Weighted median guided filtering method for single image rain removal

Weighted median guided filtering method for single image rain removal Because there is no temporal information available, rain removal with a single image is more challenging than that with a video. In this paper, we present a weighted median guided filtering method for rain removal with a single image. It consists of two filtering operations. Firstly, a weighted median filter is convoluted with an input rainy image to obtain a coarse rain-free image; then, guided filter is employed to obtain a refined rain-free image, where the coarse rain-free image is used as a guided image and convoluted with the input rainy image via guided filter. Experimental results show that the proposed method generated comparable results to the state-of-the-art algorithms with low computation cost. Keywords: Rain removal, Weighted median filter, Guided filter 1 Introduction layer by sparse coding, with a learned dictionary from In rainy days, the performance of outdoor vision systems the histogram of oriented gradients (HOG) features. will significantly degrade due to visibility obstruction, However, the aforementioned dictionary partition-based deformation, and blurring caused by raindrops. There- rain removal methods inevitably result in reconstructed fore, it is highly desirable to remove raindrops from rainy images with either over smooth or incomplete rain images to ensure the reliability of outdoor vision systems removal. This is caused by the inaccurate decomposi- tion of the high-frequency portion into rain components [1]. For this purpose, numerous efforts have been made in past years. One common strategy is using video sequences and non-rain components, which failed to recover the [1, 2] for rain removal. The main idea of this kind of non-rain components and faulty incorporation of the methods is to explore the redundant temporal informa- rain components into the low-frequency partition. Similar tion from multiple images. Though such kind of method methodshavealsobeenproposedin[4, 5], Kang et al. [4] works well, it heavily depends on the temporal contents proposed a method that employing bilateral filter to divide in videos and cannot be applied for the case where only a the image with rain into low-frequency portions and high- single image is available. Nevertheless, in this age of ubiq- frequency portions firstly. The rain component is then uitous smart phone usage, there is an increasing need for extracted from the high-frequency portion by using a techniques where only a single image available. Motivated sparse representation-based dictionary partition in which by this need, in this paper, we instead focus on removing the dictionary is classified using HOG in each atom where rain from a single image. the bilateral filter is used to separate the low-frequency Compared with video-based rain removal, due to the part from its high-frequency part of an input image. lack of temporal information, rain removal with a sin- Though the decomposition idea is elegant, the selection gle image is more challenging. Some single-image-based of dictionaries and parameters are heavily empirical, and rain removal methods regard the problem as a layer sep- the results are sensitive to the choice of dictionaries. aration problem. For example, Huang et al. [3]attempted Moreover, all the three dictionary learning-based frame- to separate the rain streaks from the high-frequency works [3, 5] suffer from heavy computation cost. In [6], Manu uses the L0 gradient minimization approach for *Correspondence: zhenghaoshimtap@163.com rain removal. The minimization technique can globally School of Computer Science and Engineering, Xi’an University of Technology, control how many non-zero gradients are resulted in the No. 5 Jinhua South Road, 710048 Xi’an, China Full list of author information is available at the end of the article image. In [7], Kim et al. proposed a two-stage method for © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 2 of 8 rain removal. In the first stage, rain streaks are detected state-of-the-art methods in rain removal with lesser by using a kernel regression method under an assumption computation cost. of the elliptical shape and the vertical orientation of rain Our main contributions in this paper are as follows: steaks. Then, rain steaks are removed by using non-local (1) To our best knowledge, our work is the first one to mean filtering in the second stage. Though the method is apply the weighted median filter for rain removal effective for images with a simple structure, some desir- with a single image. Without any image priors (e.g., able details