B.S. Khehra, A. Sharma, Water Cycle Algorithm based Multi-Objective Contrast Enhancement Approach, Optik - International Journal for Light and Electron Optics (2017), http://dx.doi.org/10.1016/j.ijleo.2017.04.041 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
*Manuscript 1
Water Cycle Algorithm based Multi-Objective Contrast Enhancement Approach Abstract— Enhancement of hazy images or video is challenging task because of low contrast exhibited in them. Global
ip t
contrast stretching methods have been successful in restoring contrasts but problems like overcompensation, truncation of pixel values amounting to loss of information tends to creep in. Artifacts may be introduced and images may loose its colorfulness. This paper presents an evolutionary enhancement method for restoring contrast in images or videos while
cr
preserving its colorfulness and brightness. The study proposes a novel histogram equalization method inspired by principles of Water Cycle Algorithm. The proposed method first smoothes Y channel of YCbCr color space and divides input frame into
us
two components using Ostu’s 2D thresholding. A set of weighing constraints have been formulated and applied to both components individually in a controlled manner. Water Cycle Algorithm has been employed to exploit an optimal value of
an
weighing factors for enforcement of constraints on individual components. A three dimension objective function has been designed to suitably perform equalization and control enhancement process. Experimental results show that proposed
M
method is effective in removing haze like patterns in images and videos.
Index Terms— Contrast Enhancement; Image de-hazing; Histogram Equalization; Water Cycle Algorithm; Evolutionary Algorithm;
d
I. INTRODUCTION
te
A number of applications like object detection and tracking, anomaly detection are governed by amount of information that could be captured by cameras. Although some success has been attained with invent of sophisticated cameras but constraints imposed by weather like low illumination, fog and haze have not been still completely overpowered. It makes the task of
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
enhancing and sharpening images or videos a challenging and an evergreen problem. An image or video captured in bad weather is bound to have low contrast that can attenuate performance of various computer vision tasks. Most of image enhancement methods tend to improve contrast and highlight details by processing individual pixel level values either locally or globally. Out of these, histogram equalization (HE) is one of the most commonly implemented strategies amongst all because of its simplicity. In methods like HE, enhancement is obtained by redistribution of pixel intensities over a dynamic range. Dynamic range of histogram is stretched by histogram equalization in order to provide contrast improvement. Histogram equalization can be performed either globally or locally. The major difference between the two is that global HE methods use histogram of entire image as compared to local methods which divide image into sub-blocks and use histogram of sub-blocks independently [1]. The final enhanced image is obtained by fusing together sub-blocks using bilinear interpolation method. Local methods suffer from check board effect introduced near boundaries of blocks. Some authors have also tried to extend principles of Histogram Equalization to color images while working in RGB space or in non linear color space like YCbCr or HSI. Few of these approaches are available at [2] [3] [4]. One of the most famous approaches for this is 3-D histogram equalization [2] proposed by P. E. Trahanias and A. N. Venetsanopoulos. A 3-D cumulative distribution function (CDF) was proposed by authors to measure correlation among R, G and B components. Another method that uses multiple 2-D
Page 1 of 18
2 histograms rather than single 3-D histogram has been given by D. Menotti et al. [3]. The method looks for correlation of color channels among two different histograms and performs equalization thereafter. J. H. Han et al. in [4] has tried to compromise over enhancement of 3-D histogram technique and low contrast of 2-D histogram ignoring correlation between R, G and B channels. HE is commonly implemented strategy but enhanced image using HE tends to have unnatural enhancement, introduction of washed-out effect and intensity saturation artifacts due to error in brightness because of mean-shifting. This paper proposes a modified histogram equalization approach that not only tends to enhance contrast of image or frame
ip t
but also aims to preserve its colourfulness and brightness. The input frame is converted to YCbCr domain and its Y channel is extracted. The Y channel of input frame is smoothened using Gaussian low pass filter to suppress useless information. The
cr
frame is then portioned into background and foreground frames using Ostu’s 2D thresholding principle. The foreground and background are then subjected to suitably designed constraints. Intensity is manipulated using an improved form of histogram
us
equalization which tends to enhance frame preserving its other details with help of water cycle algorithm(WCA). WCA helps to select an optimal value of parameters based on combination of three objective functions that has been suitably framed such that over enhancement can be negated and loss of colourfulness of image or frame could be mitigated. The approach is novel
an
as it works on YCbCr space and a novel objective function for evaluation of image quality has been carefully designed for preservation of color and naturalness. This multi objective function has been utilized in water cycle optimization (WCA) algorithm for enhancement of images or videos. The desired values obtained using WCA are used to control the contrast
M
enhancement process. This paper is organized as follows: after the introduction part, some important studies in the present field have been discussed. Section 3 provides introduction to water cycle algorithm and introduces proposed system. The formal algorithm of water cycle based histogram equalization method is also part of this section. Experimental results and
te
d
discussions are provided in section 4 and section 5 provides summarization of proposed approach. II. RELATED WORK
Consider an input image I of size M*N with intensity in the range of [X low Xhigh], the probability density function P(rk) for any
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
level is given as :
(1)
where f(k) represents frequency of occurrence of level rk in image, M*N defines the total number of pixels in image and L is total number of grey levels in image. The CDF(cumulative density function) is given as (2)
Histogram equalization maps an image from narrow range to entire intensity range of [X low Xhigh] using CDF (3)
HE is a traditional contrast enhancement approach that flattens histogram and imposes a noteworthy modification in brightness of image. This section discusses several methods used by researchers to overcome limitations of HE. Kim et al.[5] proposed brightness preservation method known as Brightness Preserving Bi-Histogram Equalization(BBHE) which performs segmentation of input image into two parts on the basis of mean and performs equalization of both components individually. Another variant of BBHE that segments image on the basis of median was proposed by Wan et al.[6] known as equal area Dualistic Sub-Image Histogram Equalization(DSIHE). Another extension of BBHE known as Minimum Mean Brightness Error BiHistogram Equalization (MMBEBHE) was proposed by Chen and Ramli [7]. These methods were found to be suitable to images
Page 2 of 18
3 having uniform intensity distribution. Chen and Ramli [8] proposed Recursive Mean Separate Histogram Equalization (RMSHE) method which partitions histogram of source image recursively [8] and performs enhancement of all segments independently. Computational cost was clearly a disadvantage of this method. Few weight based equalization methods like weighted threshold HE (WTHE)[9] had also been proposed to add adaptivity and ease of control to enhancement process. These methods tend to modify probability density function by suitable weights before equalization and performed normalization afterwards with addition of some adjustment factor. Weight Clustering HE (WCHE) [10] was also developed which had its basis on adaptive
ip t
change of weights.
There had been attempts by researchers to utilize use of soft computing techniques for controlling enhancement process.
cr
Few of these have been summarized in this section. Debdoot et al. [11] proposed modified version of HE based on fuzzy statistics of digital images known as Brightness Preserving Dynamic Fuzzy Histogram Equalization (BPDFHE) to handle
us
inexactness of gray level values. Another fuzzy logic based technique (fuzzy histogram equalization-FHE) was proposed by Magudeeswaran and Ravichandaran[12] for improvement of contrast. The approach generates fuzzy histograms for original image using fuzzy set theory and follows it with segmentation of image based on median. The sub-images were then
an
independently equalized to preserve brightness. Fuzzy logic has been exploited by Hammadulu et al. in [13][14] for proposing solutions for contrast enhancement of color images. HE has also been implemented using artificial neural network based studies and Thomas H. Hildebrandt[15] were the first to propose and implement a circuit for generalized histogram
M
equalization. Another Local HE based technique was proposed by Chitwong et al. [16] who had utilized efficacies of Hopfield networks. The method uses fuzzy c-means clustering and competitive hopfield neural network for segmentation of image while performing equalization of both components individually. The technique of Kartik et al.[17] first utilizes various statistical
d
operations to perform initial enhancement of image. It then uses two right most bits of both original and enhanced image as
te
inputs and target of designed neural network. Evolution based studies like genetic algorithm(GA) and particle swarm optimization(PSO) have also been used by various researchers to propose challenging solutions for contrast enhancement of images. In order to enhance contrast of grey scale images, problem of image enhancement has been framed as an optimization
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
problem by Apurva & Ashish[18]and PSO has been utilized to attain maximum value for an edge based contrast of image. P. Shanmugavadivu et al. in [19], [20] had used concept of histogram segmentation and multi objective functions for enhancement of images. Munteanu et al. [22] had proposed application of transformation function to each pixel of input image in which GA was used to extract parameters of transformation function. Saitosh et al. [23] used GA to represent relation between input and output gray levels using lookup table. Chromosomes were evaluated using edge intensities. Changiiang and Xiaodong [24] employed In-complete Beta Transform (IBT), in which Genetic Algorithm (GA), and Wavelet Neural Network (WNN) were used to enhance contrast for an image. In this method, a non linear transform curve is obtained using In-complete Beta Transform. IBT in the whole image is approximated using a new kind of WNN. The task of GA was to determine optimal gray levels transform parameters. A classification criterion was used to determine original contrast and discrete stationary wavelet transform (DSWT) was used to enhance local contrast of image. S. Hashemi et al. [25] method used a simple and novel chromosome representation together with corresponding operators. The method was based on a simple chromosome structure to overcome shortcomings of previous methods. Two more recent studies based on use of genetic algorithm are available at [26-27].PSO based studies are simple and provides faster convergence as compared to GA based studies which tend to make use of various operators like mutation, crossover selection etc. to generate effective solutions. Readers can refer to PSO base studies at [18-22] and GA based studies [22-27] for detailed implementations.
Page 3 of 18
4 The present study tend to utilize efficacies of new optimization algorithm named as Water Cycle algorithm to propose an adaptive solution for contrast enhancement using modified version of histogram equalization. The approach has been tested qualitatively and quantitatively and is also subjected to competition with contemporary technique. The next section introduces complete methodology of the proposed technique. III. HISTOGRAM EQUALIZATION WITH COLOR RESTORATION
ip t
The performance of almost all HE based technique degrades when a more natural and complex looking image is processed. HE based techniques suffer from problem of mean shift, preservation of details, over enhancement and loss of colourfulness for
cr
colored images. The main objective of this study is to modify the basic HE principle to generate a visually pleasing image and control effectiveness of contrast with help of water cycle based optimization algorithm. Figure 1 presents the major steps of the proposed technique. All these steps are explained in details for better understanding of readers.
us
Pre-processing: In the proposed method, we first transfer the video/image sequence to YCbCr color model and extract the luminance(Y) component to reduce computational complexity. Y component is more prone to dehazing or other artifacts since
an
it is very similar to gray scale version of image. The Y channel contains rich boundary details which cause redundant information to flow to next phase and therefore smoothening of input frame has been done using 3x3 Gaussian filter. Smoothed Y channel has been used in this method to reflect uniform distribution of real scene.
M
Segmentation of input image using Ostu’s 2D threshold: Thresholding divides an input image histogram into foreground and background which can further be equalized separately in order to improve contrast of individual regions. Ostu’s method [28] is one of the most commonly used approaches to automatically segment image into two distinct classes. Ostu’s algorithm
d
searches for threshold that maximizes inter class variance between foreground and background defined as weighted sum of
where
and
te
variances of the two classes:
whole image is given by A(Y).
