Performance Analysis of different Haze removal Algorithms

I. INTRODUCTION

Most of the time the quality of the outdoor image is degraded due to atmospheric weather conditions and scattering of large number fog, dust or water particles suspended above the ground. This fog or haze conditions cause a great deal of inconvenience to Military and Defence surveillance systems. This conditions directly effect the high level image processing applications such as face detection, recognition system and outdoor transportation systems. In order to improve the stability and robustness of a visual system, we require effective haze removal algorithms. The depth of the haze is unknown in haze image and it is a challenging problem for removal of an unknown haze depth. Haze removal algorithm is used to recover a better image from bad weather image and it is highly desired in computational photography applications. The air-light is used to recover image from colour shifts after first level of haze removal. There are different techniques for haze removal among which one is based on image enhancement [1] and the other one is based on image restoration. There are different image enhancement techniques for haze removal from the image such as contrast enhancement, histogram equalization and logarithmic transformation. But, these techniques are independent of haze present in the image and results a loss of information after haze removal. The physical process of the image was studied using image restoration techniques and form a degradation model. The haze free image is recovered after inverting the degradation model without any information loss. Therefore, haze removal based on image restoration is one of the research focus over a decade.

The rest of the paper is organized as follows: The detailed literature survey is discussed in section II, the basic atmospheric scattering model is discussed in section III , and section IV explains different haze removal algorithms. And then, the results and conclusions were discussed in section V,VI respectively.

II. LITERATURE SURVEY

To solve the problem of atmospheric illumination effects on outdoor captured images, haze removal algorithms were studied from decades. The main aim of haze removal algorithms is to recover better image from atmospherically degraded bad haze image. There are different ways to recover a better image from bad haze image and they are mainly divided into visible spectrum based analysis [2], [3], polarization of light [4], [5], statistical analysis [6], [7] and dark channel priori [8], [9]. Based on the previous work, haze is removed by finding the color variations of an image under various weather conditions proposed in [2] using atmospheric scattering model [10]. The clear day image was recovered from a set of two or more bad haze images by considering 3D structure of an image and scene colour variations [3]. However, these methods are used for fog and dense haze images based on the assumption that atmospheric light is independent of wavelength over a visible spectrum. But, in some haze conditions the scattering strongly depends on the wavelength of received light. Furthermore, this methods result in saturated images when scene and haze color are same using dichromatic scattering model. The haze is removed from the image by taking polarization as a clue proposed in [4]. According to this method haze due to light polarization is removed and this method is based on polarized and un-polarized properties of object radiation and natural illumination of scattered light respectively. Based on many works [5], the image is recovered by assuming the object radiation as un-polarized and scattering light as partially polarized. Based on un-polarized properties of object radiance, most of the photo enhancement software's use polarized filters for reducing haze from the landscape and sky images. According to [4], [5] only polarized filters will not provide good results for haze images and this method is limited to mist with sky illumination and dense fog conditions. Based on statistical assumption of maximizing the local contrast of an image for haze removal is proposed in [6]. According to this method the haze free image has higher contrast than haze image. But, increasing the contrast of haze image results in saturated images and this method doesn't consider atmospheric light component. A haze removal algorithm is proposed in [7] based on the statistical, correlation property of transmission rate of light and surface shading. This method is based on the assumption of uncorrelated property of target surface shading and propagation of light and derive the transmission map of light using the reflectivity of the scene. This method will provide better results for less haze images and it requires color information of an image. Basically haze image is loss of colour information.

A new haze removal algorithm is proposed in [8] based on the assumption of dark channel color priori. According to this algorithm the dark channel which contains dark or very small pixel intensity is estimated. The haze free image is obtained by combining the optical model of an image and transmission map with median filter and this method produce invalid results when object and sky are of same colour. Improved dark channel priori algorithm is proposed in [11] for eliminating the effects of He's algorithm [8]. This algorithm produces improved results by replacing median or soft matting filter with bilateral filter. An improved transmission formula and weaker method is used to eliminate the effect of dark channel priori algorithm. Haze removal algorithm based on luminance component of an image is proposed in [12]. According to this algorithm RGB image is converted into YCbCr colour space and a Multi-Scale Retinex on luminance component of an image is applied . The transmission map is obtained by using luminance component and median filter. This algorithm is suitable for gray scale images and produce color shifts for RGB images. In this paper a new haze removal algorithm is proposed by using the colour resolution function and luminance component of an image. The proposed method effectively eliminates the haze from image and colour shifts using colour resolution function. The colour resolution function is calculated based on YCbCr colour space and this method results a better Peak Signal to Noise Ratio (PSNR).