in images with complex structures are usually the relationship between the input and desirable removed. In [8], Zheng et al. separated the low-frequency output images), our method just takes rain steaks as part of input image using guided filter, and experiments image noise, which makes single-image-based show that the results are better than those of using bilat- applications applicable in real-world scenarios. eral filter. Apart from the abovementioned filters, the (2) The novelty of our method attributes to the use of weighted median filter [9] is also a better alternation for weighted median filtering images to preserve the median filter to effectively filter images while not geometrical details in rain-removed image via guided strongly blurring edges. A recent work of [10] exploits filter. the Gaussian mixture models to separate the rain streaks, achieving the state-of-the-art performance, however, still The remainder of this paper is organized as follows. with slightly smooth background. Section 2 describes the proposed method in detail. Basedonwhatwasmentionedabove,inanattempt to Section 3 provides experimental results on both synthetic preserve more complex structures in the rain-removed rain images and real rain images. Section 4 discusses some images, in this paper, we present a weighted median issues about the proposed method. Finally, the paper is guided filtering method for rain removal with a sin- concluded in Section 5. gle image. It consists of two main operations. Firstly, a weighted median filter is convoluted with an input 2 Proposed method rainy image to obtain a coarse rain-free image. Then, 2.1 Overview of the proposed method the coarse rain-free image is used as a guided image Figure 1 shows the framework of the proposed method. It and convoluted with the input rainy image via guided consists of two main steps: firstly, the input rainy image filter to obtain the final rain-free image. Unlike the is filtered using the weighted median filter [9], where the aforementioned methods, the proposed method does rain steaks will be excluded and the most basic informa- not rely on other image processing modules for pre- tion will be retained; then, the weighted median filtered or post-processing, which avoids the possible vulnera- image is used as a guide image and convoluted with bility of these techniques when processing images of the input rainy image to obtain a texture/edge preserved complex structures. Experimental results show that the rain-free image via guided filter. Details of each step are proposed method generates comparable outputs to the elaborated below. Fig. 1 Framework of the proposed method. It consists of two computation stages: firstly, to filter the input rainy image using the weighted median filter; then, the weighted median filtered image is used as a guide image and convoluted with the input rainy image to get the final rain-free image Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 3 of 8 2.2 Remove rain steaks via weighted median filtering image. Formally, given f (x, y) as the value of the centered Median filter has been widely used in image denoising pixel (x, y) of a local window R(x, y), then all pixels in the due to its well smoothing effect on noise with long tail window can be expressed as probability distribution as well as better preserving func- R(x, y) =|f (x + k, y + r)|k, r =−1, 0, 1| (1) tion on image details. However, the filtering window size To compute the average value of all pixels in the window has an important effect on the denoising performance of the traditional median filter. A small size of the filtering R(x, y) window, a better detail preserved with a lower denosing 1 1 performance; while a large size of the filtering window, a R(x, y) = f (x + k, y + r) (2) averaged high denosing performance with a poor detail preserved. k=−1 r=−1 To address this problem, the weighted median filter was Let Z , Z is the max and min pixel value in the max min proposed [9]. The main idea of the weighted median local window R(x, y), then for the pixel value f (x, y) of the filter is to replace the current pixel with the weighted centered pixel (x, y), if f (x, y) = Z ,or f (x, y) = Z ,or max min median of neighboring pixels within a local window, as |f (x, y) − R(x, y) | > d , then the pixel (x,y) will be averaged x,y showninFig. 