(4)
means average brightness of background and foreground thresholded by T, mean brightness of
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
and
are cumulative probabilities of background and foreground respectively defined as (5) (6)
Otsu’s method exhibits relatively good performance for bimodal histogram distribution but is not able to generate an optimal threshold when objects in frame are of small size, or if variances of object and background intensities gets big, or if image is too noisy. Due to one-dimensional Otsu algorithm's poor adaptability to noise image segmentation, a 2D Otsu algorithm [29] was introduced. 2D Ostu algorithm has been widely applied for segmenting images because of its content independent attributes. 2D Ostu algorithm makes use of values of neighbours together with its own value to extract an optimal threshold for improvement of binarizaion results. This makes it suitable to segment those images with small object and images with lots of noise. 2D ostu’s method works by forming a pair of pixel gray level(i) and average of its neighbours(j); and each such pair belongs to a 2 dimensional bin with total number of bins equal to L x L, where L is number of grey levels and neighbouring grey values. The joint probability mass function of such 2D histogram may be defined as number of occurrences f(i,j) of pair (i,j) divided by total number of pixels in image (MxN). (7)
Page 4 of 18
5
where
te
d
M
an
us
cr
ip t
(8)
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Figure 1: Proposed methodology for contrast enhancement
Let segmentation to be [S,T] and assuming that 2-D histogram consists of only two regions namely foreground and background, these regions will have different probabilities specified as (9) (10)
Mean vectors corresponding to two regions are given as (11) (12) Inter class variance of these regions can be specified as (13) The optimal threshold could be found by maximizing interclass variance of two classes given as and is given by (14) The input image is segmented into foreground (IF) and background(IB) sub images with help of
. The threshold generated
Page 5 of 18
6 by 2D ostu’s method will provide a better segmentation as compared with standard ostu’s method for images with multimodal distribution and noise. The readers may refer to study [29] for further details. Modification of Background and Foreground sub-images: Probability distributions for foreword sub-image(IF) and background sub-images(IB) obtained in previous step are calculated as
and
. These distributions are subjected to
suitably defined constraints for background and foreground respectively. The modified probability density function for and back- ground
) frames are specifed as
ip t
foreground (
cr
(15)
an
us
τ
(16)
Mean of both modified probability distribution and normal pdf is calculated for both foreground and background sub images. The relative difference among both is added to and
for
background
and
foreground
is
. The cummulative distribution functions calculated
as
M
using
and
and
and new historam equalization procedure is applied according to (17)
d
(18)
te
These constraints are defined so that effect of dominating intensities could be controlled and effect of intensities with small PDF could be ignored. For this reason PDFs of both foreground and background are clamped with respect to certain threshold values specified in equation 15 and 16. Values of controlling parameters (r,t) are fixed in the range from 0.1-1.0 as over
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
enhancement creeps in beyond this range. The degree of enhancement is controlled by power parameters (q,s), less probable levels in the corresponding sub-images can be reallocated with more probable levels when the value of these parameters is < 1.0. It helps in preservation of visual details of the image. When value of power factor approach 1.0, the procedure behaves like standard histogram equalization, when value goes beyond 1.0, more weight is added to high probability levels and these constraints impose an even stronger effect than HE which may result in over enhancement. The resolved constraints are applied to both sub-images and both are combined to generate new image I
new
which is passed to next step for constraint
optimization. The optimization of parameters is done with water cycle algorithm as explained in next section. A. Water cycle algorithm based optimization of constraints parameters For implementation of histogram equalization with color restoration we had defined four variables namely (q,r,s,t) that will help to control the effects of standard HE and the values of these parameters range from 0.1-1.0. This section discusses how water cycle algorithm has been used to extract an optimal combination of these parameters. Water Cycle Algorithm Water Cycle Algorithm(WCA) is new population based meta-heuristic algorithm based on the observation of nature’s water cycle and finds solutions mimicking flow of rivers and streams toward sea [30] [31]. WCA creates an initial population of design
Page 6 of 18
7 variables or in other words population of streams (raindrops) after assuming rainfall or precipitation. The best individual (best raindrop) i.e. one having least value of cost function is chosen as a sea. Then, a number of good raindrops i.e. solutions having cost function closest to current best are chosen as rivers and rest of the raindrops are considered as streams which flow to the rivers and sea. In order to simulate this effect an initial population with dimensions Npop × D is created where D represents
cr
ip t
number of design variables and Npop is population size.
an
us
(19)
Out of total of Npop streams created initially, Nsr represents best individuals from population designated as sea and rivers. The rest of population NStream are considered as streams flowing into rivers or into sea directly. Depending on their magnitude
M
of flow each river absorbs water from the streams and amount of water in a stream entering a river and/or sea varies from other streams. Sea is assumed to be most downhill location and rivers flow to the sea. Figure 2 illustrates complete WCA optimization process.
where
te
given
d
In order to designate/assign raindrops to the rivers and sea depending on the intensity of the flow, the following equation is
is the number of streams which flow into the specific rivers and the sea and
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
(20) is the cost of respective raindrop.
Flow of rivers to sea and streams to rivers can be modelled as
where C is number between 1 and 2. Initial value of X may be in the range
(21) (22) where d is current distance
between stream and river and X is randomly chosen distance to depict flow of streams to river along the connecting line(Fig. 2). Evaporation process(Exploitation) is also modelled to prevent premature convergence to local minima. Evaporation may lead to more precipitation and algorithm checks whether river/stream is sufficiently close to sea to enable the evaporation process to occur. Following criterion is utilized for the evaporation condition between a river and the sea: (23)
where dmax is a small number close to zero and value of dmax adaptively decreases as follows: (24) After evaporation, the raining process is applied and new streams are formed in different locations (similar to mutation in the GAs). In the raining process, new raindrops form streams in different locations (acting similar to mutation operator in GA).
Page 7 of 18
cr
Figure 2: Water cycle optimization process (a) Flow of streams to rivers (b) WCA optimization process [30]
(25)
us
For specifying locations of the newly formed streams, the following equation is used:
where LB and UB are lower and upper bounds defined by given problem respectively. Termination condition of WCA may be
refer to [30] [31] for more details about water cycle algorithm.
an
fixed as maximum number of iterations or small tolerance value between consecutive iterations. The readers are advised to
Objective function for WCA: The main aim of proposed technique is to enhance contrast of input frame while preserving
M
brightness of source image. Two different functions have been used for implementing this aspect where one will try to maximize contrast and other will be used to preserve brightness. For preservation of brightness we propose difference in
d
average intensities of enhanced and original image as one of the function which can be calculated as (26) (26)
te
Several methods are available to measure contrast like discrete entropy(DE), contrast improvement index, edge based contrast etc. In this paper we had chosen a modified contrast index used by Y. Wang et al.[32] in which contrast of image is measured by combining different performance measures like entropy, sum of edge based intensities and number of edge pixels.
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
ip t
8
where I is image of M x N pixels;
(27)
is visible edge image of I which can be computed as [33]. E(.)is sum of internsities of
visible image, ne represents number of edges and H represents entropy of image. In order to preserve colorfulness of the images, a parameter that measures colorfulness of image is used(CCI). CCI reflects the richness and vivid degree of color and is measured by using work of [34]. In order to maintain colorfulness of frame, difference of original and enhanced frame is taken as third objective function to optimize. (28) The three metrics can be combined together to get a single function as (29) In order to make final image look natural and rich in contrast with preservation of colorfulness and brightness it is desirable to maximize value of E and therefore we had employed equation 29 as objective function for our problem.
and
are weighing
factors for controlling levels of brightness and color component enhancement of image. The ideal values for these weights shall be small otherwise the output image will be over enhanced. The values of these weights were set experimentally depending on
Page 8 of 18
9 nature of input image. The complete optimization process is implemented using water cycle algorithm which is detailed in the next section. B. Water Cycle Algorithm Based on above observation, we tend to formulate the problem of searching for optimal parameters as maximization of equation for E which could be stated as follows out
(30)
such that 0.1
ip t
(q*,r*,s*,t*) = arg max E (I (q,r,s,t))
This section illustrates how the problem of parameter detection has been solved using water cycle algorithm. The value of
cr
proposed objective function(E) is combination of three individual functions providing modeling of contrast, brightness and color. Since value of E is affected by parameters (p,q,r,s) and it is empirically difficult to choose an optimal combination for
us
which equation(30) will be maximized, authors tend to seek help of water cycle algorithm for solving this optimization problem. Figure 3 introduces complete algorithm to perform enhancement of input image or frame. 1. Input an image or frame of video and convert it to YCbCr color space.
an
2. Extract Y channel and use 3x3 Gaussian mask to smoothen the input image.
3. Use 2-D ostu’s thresholding defined in section 3 to extract appropriate threshold value (
M
14.
for the frame using equation
4. Apply threshold generated in previous step to extract two sub images from original image as foreground(IF) and background(IB).
d
5. Call Water cycle based algorithm for finding optimal values for (q, r, s, t)
q,r,s,t), LB =0.1, UB=1.0 etc.
te
a. Initialize WCA parameters: population size(N=30), Nsr=6,dmax,max_it(40), number of parameters (nvars =4 for
b. Generate random population (equation 19), subject to condition 0.1
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
c. For all population species, extract new probability density functions for foreground and background according to equations 15 and 16. d. Perform equalization of foreground and background separately according to equation 17 and 18. e. Find values of individual objective functions and combine these to calculate the cost of population species according to equation 29. f. Sort the population according to cost and assign first/best candidate as sea, perform initialization of rivers and streams accordingly.
Page 9 of 18
10 g. Designate streams to rivers and sea according to equation 20. h. Model flow of streams to rivers and rivers to sea according to equation 21 and 22. i. Exchange positions of river with stream which gives best solution and exchange position of sea with river if better solution is available(Figure 2). j. Perform evaporation process according to equation 23 and in case there exists evaporation perform raining process
k. Update dmax using equation 24.
ip t
according to 25.
l. Check for convergence, if stopping criteria is met return value of sea as best solution else return to step 5(g).
cr
6. Use the optimal values returned by WCA to generate enhanced Y channel image.
7. Combine CbCr channel information into enhanced Y channel and convert back to RGB space.
us
Figure 3: Proposed water cycle algorithm based histogram equalization method
The algorithm given in figure 3 inputs a hazy image or frame in first step and as part of pre-processing operation extracts Y channel of image after converting from RGB space to YCbCr color space. The input image is segmented into foreground and
an
background sub images using 2-D ostu’s thresholding principle. 2-D ostu’s method has been utilized as it is more suitable to complex or natural images. Both sub images are enhanced separately by subjecting respective probability density functions of foreground and background to specifically designed constraints. The designed constraints can be controlled with four
M
parameters (q,r,s,t) as illustrated in previous sections. Different combinations of these parameters will constitute for different enhanced versions of final image and in order to select an optimal value, water cycle algorithm has been used for this purpose.
d
WCA first generates a random population of raindrops(stream) within fixed lower and upper bound. The algorithm then proceeds to initialize sea, rivers and streams from initial population. Flow of streams to rivers and rivers to sea is modelled for
te
fixed number of iterations and best solutions are re-designated as sea or rivers. The objective function for water cycle optimization is considered to be a three dimensional function that tends to enhance contrast and preserve brightness and colorness of input image. The best solution is utilized for modification of foreground and background sub-images which are
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
eventually combined to generate final enhanced Y channel image. The CbCr components are added to the enhanced Y channel image to get the final image.