III. ATMOSPHERIC SCATTERING MODEL

Basically all work in computer vision is based on transparent medium such as air. According to this the light rays reflected from the object are received to the observer without any attenuation or scattering. The brightness of the image depends on the brightness of the actual object. Based on this, present vision systems and algorithms only work for clear day images. A study of atmospheric optics is needed for entire weather conditions such as fog, haze, rain and snow for recovering the haze free image from haze image. In this section, a lower order restoration model for haze images is discussed. According to Fig.1 the radiance received to the observer from the scene is divided into direct transmission and air-light due to water or dust particle suspended above the ground. The lower order restoration model shown in Eq.1. According to the scattering model,image J (x, y) received by camera sensor is a combination of direct transmission R (x, y) t (x, y) and air-light due to scattering (1 − (t (x, y)))A (x, y). The maximum radiance of the object is 1, then the homogeneous atmospheric transmission coefficient t (x, y) = e −βd(x,y) is transmitted along the direct transmission and remaining radiance is transmitted along the atmospheric global light A (x, y). In atmospheric transmission co-efficient t (x, y), β is called as atmospheric attenuation coefficient and d (x, y) is the distance between the object and the observer. R (x, y) is called as surface radiance vector. If the distance between the observer and the object is infinity, then the image received by the observer is only atmospheric global light (t (x, y) = e −βd = 0) and distance is very small, then the image received by the observer is absolute object radiance (t (x, y) = e −βd = 1).

IV. HAZE REMOVAL ALGORITHMS

In this paper different haze removal algorithms are discussed and along with their flow model. Most of the outdoor images are degraded due to bad weather conditions. In this weather conditions the image is partially gray and white, due to atmospheric scattering effects. This conditions bring a great deal of inconvenience to the Military and Defence surveillance systems, so the study of haze removal algorithm is very important. The main aim of haze removal algorithms is estimating the depth of transmission map. The haze free image is recovered using transmission map and global air-light.

A. He's Algorithm

He's algorithm [8] is based on the assumption of dark channel priori. According to this algorithm atleast one channel in R, G and B image contains dark pixel due to shadow of the object or tree leaves. This is one of the simple method for finding the haze free image based on transmission map obtained from dark color priori assumption. The transmission depth map is obtained by subtracting the dark colour channel from white background image and the flow model of He's algorithm is shown in Fig.2. Based on the flow model the haze free image is obtained using the refined transmission map extracted from transmission depth map using median filter and atmospheric air-light. This algorithm improves the quality and sharpness of the image. But, this algorithm will results in saturated image when object and sky are same colour.

B. Improved Dark Channel Priori Algorithm

Improved dark channel priori [11] is little modification of He's algorithm. It is also based on the concept of dark channel priori. This algorithm will eliminate effects present in the He's algorithm such as when object and sky are same color and it recover the strong edges by using bilateral filter. According to this haze removal algorithm is a combination of bilateral filter and dark channel prior. By using atmospheric model and dark channel priori, transmission map of image was derived, and then combining with gray scale image. The refined transmission map is extracted from transmission map using fast bilateral filter. The flow model of improved dark channel priori is shown in Fig.3. This algorithm greatly improves the He's algorithm [8] in Haze free regions. A improved transmission formula is used to eliminate dim present in the image. This algorithm greatly restores contrast and colour of the scene and increase the visual effects of the image. This algorithm works well for images contain a large sky region with out any distortions. To increase the adaptability and weaking the sky regions of an image weaker's method is used.