1, where the current pixel I is replaced takenasanoisepixel.Here, d x, y) is a threshold which is with the weighted median of its neighboring pixels within determined by a local window I . This filter has the following special characteristics: 1 1 d x, y) = [ f (x + k, y + r) − R(x, y) ] ( averaged (1) The filtering kernel is not separable. r=−1 k=−1 (2) It cannot be approximated by interpolation or (3) down-sampling. (2) Determining the filtering window size. In order to (3) There is no iterative solution. combine the advantage of both window size of a filter, For the reasons mentioned above, the weighted median the size of the filtering window is determined accord- filter can effectively remove noises from an noised images ing to the number of noise pixels in a local window while not strongly blurring the edges of image structure. R(x, y). Given the number of noise pixels in a local window This is why we employ the weighted median filter for rain R(x, y) as Num(R), then the size of the filtering window is streaks removal in this work. The filter used here con- determined as sists of the following three operation steps: (1) rain streaks 3 × 3, Num(R) ∈{1, 2, 3} detection, (2) determining the filtering window size, and Size(R) = 5 × 5, Num(R) ∈{4, 5, 6} (4) (3) noise filtering. Details of each operation are described 7 × 7, Num(R) ∈{7, 8, 9} as followings: (1) Rain streaks detection. This operation will provided (3) Noise filtering. Formally, given a pixel p in an image basis for rainy image pixel classification. To determine the I, and a local window R(p) of radius r centered at p,for noised pixels, a 3 × 3 window is used to slide over the each pixel q ∈ R(p), we define weighted median filter as Fig. 2 Results with different filtering method. a Input rainy images. b Filtering results using the weighted median filter. c Filtering results using guided filter Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 4 of 8 all pixels in R(p) as{(I(q), w )}. Then, by sorting all pix- pq els values in an ascending order, we can get a weighted median filtered image I(p ),where k n p = minks.t. w ≥ w (6) pq pq q=1 q=1 In order to accelerate weighted median filtering pro- cessing, ideas proposed in [9], including the joint- histogram and median tracking strategy using balance counting box for dynamically finding median, are used in this work. Figure 2 shows part results using weighted median filtering for rain removal. It can be seen that all rain steaks in the input rainy image are well removed after filtered with the weighted median filter, as shown in Fig. 2b. 2.3 Recover texture details from weighted median filtered image using guided filter As can be seen from Fig. 2b, though rain steaks in input rainy images are well removed, part regions in the input image are also over smoothed, where some edge and tex- ture details are removed, for example, the wave lines and textures in the left image of the first row and edges and textures of grass in the left image of the second row are all removed, so those regions look very unnaturally. Aiming at this problem, guided filter [11]isemployedin this section, where we used the weighted median filter- ing image as the guide image, and used input image as the input of guided filter. The main idea of guided filter is to filter input images by considering the content of the guidance image. Formally, given a guidance image I, a guided filtering output image Fig. 3 Comparison of different rain removal methods on synthetic guided I of input image I can be defined as in in rainy images. The top row is the input rainy images, follows in sequence are the ground true, results with Li’s method [10], Manu’s guided I (x) = a I (x) + b , ∀i ∈ ω (7) method [6],Luo’smethod[5], Huang’s method [3], and our method k k k in where ω is a window centered at the pixel x, a and b are k k k assumed to be constant in ω , and determined as w = G(f (p), f (q)) (5) pq i∈ω I p −μ p n i ω k k k where w is a weight based on the affinity of pixel p and q in pq a = (8) σ + ε in corresponding feature map f. f (p) and f (q) are features at pixels p and q in f,respectively. g is a typical influ- ence function between neighboring pixels, which can be b = p − a μ (9) k k k in Gaussian exp−||f (p) − f (q)|| or other forms. Let I(q) denote the value of a pixel at the location q in where μ and σ are the average value and the vari- image I, n = (2r + 1) denote the number of pixels in ance of the input image I in window ω ,respectively. in R(p), we can describe the values and weight elements for n is the number of pixels in window ω . p is the k k Table 1 Quantitative comparison of different methods on synthetic rainy images Li’s method [10] Manu’smethod[6] Luo’s method [5] Huang’s method [3] Our method PSNR 31.93 ± 0.01 24.05 ± 0.01 33.34 ± 0.01 33.98 ± 0.01 35.46 ± 0.01 SSIM 0.81 ± 0.01 0.72 ± 0.01 0.80 ± 0.01 0.83 ± 0.01 0.90 ± 0.01 Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 5 of 8 is why we employed guided filter as a tool for image texture