Effective Image Haze Removal Using Dark Channel

Effective Image Haze Removal Using Dark Channel Prior And Post-processingSoo-Chang Pei1, Tzu-Yen Lee2l Department of Electrical Engineering, National Taiwan University, Taipei, Taiwan R.O.C.2 Graduate Institute of Communication Engineering, National Taiwan University, Taipei, Taiwan R.O.C.E-mail : 1 pei@.tw, 2 r99942049@.twAbstract—Removing haze technique is an important and necessary procedure to avoid ill-condition visibility of human eyes. To improve the dehazing quality, a new haze-removing method is proposed in this paper. Specifically, this method can be properly applied to various dense and distribution haze images without sacrificing color naturalness by using refined Dark Channel Prior(DCP) and adding post processing. Experimental results demonstrate that the new method provides higher dehazing quality and color naturalness than other existing techniques.I.INTRODUCTIONHaze removal is a challenging problem because of the unknown transmission and ill-posed conditions. The key characteristics of light such as intensity and color are changed by its interactions with the atmosphere. These interactions can be classified into three categories: scattering, absorption, and emission. As a result, haze occurs due to atmospheric absorption and scattering, at which the irradiance received by the camera from the scene point is attenuated along the line of sight; the incoming light is blended with the airlight.Dehazing is the process of removing haze effects from capturing images and recovering the original details and colors of natural scenes. A computationally-efficient haze-removal algorithm was proposed in [2] which has the problem of the unknown depth information, and it needs multiple images or extra information to estimate the unknown depth information. In order to overcome the drawbacks, the single image algorithms have been investigated in [1, 3, 4]. Fattal [4] estimated the albedo of scene and inferred the medium transmission. However, this approach cannot well handle heavy haze images and may fail in the cases that the assumption is invalid. Tan [3] maximizes the local contrast of the restored image which assumes that a dehazed image should have high contrast. Unfortunately, Tan’s algorithm often produces halo artifacts and overstretches contrast. The approach of He et al. [1] is the most attracting one based on the dark channel prior. The dark channel prior means that at least one color channel should have almost zero pixel values within a haze-free image. By applying the method, we can obtain impressive results, but the color tone might be modified due to inaccurate estimation of airlight and the patch operation. Unlike the conventional methods mentioned before, the proposed method can avoid the block and contouring effects. Moreover, it can reduce the complexitythat increases from refining the transmission. Thus, the restoration performance of the proposed method yields high quality images, and it is more robust than the approach ofHe et al. [1].This paper is organized as follows: Section Ⅱ briefly reviews the haze image model and dark channel prior. Section III presents the proposed dehazing algorithm applicable to all conditions of haze images. The experiment results are presented in Section Ⅳ. Finally, Section Ⅴconcludes the paper.II.PRELIMANARIESThe proposed method of this paper is based on the haze image model [2, 3, 4, 5] and the haze removing method named dark channel prior [1]. In the following sub-sections, these fundamental concepts of the proposed method will be briefly introduced.A. Haze Image ModelThe haze image model proposed by McCartney [2, 3, 4,5] is widely used in computer vision described as()()(1-())I x t x J t x A=+(1), where I is the observed intensity of the haze image, x is the pixel’s index, J is the scene radiance of the haze-free image which is desired to be obtained by haze-free techniques, t isthe medium transmission expressed part of light that is not scattered and reaches the observer, and A is the global atmospheric light.B.Dark Channel PriorThe dark channel prior is based on the assumption thatin the most of the non-sky patches contain some pixels which have very low intensities in at least one color channel. Therefore, the darkest scene radiance (J) of the x thpixel is expressed as{,,}()()min(min(()))dark cc r g b y xJ x J y∈∈Ω=(2), where c J is a color channel of J and Ω(x) is a local patch centered at x. Thus, the darkJ is so called the dark channelof J, and the statistical observation is named as dark channel prior.The flow chart of the dark channel prior [1] is given inFig.1. In the 1st step, the estimated atmospheric light (A) is acquired by choosing the top 0.1% brightest pixels in the dark channel which are the most haze-opaque pixels. In the 2nd step, the coarse estimated transmission is()()'()1min(min ())cc c y x I y t x w A∈Ω=− (3), where w (01w <≤) is a constant to prevent the full elimination of haze; the full elimination of haze would result