C. Haze removal algorithm based on Luminance Component

This approach is based on the observation that when the haze image transforms from RGB to YCbCr color space, and uses Multi-Scale Retinex (MSR) algorithm to the luminance component [13]. Retinex theory of images gives compensation for illumination effects. The main aim Retinex theory is to decompose a given image P into two images, the reflectance image R, and the illumination image L at each point (x, y) in the image domain, P(x, y) = R(x, y) *L(x, y). The Retinex methodology was proposed by Land's for visual system [14]. The main advantage of this decomposition includes enhancing the image edges and correcting the color of an image induced

D. Haze removal algorithm based on Luminance Component and Color Resolution Function

The image colors tend to be desaturated greyish due to original MSR.This is due to MSR applied on gray scale image and color differentiation or constancy processing on relatively small mask. Each Pixel's color is compared with the Gaussian distribution of surrounding pixels and each constant color channel produces a greyish image due to MSR. Here, a new Haze removal algorithm is implemented based on color resolution function for eliminating the effects of color constancy problem. The color resolution function is obtained using YCbCr color space. The main aim of this algorithm is estimating the transmission map. The flow model of this algorithm is shown in Fig.5. The basic effect of MSR algorithm on gray scale image produces desaturated images either globally or specific regions. A color resolution function is used to recover a good images from desaturated images by providing color rendition. The algorithm for color resolution method is given below,

is the i th band of color resolution function and R imcf (x, y) is i th spectral band of MSR. The color resolution function is combination of Y, Cb and Cr components shown below,

where

Here the color resolution function is depended on the YCbCr colour space and varies automatically based on input haze image. After applying the color resolution function on MSR image results greyish image. In order to display the image properly with out any logarithmic effects a gain G and offset b is used. The Haze free image was recovered from airlight of the image and refined transmission map extracted from the combination of color resolution function along with MSR on luminance component. This algorithm results better Peak Signal to Noise Ratio (PSNR) then compared to other methods and this method is effectively eliminating the color shifts problem present in the previous algorithm [13].

V. RESULTS AND DISCUSSIONS

In this section, we will evaluate the performance of the proposed method on different test images roadways and buildings using OpenCV. The roadways shown in Fig.6 contains a strong depth of haze with fog and buildings shown in Fig.7 contains small depth of fog and haze. The proposed method is compared with other haze removal algorithm such as He's algorithm [8], Improved dark channel priori [11] and haze removal algorithm based on luminance component of an image [13] using image quality assessment techniques.

Basically, image quality measures are classified as objective quality measures based on statistical analysis and subjective quality measures based on human visual system observation. Most of the time the human eye is more sensitive to edges rather then information present in the image. The subjective quality of the image is based on human observation. The performance of different haze removal algorithms based on subjective quality measure are verified by taking different observations. Based on this observation the human will provide a rating to images as best, good and worst. The subjective quality measures of different haze removal algorithms for roadways and buildings shown in Fig.6 and Fig.7 respectively. In Fig.6 and Fig.7 a) haze image contains strong depth of haze, b) output haze free image using He's algorithm [8], c) output haze free image using improved dark channel priori algorithm [11], d) output haze free image using luminance component based haze removal algorithm [13] and f) output of proposed method. It clearly shows that proposed method gives higher PSNR (peak Signal to Noise Ratio) and good quality image compared to other methods. The different haze removal algorithm are analysed using the objective quality measures of an image such as average gradient (AG), contrast of an image Table 1 shows the performance analysis of different haze removal algorithm for images shown in Fig.6. According to table 1, proposed method results higher PSNR, less error and less number of dark pixels compared to all other methods. The running time of proposed method is more compared others. The image fidelity is more and peak mean square error is less for proposed method compared to other methods. Table 2 shows the performance analysis of different haze removal algorithm for images shown in Fig.7. Based on this object quality measures, proposed method results good PSNR compared to all other methods. The contrast of image obtained from proposed method is sightly more than other methods. Proposed method results a less artefacts and results a less structural content of an image compared to other haze removal algorithms. The proposed method results less errors and high cross correlation coefficient value and more running time compared to other methods. Based on the performance analysis of table 1 and table 2, these haze removal algorithm are used for different applications by neglecting the some of the image quality measures. He's algorithm is used for low processing power applications with out considering the structural content of an image. The haze removal algorithm based on luminance component is used for medium processing applications for the requirement of less errors. Applications require more PSNR and less structural content of an image with out considering the running time of the algorithm, for this applications haze removal based on luminance component and colour resolution function will provide better results. Improved dark channel priori algorithm is used for medium processing applications for the requirement of better PSNR, strong edges, more structural distortion measure and less structural content of an image. Each haze removal algorithm is used for different applications based on the requirement of the user. Each and every algorithm has different applications, the selection of algorithm is based on image objective and subjective quality measures.