recover. Figure 2c shows the recovered images with guided fil- ter. As can be seen that the removed image texture and edge details by the weighted median filter are well recov- ered, for example, as shown in Fig. 2c, the wave lines and textures in the left image of the first row and edges and textures of the grass in the left image of the second row are very clear to observe, and those regions look very naturally compared with that in Fig. 2b. 3 Experimental results 3.1 Experiment setting To evaluate the proposed method, extensive tests have been done using MATLAB R2015 on a PC with a 2.60 GHz Intel Dual Core Processor and 4G RAM. To the best of our knowledge, there is no standard rainy image data set available for bench marking currently. Hence, we collected totally 100 natural/synthetic rainy images from the Inter- net and also from the test image data set released from [10]. We compare the proposed method with several state- of-the-art methods, including Li’s method [10], Manu’s method [6], Luo’s method [5], and Huang’s method [3], the results of these methods are generated in MATLAB with suggested parameter setting by the authors. For quantitative evaluation of different methods on syn- thetic rainy images, because the ground-truth images are available, we employ the indexes of peak signal-to-noise ratio (PSNR) [12] and Structure Similarity Index (SSIM) Fig. 4 Comparison of different rain removal methods on real rainy images. The top row is the input rainy images, follows in sequence [13] on the luminance channel as evaluation measures. are the results with Li’s method [10], Manu’s method [6], Luo’s For quantitative evaluation of different methods on nat- method [5], Huang’s method [3], and our method ural rain images, because the ground-truth images are usually unavailable, PSNR and SSIM are unable to be used as evaluation measures; we employ an overall qual- ity index proposed in [14] as evaluation measure. This mean of the guided image in window ω . ε is the regu- indicator estimates the average visibility enhancement larization parameter which is used to control the struc- obtained by the restoration method. The higher the value tural similarity. The larger ε is, the smoother the output of the overall quality index is, the better the enhanced will be. visibility is. Edges in an image after using guided filter will change differently. For step edges, it is still step edges after using 3.2 Evaluation on synthetic rainy images guided filter, but their ranges become smaller, which Figure 3 shows part experiment results on synthetic rainy means that the step edges become smoother after guided images [10]. It can be seen that though Li’s method [6] filter. For ridge edges, if the ridge edges with small size are works well on rain removal, the contrast of the output unaffected by the other edges, their variances are close to images is obviously decreased, and details in the dark part 0, then the ridge edges will disappear and tend to the back- of the output images are difficult to observe, as shown in ground; for valley edges will become larger than the input. the third row in Fig. 3, for example, the feet of the old man From what was mentioned above, we can see that the in the third row in Fig. 3b are not clear to be seen. For guided filter has well preserving ability on image edges. Manu’s method [10], as shown in the fourth row in Fig. 3, Therefore, it can be used for image texture recovery. This Table 2 Quantitative comparison of different methods on real rainy images Li’s method [10] Manu’s method [6] Luo’s method [5] Huang’s method [3] Our method 0.2854 ± 0.01 0.1665 ± 0.01 0.3287 ± 0.01 0.1823 ± 0.01 0.3925 ± 0.01 Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 6 of 8 it works well on rain removal. However, the whole image 4 Discussion is over smoothed, the output image looks unnatural, and In this paper, we present a weighted median guided fil- part texture information is lost. For example, because of tering method for rain removal with a single image. To the lost of the textures information in the output images, several mountains become one in the first image in the fourth row, and clouds in the third image in the fourth row are also disappeared. For Luo’s method [5]andHuang’s method [3], it can be seen that there are many rain streaks remained in the output images, as shown in the fifth row and sixth row in Fig. 3. Whereas, with respect to the afore- mentioned methods, the proposed method performs well on these synthetic images, as shown in the seventh row in Fig. 3, and is able to successfully remove majority of the rain streaks while maintaining most details of the original images and achieved a good visual effect, which are very close to the ground true (the second row in Fig. 3). Table 1 shows quantitative evaluation results of differ- ent methods on synthetic rainy images. It can be clearly observed that theproposedmethodisabletoachieve bet- ter quantitative performance compared to all of other four mentioned methods in terms of both of PSNR and SSIM, and this is consistent with what we observed in Fig. 3. 