IV. RESULTS AND DISCUSSION
In order to test performance of proposed method, experiments were conducted on number of hazy and unclear images downloaded from [35]. In addition, the proposed method was also tested on hazy videos used by [36], object detection videos from [37]. The frames or images were subjected to proposed approach for 40 iterations and best results obtained are reported in this section. In order to compare performance of proposed algorithm, same images were enhanced with contemporary techniques like Histogram Equalization(HE), BBHE(Bilevel histogram equalization), CLAHE(contrast limited histogram equalization) and RMSHE. The performance of all these methods has been evaluated in terms of human visual perception and some qualitative measures like entropy, edge based contrast, colorfulness, contrast improvement index(CII), MSE and PSNR. Although tests were conducted on all sorts of images we hereby demonstrate only color, complex and natural images except lenna image (figure 4) which has been included for completeness. Figures 4-9 presents visual results of some sample images on which tests were conducted. Enhanced images obtained by HE, BBHE, CLAHE, RMSHE and proposed method are shown in corresponding images from 4(b)-(f) to 9(b)-9(f) respectively.
Page 10 of 18
11
a
b
c
d
e
f
b
c
d
e
us
a
cr
ip t
Figure 4 (a)Original Lenna image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
f
Figure 5 (a) Original camel image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f)Proposed
HE is not able to preserve original color of image and over-enhances can be seen in figure 4(b); the results obtained by CLAHE
an
and RMSHE are almost same but improved results can be seen for BBHE (fig 4(c)). The proposed method was not only able to provide a contrast rich image but color and details of image were also preserved. The final obtained image is more natural and
M
colourful. In Figure 5, we can see that none of the contemporary technique was able to perform enhancement effectively. Artifacts can be observed in images enhancements by HE, BBHE, CLAHE and RMSHE where as image shown in fig 5(f) is more natural and bright. The color of camel seat has been turned to black by HE, BBHE and RMSHE. The results of CLAHE are better
d
than these methods however CLAHE was not able to perform enhancement of sand and sky regions that look far better in proposed method. In Figure 6, we see that original color of sand and grass are lost after performing enhancement using HE,
te
BBHE, CLAHE and RMSHE methods. Some white spots can be observed on portions of image covering bench in HE, BBHE and RMSHE but no such area can be seen in final image generated by proposed technique. The enhanced image looks rich in contrast and is able to remove haziness from background in a far better way than any of the other methods. In Figure 7, a
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
redundant bright spot has been generated in background for HE and BBHE methods. In addition HE, BBHE and RMSHE were not able to perform enhancement of lower sub areas of image as highlighted in circles. Results of CLAHE and proposed method are comparable however proposed method is more successful in preserving brightness of original image. In figure 8, we can easily observe that proposed method was able to preserve details of aircraft while others have failed completely. This is another demonstration that proposed technique has successfully performed enhancement of image while preserving color, brightness and details of original image. In figure 9, contemporary techniques like HE, BBHE, RMSHE failed to provide enhancement in regions that are far away from the source, in addition color information of truck is also lost in HE and BBHE methods. Results of CLAHE seem to be better but whiteness and artifacts are visible in portions of image which is encircled. The proposed method was also tested for enhancement of video frames on hazy video sequences and moving objects videos from CAVAIR dataset. The visual results are shown in figure 10 and 11 of frame number 55 (hazing road) and frame number 3(video Br3 of CAVIAR). The results of the proposed method are far better than all the contemporary technique which justifies applicability of proposed technique to videos also.
Page 11 of 18
12
b
c
d
e
f
b
us
cr
ip t
Figure 6 (a) Foggy bench image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
d
c
e
f
M
an
Figure 7 (a) Tienamen image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
c
d
e
d
b
f
te
Figure 8(a) Aircraft image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
Ac ce p
1 2 3 4 5 6 7 8 9 a 10 11 12 13 14 15 16 17 18 19 20 21 22 a 23 24 25 26 27 28 29 30 31 32 33 a 34 35 36 37 38 39 40 41 42 43 44 a 45 46 47 48 49 50 51 52 53 54 a 55 56 57 58 59 60 61 62 63 64 65
b
c
d
e
f
Figure 9 (a) Foggy Bus image (b)HE (c)BBHE (d)CLAHE (e) RMSHE (f)Proposed
b
c
d
e
f
Figure 10 (a) Frame #55 of hazing road (b)HE(c)BBHE (d)CLAHE (e) RMSHE (f)Proposed
Page 12 of 18
13
a
b
c
d
e
f
Size
Original
HE
BBHE
512 x 512
5.51955
7.07394
7.14725
Tienamen
600 x 450
7.64235
7.22936
7.30205
foggy bench
800 x 600
7.41689
7.36125
foggy bus
820 x 553
7.10801
7.29976
3096_airplane
481 x 321
6.23429
6.88136
Camel
481 x 321
6.92767
mountains
481 x 321
7.17124
Haze
400 x 300
7.26217
Foggy Trees
1076 x 723
7.42706
Frame 55
--
7.2686
Frame #3
--
7.2608
CLAHE
RMSHE
Proposed
6.83556
6.24508
7.21614
7.60426
7.68047
7.69592
7.39041
7.48466
7.8954
7.76138
7.4575
7.39229
7.54153
7.43937
6.98299
6.95618
6.99534
6.55078
7.41832
7.42756
7.55185
7.78121
7.52567
7.21629
7.23563
7.3936
7.36526
7.29593
7.33521
7.46763
7.31939
7.33572
7.54303
7.3476
7.31067
7.72052
7.79199
7.27353
7.0578
7.3737
7.5412
7.3526
7.3645
7.3564
7.4908
7.5527
7.7714
7.7812
te
d
an
us
Lenna
M
Image
cr
TABLE 1: COMPARISON OF DISCRETE ENTROPY VALUES
ip t
Figure 11 (a) Original Frame #3 of Br3 (b)HE(c)BBHE (d)CLAHE (e) RMSHE (f) Proposed
TABLE 2: COMPARISON OF CONTRAST VALUES
Image Lena
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Original
HE
BBHE
CLAHE
RMSHE
Proposed
85
41.81753
43.41483
54.66605
53.6942
57.53099
50.2959
40.49902
44.82648
48.77289
45.913
51.12503
53.25074
43.39474
52.37272
43.39441
54.8284
55.08023
7.10801
7.29976
7.4575
7.39229
7.54153
7.4393
3096_airplane
39.25515
42.18912
51.00271
48.18964
49.1758
48.579
Camel
13.16601
36.04692
30.37829
25.40675
29.32267
47.37631
Mountains
38.61747
40.7504
37.4932
42.65657
38.35404
45.3334
Haze
47.66814
41.47838
48.30346
53.19276
46.33241
55.1931
Foggy Trees
19.60666
46.0364
30.5084
38.82107
30.92912
42.5789
Frame 55
40.8643
40.2576
40.1619
39.8530
45.3518
43.8153
Frame #3
32.4323
36.7071
32.8179
39.412
33.1266
39.487
Tienamen foggy bench foggy bus
Page 13 of 18
14 TABLE 3: COMPARISON OF COLORFULNESS VALUES Image
Original
HE
BBHE
CLAHE
RMSHE
Proposed
17.2714
17.2619
16.5929
17.2256
17.26335
18.00179
Tienamen
31.84715
31.82578
33.4687
31.8983
31.88933
31.924
foggy bench
16.3323
15.1656
15.5917
15.0338
16.06058
16.29719
foggy bus
7.50751
8.07108
8.07778
8.11048
8.11322
8.17545
12.5687
11.9457
12.5974
12.5517
12.4877
12.57478
Camel
39.9830
36.4172
37.0909
39.8726
39.9964
40.03432
Mountains
41.7874
42.8050
42.7777
41.9980
41.8426
42.97118
Haze
11.2910
13.1564
13.0575
11.5282
Foggy Trees
20.6650
21.7223
21.9232
21.0708
Frame 55
14.9522
15.0578
15.870
15.0322
14.8658
15.7904
Frame #3
18.1431
16.3281
16.684
17.5015
17.3402
17.5755
cr 11.5338
11.53948
20.7415
21.93014
us
an
3096_airplane
ip t
Lenna
TABLE 4: COMPARISON OF CII VALUES HE
BBHE
1.4521
1.35789
Tienamen
1.1689
1.1789
foggy bench
1.5781
0.9874
foggy bus
0.94785
3096_airplane
1.0178
Camel
0.1897
Mountains
0.00456
Haze
2.417621
RMSHE
Proposed
1.566949
3.49328
2.013815
1.181047
2.134998
2.12056
2.054997
2.37136
0.9268
1.302322
0.85533
1.443819
1.2713
1.561
1.705542
1.344101
0.21056
2.166294
1.797936
2.602408
0.00712
0.189981
0.121619
0.144246
te
d
Lenna
CLAHE
M
Image
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
0.0896
0.10489
2.064851
1.226329
2.001621
Foggy Trees
0.03147
0.03258
1.396552
1.678399
1.672023
Frame 55
0.00124
0.00256
0.0897
0.0344
0.1025
Frame #3
0.0192
0.0574
1.6982
1.69054
2.1789
The qualities of test images enhanced using various algorithms were also measured in terms of discrete entropy, edge based contrast, contrast improvement index, colorfulness, MSE and PSNR values. Tables 1 through 6 demonstrate results obtained using these algorithms for images and videos. In Table 1, we can see that proposed method either led to increase in entropy of an image or maintain the same degree of entropy values. This means that details of source image were preserved by the proposed objective function. The effectiveness of proposed technique is also justified when we observe the values obtained for contrast and colorfulness of various test images demonstrated in table 2 and table 3. The proposed method is efficient in terms of improving contrast for almost all images barring a few in which values observed are rather competitive. The proposed method has also justified that it is able to preserve colorfulness of original image as obtained results are quite promising.