VI. CONCLUSION Haze removal algorithms plays a major role in the defence, military surveillance systems. The study of haze removal algorithm is very important for outdoor scenarios. Haze removal algorithm based on luminance component and colour resolution function provides better results compared to other methods. The proposed methods provide good PSNR and less error compared to other methods. The performance of different haze removal algorithms are verified using the subjective and objective quality measures of an image. Based on the experimental results proposed method results better performance than the other haze removal algorithms. Each and every algorithm has specific applications and selection of particular algorithm is based on objective quality measures of an image.

APPENDIX

Number of dark pixels present in the restored imagê C K (i, j) can be obtained by finding the number of pixels having intensity less than 10.

The number of dark pixels are more the quality of the image is less.

Contrast of the restored imageĈ K (i, j) is given by

Here L max and L min are the maximum and minimum luminance values of a restored image and maximum value indicates image is more quality Average gradient of the restored imageĈ K (i, j) is given by [15]

Here G x and G y are the gradient in x, y directions respectively and higher value indicates image has good quality Normalized Cross Correlation (NCC) [15] between restoration image C K (i, j) and original imageĈ K (i, j) is given by K is for number of components Ex. RGB and M, N are the frame height and width respectively and higher value indicates the image is more quality Normalized absolute error (NAE) between restoration image C K (i, j) and original imageĈ K (i, j) is given by

Here M, N are the frame height and width respectively. The large value of Normalized Absolute Error (NAE) means that image is poor quality. Normalized mean square error (NMSE) between restoration image C K (i, j) and original imageĈ K (i, j) is given by

Here O (C (i, j)) = (C (i, j)) 1 3 and C K (i, j) andĈ K (i, j) are the restoration and original images respectively and M, N are the frame height and width respectively. The large value of Mean Square Error (MSE) means that image is poor quality. Peak signal to noise ratio (PSNR) [15] between restoration image C K (i, j) and original imageĈ K (i, j) is given by Here M, N are the frame height and width respectively. The small value of Peak Signal to Noise Ratio (PSNR) means that image is poor quality. Image fidelity (IF) between restoration image C K (i, j) and original imageĈ K (i, j) is given by

j=0 ((C (x, y)) − (Ĉ (x, y))) 2

Here M, N are the frame height and width respectively. The value is as high as possible and maximum is one and higher value indicates image is good quality Peak mean square error (PMSE) [15] between restoration image C K (i, j) and original imageĈ K (i, j) is given by

j=0 ((C k (x, y)) − (Ĉ k (x, y))) 2 M −1 i=0 N −1 j=0 (max(C k (x, y))) 2 (19) Here M, N are the frame height and width respectively. The large value gives poor image quality. Structural content of an image (SC) [16] between restoration image C K (i, j) and original imageĈ K (i, j) is given by

Here K is for number of components Ex. RGB and M, N are the frame height and width respectively. The large value of Structural Content (SC) means that image is poor quality.

Percentage of similarity (PS) [17] between restoration image C K (i, j) and original imageĈ K (i, j) is given by

k=0 min (C k (i, j)), (Ĉ k (i, j)) K k=0 (C k (i, j)) + (Ĉ k (i, j))   (21) Here K is for number of components Ex. RGB and M, N are the frame height and width respectively. This value is as high as possible and higher value indicates image is good quality Structural distortion measure (Q) or (SSIM) [18] between restoration image C K (i, j) and original imageĈ K (i, j) is given by

Herex,ȳ are the mean of original and restoration images and σ xy is the co-variance between original and restoration images and σ 2 x , σ 2 y are the variance of original and restoration images. The maximum value is one and less value of index will give poor result.