3.3 Evaluation on real rainy images Figure 4 shows part results of this experiment. It is obvi- ously the result with the proposed method which out- performs most of the other state-of-the-art methods in terms of both the effectiveness of removing rain streaks and the visual quality of recovered images. For example, Manu’s method [6] removes many image details and tex- tures from the input image and that makes the image unnatural, as shown in the third row in Fig. 4.Incon- trast, Li’s method [10] (the second row in Fig. 4.) and our method (the last row in Fig. 4.) are able to remove most of the rain streaks and meanwhile produce lesser artifacts on the recovered images. But the results with Li’s method [10] look a little dark and some details are not clear to see, asshowninthe secondrowinFig. 4, whereas that with our method has the sharpest appearance of the trees and leaves, as well as clear people faces, and there no afore- mentioned issues appear, as shown in the last row in Fig. 4. For Luo’s method [5]and Huang’smethod[3], as shown in the fourth row and fifth row in Fig. 4, there are not only many rain streaks that remained in the output images, but also the background is blur which lead difficult to observe image details. Table 2 shows the quantitative evaluation results of dif- ferent methods on real rainy images in terms of the overall quality index. It can be seen that the average overall qual- ity index value of the proposed method are higher than that of other methods, which indicates that the proposed Fig. 5 Comparison of different noising filters for removing raining. The top row is the input rainy images, follows in sequence are the method has better ability in recovering the visibility of results with Gaussian filter, median filter, bilateral filter, guided filter, rainy images compared to other methods. This confirms the weighted median filter, and the ground true our observations on Fig. 4. Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 7 of 8 Table 3 Running time (seconds) for different methods on a 321× 481 image Li’s method [10] Manu’s method [6] Luo’s method [5] Huang’s method [3] Our method 12.74 ± 0.01 31.40 ± 0.01 61.35 ± 0.01 26.86 ± 0.01 1.22 ± 0.01 our best knowledge, our work is the first one to apply 5Conclusions the weighted median filter for rain removal with a single In this paper, we present a simple and effective weighted image. Compared to most of the existing methods, our median guided filter for rain removal with a single image. method does not rely on other image processing modules Different from most existing methods, the proposed for pre- or post-processing, just takes rain steaks as image method does not rely on other image processing mod- noise, which avoids the possible vulnerability of these ules for pre- or post-processing, which avoids the likely techniques when processing images with complex struc- vulnerability of these techniques when processing images tures and makes single-image-based applications applica- with complex structures. Experimental results show that ble in real world. the proposed method produces comparable outputs to the To verify whether the weighted median filter employed state-of-the-art algorithms with low computation cost. in this method can be replaced by other de-noising filters, Abbreviations such as Gaussian filter, median filter, bilateral filter, and HOG: Histogram of oriented gradients; PSNR: Peak signal-to-noise ratio; SSIM: guided filter, we compare their abilities in rain removal Structure similarity index with that of the weighted median filter in this section. Acknowledgements Part experimental results are shown in Fig. 5.FromFig. 