Page 14 of 18
15 TABLE 5: COMPARISON OF MSE VALUES HE
BBHE
CLAHE
RMSHE
Proposed
Lenna
132.5898
124.3801
98.82253
0.39547
92.5898
Tienamen
141.6714
113.3389
49.79701
146.0021
36.81183
foggy bench
153.7654
112.5811
160.3484
103.735
117.9389
foggy bus
110.9421
111.0487
101.3295
84.71749
86.3038
3096_airplane
108.5361
80.846
26.72487
67.24598
26.73755
Camel
88.74955
114.2815
49.70072
87.68683
14.22405
Mountains
67.20057
110.2334
45.5466
28.22699
57.80656
Haze
131.6577
86.45403
23.69154
Foggy Trees
56.75089
127.6994
32.06301
cr
ip t
Image
15.70304
113.961
50.5162
us
26.67682
BBHE
Lenna
26.9397
27.21729
Tienamen
26.65202
27.62104
foggy bench
26.29622
27.65015
foggy bus
27.71385
3096_airplane
CLAHE
Proposed
28.21624
52.19731
28.46517
31.19285
26.52125
32.47093
26.11417
28.00555
27.41423
27.70968
28.10757
28.88507
28.7705
27.80936
29.08868
33.89567
29.88823
33.85959
Camel
28.68314
27.58505
31.2012
28.7355
36.60057
Mountains
29.91631
27.75003
31.58071
33.65871
30.51103
Haze
26.97034
28.79695
34.41893
33.90347
36.17097
Foggy trees
30.62668
27.10305
33.10642
27.59737
31.0965
M
HE
te
Image
an
RMSHE
d
TABLE 5: COMPARISON OF PSNR VALUES
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Another quantization measure of contrast enhancement can be defined by contrast improvement index(CII), and its formula can be expressed by the following
where
and
(31)
are the contrasts of the processed and original images respectively. C is the average value of
the local region contrast in the processed or original image which can be calculated as (32) where Xf and Xb are maximum and minimum luminance values of foreground and background regions in local window size of 5*5. Higher value of cii represents better image quality and values of CII have been normalized according to their size in the present study. Table 4 lists values of CII obtained for various methods, value of CII for original image comes to be 1. The table depicts an increase in values of CII over other methods suggesting effectiveness of proposed algorithm . MSE and PSNR measures which have been borrowed from signal processing have also been used historically for evaluation of image quality. Although their findings cannot be considered reliable nevertheless, their use is still common practice and for the
Page 15 of 18
16 sake of completeness these have also been included in our study. The mean square error and PSNR between two images of size(MxN) can be defined as: (33) (34) L reflects the range of values that a pixel can take; The values of MSE and PSNR for test images are tabulated in tables 5 and
ip t
6 respectively. A small mean square error results in a high signal to noise ratio, if MSE tends to zero, then PSNR tends to infinity. More similarity between original and final image will result in minimization of MSE parameter and enhancement of
cr
PSNR parameter and the observed values are in correlation to the values for other parameters. V. CONCLUSIONS
us
This paper has presented an evolutionary approach based on water cycle algorithm for enhancement of images and videos. The proposed method performs preprocessing of input frames and then runs a controlled version of equalization for
an
performing contrast stretching. Enhancement is controlled by utilizing powers of WCA by finding an optimal combination of variables according to specific image. Suitable membership function has been designed to prevent loss of colourfulness and truncation of pixel values in enhanced image. This has resulted in generation of images which are rich in contrast, colour and
M
free from information loss. In order to test effectiveness of proposed technique, different levels of images comprising standard, natural or complex and frames from hazy videos were used and a number of parameters were evaluated. The proposed technique was also put to competition tested with other contemporary techniques of contrast stretching and results
d
demonstrate that the proposed technique is capable of removing haze effects and preserving colourfulness and brightness of
te
original frames. REFERENCES
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
[1] Gonzalez Rafael C, Richard E. Woods, “Digital image processing”. 2nd ed. India: Prentice Hall; 2002. th
[2] P. E. Trahanias, A.N. Venetsanopoulos. “Color image enhancement through 3-d histogram equalization”, in Proc. 11 IAPR Conf. Pattern Recogn., vol. III, 1992,pp. 545–548.
[3] D. Menotti, L. Najman, A. de Araújo, J. Facon. “A fast hue-preserving histogram equalization method for color image enhancement using a Bayesian framework”, in: Proc. 14th International Workshop on Systems, Signal and Image Processing (IWSSIP), June 2007, pp. 414–417.
[4] J.H. Han, S. Yang, B.U. Lee.(2011) A novel 3-D color histogram equalization method with uniform 1-D gray scale histogram, IEEE Trans. Image Process. 2 (2) pp. 506–512. [5] Y. Kim(1997) Contrast enhancement using brightness preserving bi-histogram equalization. IEEE Tran. Consumer Electronics ;43,pp. 1–8. [6] Y. Wan, Q. Chen, B. Zhang. (1999). Image enhancement based on equal area dualistic sub-image histogram equalization method. IEEE Tran. on Consumer Electronics 45, pp 68–75. [7] S. Chen, A. R. Ramli.(2003). Minimum mean brightness error bi-histogram equalization in contrast enhancement. IEEE Tran. on Consumer Electronics 49,pp. 1310–29.
Page 16 of 18
17 [8] S. Chen, A. R. Ramli.(2003) Contrast enhancement using recursive mean-separate histogram equalization for scalable brightness preservation. IEEE Trans. Consumer Electronics. 49, pp 1323–32. [9] Q. Wang, R. K. Ward.(2008) Fast image/video contrast enhancement based on weighted thresholded histogram equalization. IEEE Trans. Consumer Electronics. 53, pp. 757–64. [10] N. Sengee, H. K. Choi.(2008). Brightness preserving weight clustering histogram equalization. IEEE Transactions on Consumer Electronics. 54, pp. 1329–37.
ip t
[11]D. Sheet, H. Garud, A. Suveer, M. Mahadevappa, J. Chatterjee.(2011). Brightness preserving dynamic fuzzy histogram equalization. IEEE trans. consumer electronics. 56(4), pp. 2475-2480.
cr
[12] V. Magudeeswaran, C. G. Ravichandran.(2013). Fuzzy Logic Based Histogram Equalization for Image Contrast Enhancement. Mathematical Problems in Engineering. Available: http://dx.doi.org/10.1155/2013/891864
us
[13] H. Madasu,D. Jha.(2006). An optimal fuzzy system for color image enhancement. IEEE Transactions on Image Processing. 15, pp. 2956–66.
[14] M. Hanmandlu, O. P. Verma, K. K. Nukala, M. Kulkarni.(2009)A Novel Optimal Fuzzy System for Color Image Enhancement
an
Using Bacterial Foraging. IEEE trans. instrumentations and measurement.58 (8), pp. 2867-2879. [15] T. H. Hildebrandt.(1993) A Local Neural Implementation of Histogram Equalization. Available: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=298809
M
[16] S. Chitwong, T. Boonmee, F. Cheevasuvit. “Local area histogram equalization based multispectral image enhancement from clustering using competitive Hopfield neural network” in Can. Conf. Electrical and Computer Engg. June 2003, pp. 1715 1718 vol.3 ·
d
[17] K. Ch. Jena, N. Dash, N.K. Kamila. (2015). A neural network based approach for Image Enhancement using Histogram
te
Equalization. International Journal of computer Systems. 02(11), pp. 460-463 [18] A. Gorai, A. Ghosh. “Gray-levelimage enhancement by particle swarm optimization”. in proc. of world Cong. on Nature and Biologically Inspired Computing 2009, pp. 72–77.
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
[19] P.Shanmugavadivu, K.Balasubramanian.(2014) Particle swarm optimized multi-objective histogram equalization for image enhancement, J. optic and laser technology, 57, pp. 243-251. [20]P. Shanmugavadivu, K. Balasubramanian, Muruganandam, (2014). Particle swarm optimized bi-histogram equalization for contrast enhancement and brightness preservation of images. A. Vis. Comput. 30, 387-392. doi:10.1007/s00371-013-0863-8 [21] S. M. W. Masra, P. K. Pang, M. S. Muhammad, K. Kipli. “Application of Particle Swarm Optimization in Histogram Equalization for Image Enhancement”. In proc. IEEE Colloquium on Humanities, Science & Engineering Research (CHUSER 2012), December 3-4, 2012, Kota Kinabalu, Sabah, Malaysia, pp. 294-299 [22] C. Munteanu, A. Rosa, “Towards automatic image enhancement using genetic algorithms”. in proc. of the Congress on Evolutionary Computation, vol. 2, pp. 1535–1542. [23] F. Saitoh. “Image contrast enhancement using genetic algorithm”. In proc. IEEE Internat. Conf. Syst. Man Cybern., 4 (1999), pp. 899–904 [24] Z. Changjiang, W. Xiaodong. Global and local contrast enhancement for image by genetic algorithm and wavelet neural network. LNCS Book Series, vol. 4234, Springer (2006) pp. 910–919 [25] H. Sara, K. Soheila,N. Navid, E. M. Mohsen.(2010). An image contrast enhancement method based on genetic algorithm. Pattern Recognition Letters, 31pp. 1816–24. doi: 10.1109/ICDIP.2009.87
Page 17 of 18
18 [26] P. Babu, V. Rajamani. Contrast Enhancement using Real Coded Genetic Algorithm Based Modified Histogram Equalization for Gray Scale Images.2015. Int. J. Imaging Syst. Technol. (25), pp. 24-32. [27] S. Sedighi, M. Roopaei, S. Agaian. “Genetic-Based Thresholds for Multi Histogram Equalization Image Enhancement” in proc. Of 12th Int. Conf., MLDM 2016, New York, USA, July 16-21, pp. 483-490 [28] N. Otsu (1979). A threshold selection method from gray-level histograms. IEEE Trans. Sys., Man., Cyber.9 (1), pp. 62– 66. doi:10.1109/TSMC.1979.4310076.
ip t
[29] J. Liu, W. Li. (1993) The automatic thresholding of gray-level pictures via two-dimensional Otsu method, Acta Automatica Sinica 19 (1), pp. 101–105.
cr
[30] H. Eskandar, A. Sadollah, A. Bahreininejad, M. Hamdi.(2012) Water cycle algorithm – A novel metaheuristic optimization method for solving constrained engineering optimization problems. Computers and Structures (110–111), pp. 151–166.
us
[31] A. Sadollah, H. Eskandar, H. M. Lee, D. G. Yoo, J. H. Kim. (2016). Water cycle algorithm: A detailed standard code. SoftwareX(2016), http://dx.doi.org/10.1016/j.softx.2016.03.001.
[32] Y. Wang, H. Wang, C. Yin, M. Dai.(2016). Biologically inspired image enhancement based on Retinex., Int. J.
an
Neurocomputing. (177), pp. 373-384. http://dx.doi.org/10.1016/j.neucom.2015.10.124
[33] N. Hautière, J. P. Tarel, D. Aubert, É Dumont.(2008). Blind contrast enhancement assessment by gradient ratioing at visible edges. Image Anal. Stereol. J., 27 (2), pp. 87–95
M
[34] D. Hasler, S.E. Suesstrunk, Measuring colorfulness in natural images, in: Electronic Imaging, International Society for Optics and Photonics, Santa Clara, 2003, pp. 87–95.
[35] Haze Removal using matlab. Available: https://github.com/akutta/Haze-Removal. last accessed on 11 october 2016.
te
Vis. Comm. and Image Repre. 24(3) pp. 410-425.
d
[36] J. H. Kim, W. D. Jang, J. Y. Sim, C.S. Kim.(2013) Optimized contrast enhancement for real-time image and video dehazing, J.
[37] CAVIAR: Context Aware Vision using Image-based Active Recognition. http://homepages.inf.ed.ac.uk/rbf/CAVIAR/ (2010).