in an unnatural image and the loss of depth perception. In the 3rdstep, estimated transmission map is refined by applying the soft matting method [6], and we express the refined '()t x as ()t x . Finally, the estimated scene radiance is0()()max((),)I x A Jx A t x t −=+ (4), where t 0 is the lower bound of the refined medium transmission. By applying the steps of dark channel prior method [1], the haze removal results can be obtained. However, the block effect and soft matting complex computation are disadvantages of this method, then the proposed method is provided to solve these problems. III. PROPOSED METHODThe flow chart of the proposed dehazing technique is shown in Fig. 2 which refines the 1st to 3rd steps in Fig. 1, and the post-processing is added for increasing the robustness of the proposed method. A. Refined Dark Channel Prior MethodFor acquiring higher quality haze-free images, the 1st to 3rd steps of the dark channel prior method [1] are modified in this sub-section.First, we replace the patch operation with pixel-wise operation to overcome the block effect, contouring effect and the inaccurate estimation of the sky-region in (2). Thus, the pixel-wise operating form of (2) is expressed as{,,}()()min (min (()))dark c c r g b y pixel x J x J y ∈∈= (5). After replacing the operating mode of (2) by (5), the 2nd step of the dark channel prior [1] method is modified as ()()''()1min (min ())c c c y pixel x I y t x w A ∈=−(6).Fig. 1. Flow chart of the dark channel prior[1].Fig. 2. Flow chart of the proposed dehazing technique.where c A is the global atmospheric light estimated from the 1st step in Fig. 1. In this paper, all pixels from the darkchannel are sorted initially in ascending order as _dark order J by intensity values. Then, we choose 96% intensity of the pixel values (i.e., 96% value in _dark order J ) as the airlight value to prevent from choosing some wrong high-light objects as white cars and buildings. Note that the modified method can solve problems caused by the patch operation which are mentioned before.For decreasing the complexity of the soft matting in the 3rd step, the guided filter [9] is employed to smooth the transmission map while capturing the sharp edgediscontinuities and outline the profile of objects. Guidedfilter is an efficient linear-time algorithm that can provide exact solutions and the computational complexity depends on its size of kernel. The refined transmission of (6) is represented as:1'()(''())||kk k k x W t x a t x b W ∈=+∑ (7),where |W| is the number of the pixels in W k , W k is a window centered at the pixel k , and (a k ,b k ) are some linear coefficients assumed to be constant in W k . By applying the guided filter, the refined transmission map '()t x can be obtained, then the estimated and improved scene radiance '()J x c an be acquired by (4).0()'()I x AJx A −=+ (8).B. Post-ProcessingFor making the proposed haze-removal method morerobust, the post-processing is added for handling the situation that the haze of the input image is non-uniformdistribution with insufficient brightness. Thus, the multi-scale retinex (MSR) [7] and bilateral filter in local contrastcorrection (BFLCC) [8] are applied to perform the post-processing. The MSR method suits for handling the casethat the sky is covered most of the haze image and theBFLCC method suits for handling the case that the sky is not the main component of the haze image. Note that bothof the post processing approaches can remove the haze clearly. In the sky region, the color of MSR would be more natural than the BFLCC. However, the BFLCC would be better than MSR in non-sky region since it can avoid some artifacts simultaneously. Fig. 3 compares the dehazed results with our post processing approaches: MSR andBFLCC. Choosing appropriate post processing method isfurther investigated under two different cases.Case 1: The MSR method [7] is composed of the different scales of the Gaussian functions, and it adds different weights to the adjacent region with various sizes. For keeping the image edge and the color information simultaneously, the MSR method combines N various surround constants expressed as1(,)(,) {,,}i Ni n n n M x y w R x y i R G B ==∈∑ (9) where M is the output image, x and y are the pixel’s vertical and horizontal indexes, i is the index of the spectral band,w n denotes the weight associated with the n th scale, and i n Rdenotes the single-scale retinex of the i th component of the n th scale represented as(,)log((,))log((,)(,))''i n i i n R x y x y x y F x y J J =−∗ (10), where F n (x,y) denotes the Gaussian surround function givenas{}222(,)exp ()/n n F x y K x y c =−+ (11), where c is the Gaussian surround space constant, and K is applied to normalize the surround function (11) . Case 2: The BFLCC method[8] is improved from the gamma correction which replaces an inverted Gaussian low-pass filtered by an inverted low-pass, filtered with a bilateral filter which is expressed as(,)(128)128(,)(,)255[]255'BFmask x y x y O x y J