5, The authors thank the editor and anonymous reviewers for their helpful it can be seen although all de-noising filters can remove comments and valuable suggestions. most rain streaks, Gaussian filter (shown in the second Funding row in Fig. 5) also smoothed some details and that make This work was supported in part by a grant from the National Natural Science the image looks a little blur, the results both with median Foundation of China (nos. 61202198, 61401355, and 41601353), a grant from the China Scholarship Council (no. 201608610048), the Key Laboratory filter (shown in the third row in Fig. 5) remained too Foundation of Shaanxi Education Department (no. 14JS072), and the Nature many rain steaks. Though the result with bilateral filter Science Foundation of Science Department of PeiLin count at Xi’an (no. and that with the guide filter look better than that with GX1619). Gaussian filter and that with median filter, the results look Availability of data and materials a little blur compared to that with the weighted median We can provide the data. filter. It is obvious that the weighted median filter suc- Authors’ contributions cessfully removes most rain streaks while preserving most All authors take part in the discussion of the work described in this paper. The non-rain image details in these test cases. The main rea- author ZS wrote the first version of the paper. The author LY did part son is that these de-noising methods are mainly designed experiments of the paper. ZM, JB, FY, and CZ revised the paper in different versions, respectively. The contributions of the proposed work are mainly in for removing Gaussian noise with known standard devi- two aspects: (1) To our best knowledge, our work is the first one to apply the ation. However, it is not easy to model the rain streaks weighted median filter for rain removal with a single image. Without any to be removed as Gaussian noise because of the very image priors (e.g., the relationship between the input and desirable output images), our method just takes rain steaks as image noise, which makes different characteristics between them. The main reason single-image based applications applicable in real-world scenarios. (2) The lead to this might be that the most de-noising filters are novelty of our method attributes to the use of weighted median filtering mainly designed for removing Gaussian noise with known images to preserve geometrical details in rain-removed image via guided filter. All authors read and approved the final manuscript. standard deviation, and it is not easy to model the rain streaks to be removed as Gaussian noise because of the Authors’ information very different characteristics between them. For what was Zhenghao Shi received the BS degree in Material Science and Engineering from Dalian Jiaotong University, Dalian, China, in 1995, the MS degree in mentioned above, we select the weighted median filter as computer application technology from Xi’an University of Technology, Xi’an, our choice in this work for rain removing. China, in 2000, and the Ph.D degree in computer architecture from Xi’an We also compare the running time of the proposed Institute of Microelectronics, Xi’an, China, in 2005. In 2000, he joined Xi’an University of Technology, Xi’an, China. From 2000 to 2005, he was an assistant method with that of Manu’s method [6], Li’s method [10], professor in the Department of Computer Science and Engineering at Xi’an Luo’s method [5], and Huang’s methhod [3]onimages University of Technology. From 2006 up to now, he is an associate professor of with a size of 321× 481, as shown in Table 3.Itcan be the Department of Computer Science and Engineering at the same University. Since December 2016, he joined IDBE laboratory, University of North Carolina seen that it only take 1.22 s to obtain the final rain removal at Chapel Hill as a visiting associate professor. During the period of 2006 to image, whereas that with method [10] is 12.74 s, with 2007, of 2008 to 2009, he was on leave with the Department of Computer method [6] is 31.40 s, with Luo’s method [5] is 61.35 s, Science and Engineering at Nagoya Institute of Technology, Nagoya, Japan, for image research as a postdoctral researcher, respectively. From 2007 to and that with Huang’s method [3] is 26.86 s. It is obvious 2008, he was a research associate in the Kurt Rossmann Laboratories for that our method outperforms all of other state-of-the-art Radiologic Image Research, the Department of Radiology, the Division of methods in terms of running time. Biological Sciences, the University of Chicago. His research interests include Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 8 of 8 neural networks for image processing and pattern recognition, 7. JH Kim, C Lee, JY Sim, CS Kim, in 20th International Conference on Image computer-aided diagnosis, and image processing suggested by the human Processing: Sep 15, 2013 - Sep 18, 2013; Melbourne, Australia. Single-image visual systems. He is a member of the IEEE, also a member of ACM. Yaowei Li is deraining using an adaptive nonlocal means filter (IEEE, Australia, 2013), currently study for his master degree in computer science and technology in pp. 914–917 Xi’an University of Technology, Xi’an, China. Her