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Page 18 of 18
S0030-4026(17)30436-9 http://dx.doi.org/doi:10.1016/j.ijleo.2017.04.041 IJLEO 59083
To appear in: Received date: Accepted date:
19-1-2017 11-4-2017
Please cite this article as: M. Kaushal, B.S. Khehra, A. Sharma, Water Cycle Algorithm based Multi-Objective Contrast Enhancement Approach, Optik - International Journal for Light and Electron Optics (2017), http://dx.doi.org/10.1016/j.ijleo.2017.04.041 This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
*Manuscript 1
Water Cycle Algorithm based Multi-Objective Contrast Enhancement Approach Abstract— Enhancement of hazy images or video is challenging task because of low contrast exhibited in them. Global
ip t
contrast stretching methods have been successful in restoring contrasts but problems like overcompensation, truncation of pixel values amounting to loss of information tends to creep in. Artifacts may be introduced and images may loose its colorfulness. This paper presents an evolutionary enhancement method for restoring contrast in images or videos while
cr
preserving its colorfulness and brightness. The study proposes a novel histogram equalization method inspired by principles of Water Cycle Algorithm. The proposed method first smoothes Y channel of YCbCr color space and divides input frame into
us
two components using Ostu’s 2D thresholding. A set of weighing constraints have been formulated and applied to both components individually in a controlled manner. Water Cycle Algorithm has been employed to exploit an optimal value of
an
weighing factors for enforcement of constraints on individual components. A three dimension objective function has been designed to suitably perform equalization and control enhancement process. Experimental results show that proposed
M
method is effective in removing haze like patterns in images and videos.
Index Terms— Contrast Enhancement; Image de-hazing; Histogram Equalization; Water Cycle Algorithm; Evolutionary Algorithm;
d
I. INTRODUCTION
te
A number of applications like object detection and tracking, anomaly detection are governed by amount of information that could be captured by cameras. Although some success has been attained with invent of sophisticated cameras but constraints imposed by weather like low illumination, fog and haze have not been still completely overpowered. It makes the task of
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
enhancing and sharpening images or videos a challenging and an evergreen problem. An image or video captured in bad weather is bound to have low contrast that can attenuate performance of various computer vision tasks. Most of image enhancement methods tend to improve contrast and highlight details by processing individual pixel level values either locally or globally. Out of these, histogram equalization (HE) is one of the most commonly implemented strategies amongst all because of its simplicity. In methods like HE, enhancement is obtained by redistribution of pixel intensities over a dynamic range. Dynamic range of histogram is stretched by histogram equalization in order to provide contrast improvement. Histogram equalization can be performed either globally or locally. The major difference between the two is that global HE methods use histogram of entire image as compared to local methods which divide image into sub-blocks and use histogram of sub-blocks independently [1]. The final enhanced image is obtained by fusing together sub-blocks using bilinear interpolation method. Local methods suffer from check board effect introduced near boundaries of blocks. Some authors have also tried to extend principles of Histogram Equalization to color images while working in RGB space or in non linear color space like YCbCr or HSI. Few of these approaches are available at [2] [3] [4]. One of the most famous approaches for this is 3-D histogram equalization [2] proposed by P. E. Trahanias and A. N. Venetsanopoulos. A 3-D cumulative distribution function (CDF) was proposed by authors to measure correlation among R, G and B components. Another method that uses multiple 2-D
Page 1 of 18
2 histograms rather than single 3-D histogram has been given by D. Menotti et al. [3]. The method looks for correlation of color channels among two different histograms and performs equalization thereafter. J. H. Han et al. in [4] has tried to compromise over enhancement of 3-D histogram technique and low contrast of 2-D histogram ignoring correlation between R, G and B channels. HE is commonly implemented strategy but enhanced image using HE tends to have unnatural enhancement, introduction of washed-out effect and intensity saturation artifacts due to error in brightness because of mean-shifting. This paper proposes a modified histogram equalization approach that not only tends to enhance contrast of image or frame
ip t
but also aims to preserve its colourfulness and brightness. The input frame is converted to YCbCr domain and its Y channel is extracted. The Y channel of input frame is smoothened using Gaussian low pass filter to suppress useless information. The
cr
frame is then portioned into background and foreground frames using Ostu’s 2D thresholding principle. The foreground and background are then subjected to suitably designed constraints. Intensity is manipulated using an improved form of histogram
us
equalization which tends to enhance frame preserving its other details with help of water cycle algorithm(WCA). WCA helps to select an optimal value of parameters based on combination of three objective functions that has been suitably framed such that over enhancement can be negated and loss of colourfulness of image or frame could be mitigated. The approach is novel
an
as it works on YCbCr space and a novel objective function for evaluation of image quality has been carefully designed for preservation of color and naturalness. This multi objective function has been utilized in water cycle optimization (WCA) algorithm for enhancement of images or videos. The desired values obtained using WCA are used to control the contrast
M
enhancement process. This paper is organized as follows: after the introduction part, some important studies in the present field have been discussed. Section 3 provides introduction to water cycle algorithm and introduces proposed system. The formal algorithm of water cycle based histogram equalization method is also part of this section. Experimental results and
te
d
discussions are provided in section 4 and section 5 provides summarization of proposed approach. II. RELATED WORK
Consider an input image I of size M*N with intensity in the range of [X low Xhigh], the probability density function P(rk) for any
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
level is given as :
(1)
where f(k) represents frequency of occurrence of level rk in image, M*N defines the total number of pixels in image and L is total number of grey levels in image. The CDF(cumulative density function) is given as (2)
Histogram equalization maps an image from narrow range to entire intensity range of [X low Xhigh] using CDF (3)
HE is a traditional contrast enhancement approach that flattens histogram and imposes a noteworthy modification in brightness of image. This section discusses several methods used by researchers to overcome limitations of HE. Kim et al.[5] proposed brightness preservation method known as Brightness Preserving Bi-Histogram Equalization(BBHE) which performs segmentation of input image into two parts on the basis of mean and performs equalization of both components individually. Another variant of BBHE that segments image on the basis of median was proposed by Wan et al.[6] known as equal area Dualistic Sub-Image Histogram Equalization(DSIHE). Another extension of BBHE known as Minimum Mean Brightness Error BiHistogram Equalization (MMBEBHE) was proposed by Chen and Ramli [7]. These methods were found to be suitable to images
Page 2 of 18
3 having uniform intensity distribution. Chen and Ramli [8] proposed Recursive Mean Separate Histogram Equalization (RMSHE) method which partitions histogram of source image recursively [8] and performs enhancement of all segments independently. Computational cost was clearly a disadvantage of this method. Few weight based equalization methods like weighted threshold HE (WTHE)[9] had also been proposed to add adaptivity and ease of control to enhancement process. These methods tend to modify probability density function by suitable weights before equalization and performed normalization afterwards with addition of some adjustment factor. Weight Clustering HE (WCHE) [10] was also developed which had its basis on adaptive
ip t
change of weights.
There had been attempts by researchers to utilize use of soft computing techniques for controlling enhancement process.
cr
Few of these have been summarized in this section. Debdoot et al. [11] proposed modified version of HE based on fuzzy statistics of digital images known as Brightness Preserving Dynamic Fuzzy Histogram Equalization (BPDFHE) to handle
us
inexactness of gray level values. Another fuzzy logic based technique (fuzzy histogram equalization-FHE) was proposed by Magudeeswaran and Ravichandaran[12] for improvement of contrast. The approach generates fuzzy histograms for original image using fuzzy set theory and follows it with segmentation of image based on median. The sub-images were then
an
independently equalized to preserve brightness. Fuzzy logic has been exploited by Hammadulu et al. in [13][14] for proposing solutions for contrast enhancement of color images. HE has also been implemented using artificial neural network based studies and Thomas H. Hildebrandt[15] were the first to propose and implement a circuit for generalized histogram
M
equalization. Another Local HE based technique was proposed by Chitwong et al. [16] who had utilized efficacies of Hopfield networks. The method uses fuzzy c-means clustering and competitive hopfield neural network for segmentation of image while performing equalization of both components individually. The technique of Kartik et al.[17] first utilizes various statistical
d
operations to perform initial enhancement of image. It then uses two right most bits of both original and enhanced image as
te
inputs and target of designed neural network. Evolution based studies like genetic algorithm(GA) and particle swarm optimization(PSO) have also been used by various researchers to propose challenging solutions for contrast enhancement of images. In order to enhance contrast of grey scale images, problem of image enhancement has been framed as an optimization
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
problem by Apurva & Ashish[18]and PSO has been utilized to attain maximum value for an edge based contrast of image. P. Shanmugavadivu et al. in [19], [20] had used concept of histogram segmentation and multi objective functions for enhancement of images. Munteanu et al. [22] had proposed application of transformation function to each pixel of input image in which GA was used to extract parameters of transformation function. Saitosh et al. [23] used GA to represent relation between input and output gray levels using lookup table. Chromosomes were evaluated using edge intensities. Changiiang and Xiaodong [24] employed In-complete Beta Transform (IBT), in which Genetic Algorithm (GA), and Wavelet Neural Network (WNN) were used to enhance contrast for an image. In this method, a non linear transform curve is obtained using In-complete Beta Transform. IBT in the whole image is approximated using a new kind of WNN. The task of GA was to determine optimal gray levels transform parameters. A classification criterion was used to determine original contrast and discrete stationary wavelet transform (DSWT) was used to enhance local contrast of image. S. Hashemi et al. [25] method used a simple and novel chromosome representation together with corresponding operators. The method was based on a simple chromosome structure to overcome shortcomings of previous methods. Two more recent studies based on use of genetic algorithm are available at [26-27].PSO based studies are simple and provides faster convergence as compared to GA based studies which tend to make use of various operators like mutation, crossover selection etc. to generate effective solutions. Readers can refer to PSO base studies at [18-22] and GA based studies [22-27] for detailed implementations.
Page 3 of 18
4 The present study tend to utilize efficacies of new optimization algorithm named as Water Cycle algorithm to propose an adaptive solution for contrast enhancement using modified version of histogram equalization. The approach has been tested qualitatively and quantitatively and is also subjected to competition with contemporary technique. The next section introduces complete methodology of the proposed technique. III. HISTOGRAM EQUALIZATION WITH COLOR RESTORATION
ip t
The performance of almost all HE based technique degrades when a more natural and complex looking image is processed. HE based techniques suffer from problem of mean shift, preservation of details, over enhancement and loss of colourfulness for
cr
colored images. The main objective of this study is to modify the basic HE principle to generate a visually pleasing image and control effectiveness of contrast with help of water cycle based optimization algorithm. Figure 1 presents the major steps of the proposed technique. All these steps are explained in details for better understanding of readers.
us
Pre-processing: In the proposed method, we first transfer the video/image sequence to YCbCr color model and extract the luminance(Y) component to reduce computational complexity. Y component is more prone to dehazing or other artifacts since
an
it is very similar to gray scale version of image. The Y channel contains rich boundary details which cause redundant information to flow to next phase and therefore smoothening of input frame has been done using 3x3 Gaussian filter. Smoothed Y channel has been used in this method to reflect uniform distribution of real scene.
M
Segmentation of input image using Ostu’s 2D threshold: Thresholding divides an input image histogram into foreground and background which can further be equalized separately in order to improve contrast of individual regions. Ostu’s method [28] is one of the most commonly used approaches to automatically segment image into two distinct classes. Ostu’s algorithm
d
searches for threshold that maximizes inter class variance between foreground and background defined as weighted sum of
where
and
te
variances of the two classes:
whole image is given by A(Y).