α−= (13). where x and y are the pixel’s vertical and horizontal indexes, O is the output image, α is a parameter depending on the image properties and BF mask . Note that BF mask is defined over a window of size (21)(21)K K +×+ which is given by the bifiltered image of the inverted version of the input'(,)255(,)'inv x y x y J J=− (14). IV. EXPERIMENTAL RESULTSFor comparing the experimental results, the source ofthe haze images are taken from [1] and the ROHM Semi-conductor. Furthermore, we also present dehazing results for real hazed pictures of Huang Shan by applying different post-processing method in Fig. 3. Besides, the experiment is separated into 2 cases: Case 1: the input image has sufficient brightness and uniformly distributed haze. Case 2: the input image has insufficient brightness and non-uniformly distributed haze.In Case 1, the input image is dehazed by applying only the refined dark channel prior method shown in Fig. 2without post processing. The haze removal images are shown in Fig. 4 and Fig. 5. In Fig. 4, the proposed and He et al. [1] approaches are compared, and it shows that the result of the He et al. is more fuzzy and darker (Ex: brick house part), but the proposed method would not have such appearance. In Fig. 5, the proposed approach is compared with the ROHM Company’s haze removal image, and it can be observed that the ROHM Company’s result wouldlead the over-exposure happened in the back of the cable car. Besides, the color of the mountains is unnatural. Theproposed method recovers the problems. In Case 2, the input image is dehazed by applying the refined dark channel prior method with post processingshown in Fig. 2 for ensuring robustness. The experimental results are compared with He et al. [1]. As shown in Fig. 6,the results of He et al. are more fuzzy and darker. Some of the regions could not be recovered clearly (Ex: far region), that is, the quality of the output haze removal images would be lower. In contrast, the proposed method can acquire clearer and brighter dehazed images. The recovered degreeof the far region is better than He et al. results[1] (Fig.6(b)).In Fig. 6, we also compare the results of applying the post-processing (BFLCC) without dehazing (Fig.6(c)) and with dehazing procedure. (Fig.6(d))V. CONCLUSIONSAn attractive haze-removing method with high dehazing quality is proposed in this paper. Specifically, the proposed method is robust and provides a reliable hazeremoval quality for various cases of haze images. Compared with other existing dehazing methods, theproposed method is superior in dehazing quality and colornaturalness. In the future, some algorithmic refinements could be applied and then integrated into the haze-removing method to decrease the complexity and improve the dehazing image quality.(a) (b) (c)Fig. 3. Insufficient brightness and non-uniform distributed haze: comparison with MSR and BFLCC post-processing. (a) input haze image of Huang Shan picture. (b) final haze-free image after applying MSR post-processing. (c) final haze-free image after applying BFLCC post-processing.(a) (b) (c)Fig. 4 : Sufficient brightness and uniform distributed haze: comparison with He’s work[1]. (a) input haze image. (b) He et al.’s result. (c) proposed methodresult.(a) (b) (c)Fig. 5: Sufficient brightness and uniform distributed haze: comparison with ROHM’s IC-BU6521KV. (a) haze image. (b) ROHM Company’s result. (c)proposed method result.(a)(b)(c) (d) Fig. 6: Insufficient brightness and non-uniform distributed haze: comparison with He’s work [1]. (a) input haze images. (b) He et al.’s results. (c) the results ofapplying the BFLCC without dehazing procedure. (d) the results of applying the proposed method.R EFERENCES[1]Kaiming.He, J.Sun, and X.Tang, “Single Image Haze RemovalUsing Dark Channel Prior,” IEEE Conference on Computer Visionand Pattern Recognition, pp. 1956-1963, 2009.[2]S. G. Narasimhan and S. K. Nayar, “Vision and the atmosphere,”International Journal of Computer Vision, vol. 48, no. 3, pp. 233-254, 2002.8.[3]R. Tan., “Visibility in bad weather from a single image,” IEEEConference on Computer Vision and Pattern Recognition, pp. 1–8,2008[4]R. Fattal, “Single image dehazing,” ACM SIGGRAPH, pp. 1–9,2008.[5]S. G. Narasimhan and S. K. Nayar, “Chromatic framework forvision in bad weather,” IEEE Conference on Computer Vision andPattern Recognition, pp. 598–605, 2000.[6] A. Levin, D. Lischinski, and Y. Weiss, “A closed form solution tonatural image matting,” IEEE Conference on Computer Vision andPattern Recognition, pp. 61–68, 2006.[7] D. Jobson, Z. Rahman, and G. Woodell, “A multiscale retinex forbridging the gap between color images and the human observationof scenes,” IEEE Transaction on Image Processing, pp. 965–976,1997.[8]Raimondo Schettini, Francesca Gasparini, Silvia Corchs andFabrizio Marini, “Contrast image correction method,” Journal ofElectronic Imaging 19(2), 023005, 2010.[9]Kaiming He, Jian Sun and Xiaoou Tang, “Guided Image Filtering,”Lecture notes in Computer Science, Vol 6311, 2010.。