(4)
means average brightness of background and foreground thresholded by T, mean brightness of
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
and
are cumulative probabilities of background and foreground respectively defined as (5) (6)
Otsu’s method exhibits relatively good performance for bimodal histogram distribution but is not able to generate an optimal threshold when objects in frame are of small size, or if variances of object and background intensities gets big, or if image is too noisy. Due to one-dimensional Otsu algorithm's poor adaptability to noise image segmentation, a 2D Otsu algorithm [29] was introduced. 2D Ostu algorithm has been widely applied for segmenting images because of its content independent attributes. 2D Ostu algorithm makes use of values of neighbours together with its own value to extract an optimal threshold for improvement of binarizaion results. This makes it suitable to segment those images with small object and images with lots of noise. 2D ostu’s method works by forming a pair of pixel gray level(i) and average of its neighbours(j); and each such pair belongs to a 2 dimensional bin with total number of bins equal to L x L, where L is number of grey levels and neighbouring grey values. The joint probability mass function of such 2D histogram may be defined as number of occurrences f(i,j) of pair (i,j) divided by total number of pixels in image (MxN). (7)
Page 4 of 18
5
where
te
d
M
an
us
cr
ip t
(8)
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Figure 1: Proposed methodology for contrast enhancement
Let segmentation to be [S,T] and assuming that 2-D histogram consists of only two regions namely foreground and background, these regions will have different probabilities specified as (9) (10)
Mean vectors corresponding to two regions are given as (11) (12) Inter class variance of these regions can be specified as (13) The optimal threshold could be found by maximizing interclass variance of two classes given as and is given by (14) The input image is segmented into foreground (IF) and background(IB) sub images with help of
. The threshold generated
Page 5 of 18
6 by 2D ostu’s method will provide a better segmentation as compared with standard ostu’s method for images with multimodal distribution and noise. The readers may refer to study [29] for further details. Modification of Background and Foreground sub-images: Probability distributions for foreword sub-image(IF) and background sub-images(IB) obtained in previous step are calculated as
and
. These distributions are subjected to
suitably defined constraints for background and foreground respectively. The modified probability density function for and back- ground
) frames are specifed as
ip t
foreground (
cr
(15)
an
us
τ
(16)
Mean of both modified probability distribution and normal pdf is calculated for both foreground and background sub images. The relative difference among both is added to and
for
background
and
foreground
is
. The cummulative distribution functions calculated
as
M
using
and
and
and new historam equalization procedure is applied according to (17)
d
(18)
te
These constraints are defined so that effect of dominating intensities could be controlled and effect of intensities with small PDF could be ignored. For this reason PDFs of both foreground and background are clamped with respect to certain threshold values specified in equation 15 and 16. Values of controlling parameters (r,t) are fixed in the range from 0.1-1.0 as over
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
enhancement creeps in beyond this range. The degree of enhancement is controlled by power parameters (q,s), less probable levels in the corresponding sub-images can be reallocated with more probable levels when the value of these parameters is < 1.0. It helps in preservation of visual details of the image. When value of power factor approach 1.0, the procedure behaves like standard histogram equalization, when value goes beyond 1.0, more weight is added to high probability levels and these constraints impose an even stronger effect than HE which may result in over enhancement. The resolved constraints are applied to both sub-images and both are combined to generate new image I
new
which is passed to next step for constraint
optimization. The optimization of parameters is done with water cycle algorithm as explained in next section. A. Water cycle algorithm based optimization of constraints parameters For implementation of histogram equalization with color restoration we had defined four variables namely (q,r,s,t) that will help to control the effects of standard HE and the values of these parameters range from 0.1-1.0. This section discusses how water cycle algorithm has been used to extract an optimal combination of these parameters. Water Cycle Algorithm Water Cycle Algorithm(WCA) is new population based meta-heuristic algorithm based on the observation of nature’s water cycle and finds solutions mimicking flow of rivers and streams toward sea [30] [31]. WCA creates an initial population of design
Page 6 of 18
7 variables or in other words population of streams (raindrops) after assuming rainfall or precipitation. The best individual (best raindrop) i.e. one having least value of cost function is chosen as a sea. Then, a number of good raindrops i.e. solutions having cost function closest to current best are chosen as rivers and rest of the raindrops are considered as streams which flow to the rivers and sea. In order to simulate this effect an initial population with dimensions Npop × D is created where D represents
cr
ip t
number of design variables and Npop is population size.
an
us
(19)
Out of total of Npop streams created initially, Nsr represents best individuals from population designated as sea and rivers. The rest of population NStream are considered as streams flowing into rivers or into sea directly. Depending on their magnitude
M
of flow each river absorbs water from the streams and amount of water in a stream entering a river and/or sea varies from other streams. Sea is assumed to be most downhill location and rivers flow to the sea. Figure 2 illustrates complete WCA optimization process.
where
te
given
d
In order to designate/assign raindrops to the rivers and sea depending on the intensity of the flow, the following equation is
is the number of streams which flow into the specific rivers and the sea and
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
(20) is the cost of respective raindrop.
Flow of rivers to sea and streams to rivers can be modelled as
where C is number between 1 and 2. Initial value of X may be in the range
(21) (22) where d is current distance
between stream and river and X is randomly chosen distance to depict flow of streams to river along the connecting line(Fig. 2). Evaporation process(Exploitation) is also modelled to prevent premature convergence to local minima. Evaporation may lead to more precipitation and algorithm checks whether river/stream is sufficiently close to sea to enable the evaporation process to occur. Following criterion is utilized for the evaporation condition between a river and the sea: (23)
where dmax is a small number close to zero and value of dmax adaptively decreases as follows: (24) After evaporation, the raining process is applied and new streams are formed in different locations (similar to mutation in the GAs). In the raining process, new raindrops form streams in different locations (acting similar to mutation operator in GA).
Page 7 of 18
cr
Figure 2: Water cycle optimization process (a) Flow of streams to rivers (b) WCA optimization process [30]
(25)
us
For specifying locations of the newly formed streams, the following equation is used:
where LB and UB are lower and upper bounds defined by given problem respectively. Termination condition of WCA may be
refer to [30] [31] for more details about water cycle algorithm.
an
fixed as maximum number of iterations or small tolerance value between consecutive iterations. The readers are advised to
Objective function for WCA: The main aim of proposed technique is to enhance contrast of input frame while preserving
M
brightness of source image. Two different functions have been used for implementing this aspect where one will try to maximize contrast and other will be used to preserve brightness. For preservation of brightness we propose difference in
d
average intensities of enhanced and original image as one of the function which can be calculated as (26) (26)
te
Several methods are available to measure contrast like discrete entropy(DE), contrast improvement index, edge based contrast etc. In this paper we had chosen a modified contrast index used by Y. Wang et al.[32] in which contrast of image is measured by combining different performance measures like entropy, sum of edge based intensities and number of edge pixels.
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
ip t
8
where I is image of M x N pixels;
(27)
is visible edge image of I which can be computed as [33]. E(.)is sum of internsities of
visible image, ne represents number of edges and H represents entropy of image. In order to preserve colorfulness of the images, a parameter that measures colorfulness of image is used(CCI). CCI reflects the richness and vivid degree of color and is measured by using work of [34]. In order to maintain colorfulness of frame, difference of original and enhanced frame is taken as third objective function to optimize. (28) The three metrics can be combined together to get a single function as (29) In order to make final image look natural and rich in contrast with preservation of colorfulness and brightness it is desirable to maximize value of E and therefore we had employed equation 29 as objective function for our problem.
and
are weighing
factors for controlling levels of brightness and color component enhancement of image. The ideal values for these weights shall be small otherwise the output image will be over enhanced. The values of these weights were set experimentally depending on
Page 8 of 18
9 nature of input image. The complete optimization process is implemented using water cycle algorithm which is detailed in the next section. B. Water Cycle Algorithm Based on above observation, we tend to formulate the problem of searching for optimal parameters as maximization of equation for E which could be stated as follows out
(30)
such that 0.1
ip t
(q*,r*,s*,t*) = arg max E (I (q,r,s,t))
This section illustrates how the problem of parameter detection has been solved using water cycle algorithm. The value of
cr
proposed objective function(E) is combination of three individual functions providing modeling of contrast, brightness and color. Since value of E is affected by parameters (p,q,r,s) and it is empirically difficult to choose an optimal combination for
us
which equation(30) will be maximized, authors tend to seek help of water cycle algorithm for solving this optimization problem. Figure 3 introduces complete algorithm to perform enhancement of input image or frame. 1. Input an image or frame of video and convert it to YCbCr color space.
an
2. Extract Y channel and use 3x3 Gaussian mask to smoothen the input image.
3. Use 2-D ostu’s thresholding defined in section 3 to extract appropriate threshold value (
M
14.
for the frame using equation
4. Apply threshold generated in previous step to extract two sub images from original image as foreground(IF) and background(IB).
d
5. Call Water cycle based algorithm for finding optimal values for (q, r, s, t)
q,r,s,t), LB =0.1, UB=1.0 etc.
te
a. Initialize WCA parameters: population size(N=30), Nsr=6,dmax,max_it(40), number of parameters (nvars =4 for
b. Generate random population (equation 19), subject to condition 0.1
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
c. For all population species, extract new probability density functions for foreground and background according to equations 15 and 16. d. Perform equalization of foreground and background separately according to equation 17 and 18. e. Find values of individual objective functions and combine these to calculate the cost of population species according to equation 29. f. Sort the population according to cost and assign first/best candidate as sea, perform initialization of rivers and streams accordingly.
Page 9 of 18
10 g. Designate streams to rivers and sea according to equation 20. h. Model flow of streams to rivers and rivers to sea according to equation 21 and 22. i. Exchange positions of river with stream which gives best solution and exchange position of sea with river if better solution is available(Figure 2). j. Perform evaporation process according to equation 23 and in case there exists evaporation perform raining process
k. Update dmax using equation 24.
ip t
according to 25.
l. Check for convergence, if stopping criteria is met return value of sea as best solution else return to step 5(g).
cr
6. Use the optimal values returned by WCA to generate enhanced Y channel image.
7. Combine CbCr channel information into enhanced Y channel and convert back to RGB space.
us
Figure 3: Proposed water cycle algorithm based histogram equalization method
The algorithm given in figure 3 inputs a hazy image or frame in first step and as part of pre-processing operation extracts Y channel of image after converting from RGB space to YCbCr color space. The input image is segmented into foreground and
an
background sub images using 2-D ostu’s thresholding principle. 2-D ostu’s method has been utilized as it is more suitable to complex or natural images. Both sub images are enhanced separately by subjecting respective probability density functions of foreground and background to specifically designed constraints. The designed constraints can be controlled with four
M
parameters (q,r,s,t) as illustrated in previous sections. Different combinations of these parameters will constitute for different enhanced versions of final image and in order to select an optimal value, water cycle algorithm has been used for this purpose.
d
WCA first generates a random population of raindrops(stream) within fixed lower and upper bound. The algorithm then proceeds to initialize sea, rivers and streams from initial population. Flow of streams to rivers and rivers to sea is modelled for
te
fixed number of iterations and best solutions are re-designated as sea or rivers. The objective function for water cycle optimization is considered to be a three dimensional function that tends to enhance contrast and preserve brightness and colorness of input image. The best solution is utilized for modification of foreground and background sub-images which are
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
eventually combined to generate final enhanced Y channel image. The CbCr components are added to the enhanced Y channel image to get the final image.
IV. RESULTS AND DISCUSSION
In order to test performance of proposed method, experiments were conducted on number of hazy and unclear images downloaded from [35]. In addition, the proposed method was also tested on hazy videos used by [36], object detection videos from [37]. The frames or images were subjected to proposed approach for 40 iterations and best results obtained are reported in this section. In order to compare performance of proposed algorithm, same images were enhanced with contemporary techniques like Histogram Equalization(HE), BBHE(Bilevel histogram equalization), CLAHE(contrast limited histogram equalization) and RMSHE. The performance of all these methods has been evaluated in terms of human visual perception and some qualitative measures like entropy, edge based contrast, colorfulness, contrast improvement index(CII), MSE and PSNR. Although tests were conducted on all sorts of images we hereby demonstrate only color, complex and natural images except lenna image (figure 4) which has been included for completeness. Figures 4-9 presents visual results of some sample images on which tests were conducted. Enhanced images obtained by HE, BBHE, CLAHE, RMSHE and proposed method are shown in corresponding images from 4(b)-(f) to 9(b)-9(f) respectively.
Page 10 of 18
11
a
b
c
d
e
f
b
c
d
e
us
a
cr
ip t
Figure 4 (a)Original Lenna image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
f
Figure 5 (a) Original camel image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f)Proposed
HE is not able to preserve original color of image and over-enhances can be seen in figure 4(b); the results obtained by CLAHE
an
and RMSHE are almost same but improved results can be seen for BBHE (fig 4(c)). The proposed method was not only able to provide a contrast rich image but color and details of image were also preserved. The final obtained image is more natural and
M
colourful. In Figure 5, we can see that none of the contemporary technique was able to perform enhancement effectively. Artifacts can be observed in images enhancements by HE, BBHE, CLAHE and RMSHE where as image shown in fig 5(f) is more natural and bright. The color of camel seat has been turned to black by HE, BBHE and RMSHE. The results of CLAHE are better
d
than these methods however CLAHE was not able to perform enhancement of sand and sky regions that look far better in proposed method. In Figure 6, we see that original color of sand and grass are lost after performing enhancement using HE,
te
BBHE, CLAHE and RMSHE methods. Some white spots can be observed on portions of image covering bench in HE, BBHE and RMSHE but no such area can be seen in final image generated by proposed technique. The enhanced image looks rich in contrast and is able to remove haziness from background in a far better way than any of the other methods. In Figure 7, a
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
redundant bright spot has been generated in background for HE and BBHE methods. In addition HE, BBHE and RMSHE were not able to perform enhancement of lower sub areas of image as highlighted in circles. Results of CLAHE and proposed method are comparable however proposed method is more successful in preserving brightness of original image. In figure 8, we can easily observe that proposed method was able to preserve details of aircraft while others have failed completely. This is another demonstration that proposed technique has successfully performed enhancement of image while preserving color, brightness and details of original image. In figure 9, contemporary techniques like HE, BBHE, RMSHE failed to provide enhancement in regions that are far away from the source, in addition color information of truck is also lost in HE and BBHE methods. Results of CLAHE seem to be better but whiteness and artifacts are visible in portions of image which is encircled. The proposed method was also tested for enhancement of video frames on hazy video sequences and moving objects videos from CAVAIR dataset. The visual results are shown in figure 10 and 11 of frame number 55 (hazing road) and frame number 3(video Br3 of CAVIAR). The results of the proposed method are far better than all the contemporary technique which justifies applicability of proposed technique to videos also.
Page 11 of 18
12
b
c
d
e
f
b
us
cr
ip t
Figure 6 (a) Foggy bench image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
d
c
e
f
M
an
Figure 7 (a) Tienamen image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
c
d
e
d
b
f
te
Figure 8(a) Aircraft image (b) HE (c) BBHE (d) CLAHE (e) RMSHE (f) Proposed
Ac ce p
1 2 3 4 5 6 7 8 9 a 10 11 12 13 14 15 16 17 18 19 20 21 22 a 23 24 25 26 27 28 29 30 31 32 33 a 34 35 36 37 38 39 40 41 42 43 44 a 45 46 47 48 49 50 51 52 53 54 a 55 56 57 58 59 60 61 62 63 64 65
b
c
d
e
f
Figure 9 (a) Foggy Bus image (b)HE (c)BBHE (d)CLAHE (e) RMSHE (f)Proposed
b
c
d
e
f
Figure 10 (a) Frame #55 of hazing road (b)HE(c)BBHE (d)CLAHE (e) RMSHE (f)Proposed
Page 12 of 18
13
a
b
c
d
e
f
Size
Original
HE
BBHE
512 x 512
5.51955
7.07394
7.14725
Tienamen
600 x 450
7.64235
7.22936
7.30205
foggy bench
800 x 600
7.41689
7.36125
foggy bus
820 x 553
7.10801
7.29976
3096_airplane
481 x 321
6.23429
6.88136
Camel
481 x 321
6.92767
mountains
481 x 321
7.17124
Haze
400 x 300
7.26217
Foggy Trees
1076 x 723
7.42706
Frame 55
--
7.2686
Frame #3
--
7.2608
CLAHE
RMSHE
Proposed
6.83556
6.24508
7.21614
7.60426
7.68047
7.69592
7.39041
7.48466
7.8954
7.76138
7.4575
7.39229
7.54153
7.43937
6.98299
6.95618
6.99534
6.55078
7.41832
7.42756
7.55185
7.78121
7.52567
7.21629
7.23563
7.3936
7.36526
7.29593
7.33521
7.46763
7.31939
7.33572
7.54303
7.3476
7.31067
7.72052
7.79199
7.27353
7.0578
7.3737
7.5412
7.3526
7.3645
7.3564
7.4908
7.5527
7.7714
7.7812
te
d
an
us
Lenna
M
Image
cr
TABLE 1: COMPARISON OF DISCRETE ENTROPY VALUES
ip t
Figure 11 (a) Original Frame #3 of Br3 (b)HE(c)BBHE (d)CLAHE (e) RMSHE (f) Proposed
TABLE 2: COMPARISON OF CONTRAST VALUES
Image Lena
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Original
HE
BBHE
CLAHE
RMSHE
Proposed
85
41.81753
43.41483
54.66605
53.6942
57.53099
50.2959
40.49902
44.82648
48.77289
45.913
51.12503
53.25074
43.39474
52.37272
43.39441
54.8284
55.08023
7.10801
7.29976
7.4575
7.39229
7.54153
7.4393
3096_airplane
39.25515
42.18912
51.00271
48.18964
49.1758
48.579
Camel
13.16601
36.04692
30.37829
25.40675
29.32267
47.37631
Mountains
38.61747
40.7504
37.4932
42.65657
38.35404
45.3334
Haze
47.66814
41.47838
48.30346
53.19276
46.33241
55.1931
Foggy Trees
19.60666
46.0364
30.5084
38.82107
30.92912
42.5789
Frame 55
40.8643
40.2576
40.1619
39.8530
45.3518
43.8153
Frame #3
32.4323
36.7071
32.8179
39.412
33.1266
39.487
Tienamen foggy bench foggy bus
Page 13 of 18
14 TABLE 3: COMPARISON OF COLORFULNESS VALUES Image
Original
HE
BBHE
CLAHE
RMSHE
Proposed
17.2714
17.2619
16.5929
17.2256
17.26335
18.00179
Tienamen
31.84715
31.82578
33.4687
31.8983
31.88933
31.924
foggy bench
16.3323
15.1656
15.5917
15.0338
16.06058
16.29719
foggy bus
7.50751
8.07108
8.07778
8.11048
8.11322
8.17545
12.5687
11.9457
12.5974
12.5517
12.4877
12.57478
Camel
39.9830
36.4172
37.0909
39.8726
39.9964
40.03432
Mountains
41.7874
42.8050
42.7777
41.9980
41.8426
42.97118
Haze
11.2910
13.1564
13.0575
11.5282
Foggy Trees
20.6650
21.7223
21.9232
21.0708
Frame 55
14.9522
15.0578
15.870
15.0322
14.8658
15.7904
Frame #3
18.1431
16.3281
16.684
17.5015
17.3402
17.5755
cr 11.5338
11.53948
20.7415
21.93014
us
an
3096_airplane
ip t
Lenna
TABLE 4: COMPARISON OF CII VALUES HE
BBHE
1.4521
1.35789
Tienamen
1.1689
1.1789
foggy bench
1.5781
0.9874
foggy bus
0.94785
3096_airplane
1.0178
Camel
0.1897
Mountains
0.00456
Haze
2.417621
RMSHE
Proposed
1.566949
3.49328
2.013815
1.181047
2.134998
2.12056
2.054997
2.37136
0.9268
1.302322
0.85533
1.443819
1.2713
1.561
1.705542
1.344101
0.21056
2.166294
1.797936
2.602408
0.00712
0.189981
0.121619
0.144246
te
d
Lenna
CLAHE
M
Image
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
0.0896
0.10489
2.064851
1.226329
2.001621
Foggy Trees
0.03147
0.03258
1.396552
1.678399
1.672023
Frame 55
0.00124
0.00256
0.0897
0.0344
0.1025
Frame #3
0.0192
0.0574
1.6982
1.69054
2.1789
The qualities of test images enhanced using various algorithms were also measured in terms of discrete entropy, edge based contrast, contrast improvement index, colorfulness, MSE and PSNR values. Tables 1 through 6 demonstrate results obtained using these algorithms for images and videos. In Table 1, we can see that proposed method either led to increase in entropy of an image or maintain the same degree of entropy values. This means that details of source image were preserved by the proposed objective function. The effectiveness of proposed technique is also justified when we observe the values obtained for contrast and colorfulness of various test images demonstrated in table 2 and table 3. The proposed method is efficient in terms of improving contrast for almost all images barring a few in which values observed are rather competitive. The proposed method has also justified that it is able to preserve colorfulness of original image as obtained results are quite promising.
Page 14 of 18
15 TABLE 5: COMPARISON OF MSE VALUES HE
BBHE
CLAHE
RMSHE
Proposed
Lenna
132.5898
124.3801
98.82253
0.39547
92.5898
Tienamen
141.6714
113.3389
49.79701
146.0021
36.81183
foggy bench
153.7654
112.5811
160.3484
103.735
117.9389
foggy bus
110.9421
111.0487
101.3295
84.71749
86.3038
3096_airplane
108.5361
80.846
26.72487
67.24598
26.73755
Camel
88.74955
114.2815
49.70072
87.68683
14.22405
Mountains
67.20057
110.2334
45.5466
28.22699
57.80656
Haze
131.6577
86.45403
23.69154
Foggy Trees
56.75089
127.6994
32.06301
cr
ip t
Image
15.70304
113.961
50.5162
us
26.67682
BBHE
Lenna
26.9397
27.21729
Tienamen
26.65202
27.62104
foggy bench
26.29622
27.65015
foggy bus
27.71385
3096_airplane
CLAHE
Proposed
28.21624
52.19731
28.46517
31.19285
26.52125
32.47093
26.11417
28.00555
27.41423
27.70968
28.10757
28.88507
28.7705
27.80936
29.08868
33.89567
29.88823
33.85959
Camel
28.68314
27.58505
31.2012
28.7355
36.60057
Mountains
29.91631
27.75003
31.58071
33.65871
30.51103
Haze
26.97034
28.79695
34.41893
33.90347
36.17097
Foggy trees
30.62668
27.10305
33.10642
27.59737
31.0965
M
HE
te
Image
an
RMSHE
d
TABLE 5: COMPARISON OF PSNR VALUES
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Another quantization measure of contrast enhancement can be defined by contrast improvement index(CII), and its formula can be expressed by the following
where
and
(31)
are the contrasts of the processed and original images respectively. C is the average value of
the local region contrast in the processed or original image which can be calculated as (32) where Xf and Xb are maximum and minimum luminance values of foreground and background regions in local window size of 5*5. Higher value of cii represents better image quality and values of CII have been normalized according to their size in the present study. Table 4 lists values of CII obtained for various methods, value of CII for original image comes to be 1. The table depicts an increase in values of CII over other methods suggesting effectiveness of proposed algorithm . MSE and PSNR measures which have been borrowed from signal processing have also been used historically for evaluation of image quality. Although their findings cannot be considered reliable nevertheless, their use is still common practice and for the
Page 15 of 18
16 sake of completeness these have also been included in our study. The mean square error and PSNR between two images of size(MxN) can be defined as: (33) (34) L reflects the range of values that a pixel can take; The values of MSE and PSNR for test images are tabulated in tables 5 and
ip t
6 respectively. A small mean square error results in a high signal to noise ratio, if MSE tends to zero, then PSNR tends to infinity. More similarity between original and final image will result in minimization of MSE parameter and enhancement of
cr
PSNR parameter and the observed values are in correlation to the values for other parameters. V. CONCLUSIONS
us
This paper has presented an evolutionary approach based on water cycle algorithm for enhancement of images and videos. The proposed method performs preprocessing of input frames and then runs a controlled version of equalization for
an
performing contrast stretching. Enhancement is controlled by utilizing powers of WCA by finding an optimal combination of variables according to specific image. Suitable membership function has been designed to prevent loss of colourfulness and truncation of pixel values in enhanced image. This has resulted in generation of images which are rich in contrast, colour and
M
free from information loss. In order to test effectiveness of proposed technique, different levels of images comprising standard, natural or complex and frames from hazy videos were used and a number of parameters were evaluated. The proposed technique was also put to competition tested with other contemporary techniques of contrast stretching and results
d
demonstrate that the proposed technique is capable of removing haze effects and preserving colourfulness and brightness of
te
original frames. REFERENCES
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
[1] Gonzalez Rafael C, Richard E. Woods, “Digital image processing”. 2nd ed. India: Prentice Hall; 2002. th
[2] P. E. Trahanias, A.N. Venetsanopoulos. “Color image enhancement through 3-d histogram equalization”, in Proc. 11 IAPR Conf. Pattern Recogn., vol. III, 1992,pp. 545–548.
[3] D. Menotti, L. Najman, A. de Araújo, J. Facon. “A fast hue-preserving histogram equalization method for color image enhancement using a Bayesian framework”, in: Proc. 14th International Workshop on Systems, Signal and Image Processing (IWSSIP), June 2007, pp. 414–417.
[4] J.H. Han, S. Yang, B.U. Lee.(2011) A novel 3-D color histogram equalization method with uniform 1-D gray scale histogram, IEEE Trans. Image Process. 2 (2) pp. 506–512. [5] Y. Kim(1997) Contrast enhancement using brightness preserving bi-histogram equalization. IEEE Tran. Consumer Electronics ;43,pp. 1–8. [6] Y. Wan, Q. Chen, B. Zhang. (1999). Image enhancement based on equal area dualistic sub-image histogram equalization method. IEEE Tran. on Consumer Electronics 45, pp 68–75. [7] S. Chen, A. R. Ramli.(2003). Minimum mean brightness error bi-histogram equalization in contrast enhancement. IEEE Tran. on Consumer Electronics 49,pp. 1310–29.
Page 16 of 18
17 [8] S. Chen, A. R. Ramli.(2003) Contrast enhancement using recursive mean-separate histogram equalization for scalable brightness preservation. IEEE Trans. Consumer Electronics. 49, pp 1323–32. [9] Q. Wang, R. K. Ward.(2008) Fast image/video contrast enhancement based on weighted thresholded histogram equalization. IEEE Trans. Consumer Electronics. 53, pp. 757–64. [10] N. Sengee, H. K. Choi.(2008). Brightness preserving weight clustering histogram equalization. IEEE Transactions on Consumer Electronics. 54, pp. 1329–37.
ip t
[11]D. Sheet, H. Garud, A. Suveer, M. Mahadevappa, J. Chatterjee.(2011). Brightness preserving dynamic fuzzy histogram equalization. IEEE trans. consumer electronics. 56(4), pp. 2475-2480.
cr
[12] V. Magudeeswaran, C. G. Ravichandran.(2013). Fuzzy Logic Based Histogram Equalization for Image Contrast Enhancement. Mathematical Problems in Engineering. Available: http://dx.doi.org/10.1155/2013/891864
us
[13] H. Madasu,D. Jha.(2006). An optimal fuzzy system for color image enhancement. IEEE Transactions on Image Processing. 15, pp. 2956–66.
[14] M. Hanmandlu, O. P. Verma, K. K. Nukala, M. Kulkarni.(2009)A Novel Optimal Fuzzy System for Color Image Enhancement
an
Using Bacterial Foraging. IEEE trans. instrumentations and measurement.58 (8), pp. 2867-2879. [15] T. H. Hildebrandt.(1993) A Local Neural Implementation of Histogram Equalization. Available: http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=298809
M
[16] S. Chitwong, T. Boonmee, F. Cheevasuvit. “Local area histogram equalization based multispectral image enhancement from clustering using competitive Hopfield neural network” in Can. Conf. Electrical and Computer Engg. June 2003, pp. 1715 1718 vol.3 ·
d
[17] K. Ch. Jena, N. Dash, N.K. Kamila. (2015). A neural network based approach for Image Enhancement using Histogram
te
Equalization. International Journal of computer Systems. 02(11), pp. 460-463 [18] A. Gorai, A. Ghosh. “Gray-levelimage enhancement by particle swarm optimization”. in proc. of world Cong. on Nature and Biologically Inspired Computing 2009, pp. 72–77.
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
[19] P.Shanmugavadivu, K.Balasubramanian.(2014) Particle swarm optimized multi-objective histogram equalization for image enhancement, J. optic and laser technology, 57, pp. 243-251. [20]P. Shanmugavadivu, K. Balasubramanian, Muruganandam, (2014). Particle swarm optimized bi-histogram equalization for contrast enhancement and brightness preservation of images. A. Vis. Comput. 30, 387-392. doi:10.1007/s00371-013-0863-8 [21] S. M. W. Masra, P. K. Pang, M. S. Muhammad, K. Kipli. “Application of Particle Swarm Optimization in Histogram Equalization for Image Enhancement”. In proc. IEEE Colloquium on Humanities, Science & Engineering Research (CHUSER 2012), December 3-4, 2012, Kota Kinabalu, Sabah, Malaysia, pp. 294-299 [22] C. Munteanu, A. Rosa, “Towards automatic image enhancement using genetic algorithms”. in proc. of the Congress on Evolutionary Computation, vol. 2, pp. 1535–1542. [23] F. Saitoh. “Image contrast enhancement using genetic algorithm”. In proc. IEEE Internat. Conf. Syst. Man Cybern., 4 (1999), pp. 899–904 [24] Z. Changjiang, W. Xiaodong. Global and local contrast enhancement for image by genetic algorithm and wavelet neural network. LNCS Book Series, vol. 4234, Springer (2006) pp. 910–919 [25] H. Sara, K. Soheila,N. Navid, E. M. Mohsen.(2010). An image contrast enhancement method based on genetic algorithm. Pattern Recognition Letters, 31pp. 1816–24. doi: 10.1109/ICDIP.2009.87
Page 17 of 18
18 [26] P. Babu, V. Rajamani. Contrast Enhancement using Real Coded Genetic Algorithm Based Modified Histogram Equalization for Gray Scale Images.2015. Int. J. Imaging Syst. Technol. (25), pp. 24-32. [27] S. Sedighi, M. Roopaei, S. Agaian. “Genetic-Based Thresholds for Multi Histogram Equalization Image Enhancement” in proc. Of 12th Int. Conf., MLDM 2016, New York, USA, July 16-21, pp. 483-490 [28] N. Otsu (1979). A threshold selection method from gray-level histograms. IEEE Trans. Sys., Man., Cyber.9 (1), pp. 62– 66. doi:10.1109/TSMC.1979.4310076.
ip t
[29] J. Liu, W. Li. (1993) The automatic thresholding of gray-level pictures via two-dimensional Otsu method, Acta Automatica Sinica 19 (1), pp. 101–105.
cr
[30] H. Eskandar, A. Sadollah, A. Bahreininejad, M. Hamdi.(2012) Water cycle algorithm – A novel metaheuristic optimization method for solving constrained engineering optimization problems. Computers and Structures (110–111), pp. 151–166.
us
[31] A. Sadollah, H. Eskandar, H. M. Lee, D. G. Yoo, J. H. Kim. (2016). Water cycle algorithm: A detailed standard code. SoftwareX(2016), http://dx.doi.org/10.1016/j.softx.2016.03.001.
[32] Y. Wang, H. Wang, C. Yin, M. Dai.(2016). Biologically inspired image enhancement based on Retinex., Int. J.
an
Neurocomputing. (177), pp. 373-384. http://dx.doi.org/10.1016/j.neucom.2015.10.124
[33] N. Hautière, J. P. Tarel, D. Aubert, É Dumont.(2008). Blind contrast enhancement assessment by gradient ratioing at visible edges. Image Anal. Stereol. J., 27 (2), pp. 87–95
M
[34] D. Hasler, S.E. Suesstrunk, Measuring colorfulness in natural images, in: Electronic Imaging, International Society for Optics and Photonics, Santa Clara, 2003, pp. 87–95.
[35] Haze Removal using matlab. Available: https://github.com/akutta/Haze-Removal. last accessed on 11 october 2016.
te
Vis. Comm. and Image Repre. 24(3) pp. 410-425.
d
[36] J. H. Kim, W. D. Jang, J. Y. Sim, C.S. Kim.(2013) Optimized contrast enhancement for real-time image and video dehazing, J.
[37] CAVIAR: Context Aware Vision using Image-based Active Recognition. http://homepages.inf.ed.ac.uk/rbf/CAVIAR/ (2010).
Ac ce p
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65
Page 18 of 18