research interests focus on 8. X Zheng, Y Liao, W Guo, X Fu, X Ding, in Neural Information Processing: image processing. Changqing Zhang received his B. S. and M. S. degrees from 20th International Conference, ICONIP 2013: November 3-7, 2013;Daegu, the College of Computer Science, Sichuan University, in 2005 and 2008, Korea, ed. by M Lee, et al. Single-image-based rain and snow removal respectively, and the Ph.D. degree in Computer Science from Tianjin University using multi-guided filter (IEEE, Korea, 2013), pp. 258–265 in 2016. He is an Assistant Professor at the School of Computer Science and 9. Q Zhang, L Xu, J Jia, in Proceedings of the IEEE Conference on Computer Technology, Tianjin University. His current research interests include machine Vision and Pattern Recognition: June 24-27, 2014; Ohio, USA. 100+ times learning, data mining, and computer vision. Minghua Zhao received her Ph.D faster weighted median filter (WMF) (IEEE, USA, 2014), pp. 2830–2837 degree in computer science from Sichuan University, Chengdu, China, in 2006. 10. Y Li, X Guo, J Lu, RT Tan, MS Brown, in IEEE Conference on Computer Vision After that, she joined Xi’an University of Technology, Xi’an, China. Currently, and Pattern Recognition: June 26th - July 1st, 2016; LAS VEGAS, USA.Rain she is an associate professor of the Department of Computer Science and streak removal using layer priors (IEEE, USA, 2016), pp. 2736–2744 Engineering at the same University. Her research interests include image 11. Kea He, Guided image filtering. IEEE Trans. Pattern Anal. Mach. Intell. 35, processing and pattern recognition. Yaning Feng received her B. S. degree 1397–1409 (2013) from Shannxi University of Technology in 1995, received her M. S. degree from 12. Q Huynh-Thu, M Ghanbari, Scope of validity of PSNR in image/video Xian University of Technology in 2004, and received her Ph.D. degree from quality assessment. Electron. Lett. 44, 800–801 (2008) Nogoya institute of Technology in 2008. She is an Assistant Professor of Xian 13. Z Wang, HR Sheikh, AC Bovik, EP Simoncelli, Image quality assessment: University of Technology. Her current research interests focus on computer from error visibility to structural similarity. IEEE Trans. Image Process. 13, vision. Bo Jiang received his Doctoral degree in electronic circuit and system 600–612 (2004) from Chinese Academy of Sciences, Shanghai, China. He is currently the 14. Z Wang, AC Bovik, Universal image quality index. IEEE Signal Process. Lett. associate professor at Northwest University, Shaanxi, China, in 2014. His 9, 81–84 (2001) research interests include aviation remote sensing and image processing. Ethics approval and consent to participate Not Applicable. Consent for publication Not Applicable. Competing interests The authors declare that they have no competing interests. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Author details School of Computer Science and Engineering, Xi’an University of Technology, No. 5 Jinhua South Road, 710048 Xi’an, China. School of Information Engineering, Northwestern University, No. 1, Xuefu Avenue, Guodu Education and Technology Industrial District, Changan district, 710127 Xi’an, China. School of Computer Science and Technology, Tianjing University, No. 135 Yaguan Road, 300350 Tianjin, China. Received: 25 February 2018 Accepted: 2 May 2018 References 1. AK Tripathi, S Mukhopadhyay, Removal of rain from videos: a review. SIViP. 8, 1421–1430 (2014) 2. Q Zhu, J Yuan, L Shao, in Proceedings of the 2015 IEEE International Conference on Robotics and Biomimetics: December 6-9, 2015; Zhuhai, China, ed. by Ning Xi, YWST Shigeki Sugano, and Q Huang. The current challenges and prospects of rain detection and removal from videos (IEEE Robotics and Automation Society, China, 2015), pp. 843–846 3. DA Huang, YC Frank Wang, LW Kang, CW Lin, Self-learning based image decomposition with applications to single image denoising. IEEE Trans. Multimed. 16, 83–93 (2014) 4. LW Kang, CW Lin, YH Fu, Automatic single-image-based rain streaks removal via image decomposition. IEEE Trans. Image Process. 21, 1742–1755 (2012) 5. Y Luo, Y Xu, H Ji, in Proceedings of the IEEE International Conference on Computer Vision: December 13-16, 2015; Santiago, Chile. Removing rain from a single image via discriminative sparse coding (IEEE, Chile, 2015), pp. 3397–3405 6. BN Manu, in Proceeding of 7th International Conference on Information Technology and Electrical Engineering: October 29-30, 2015; Chiang Mai, Thailand. Rain removal from still images using l0 gradient minimization technique (IEEE, Thailand, 2015), pp. 263–268 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png EURASIP Journal on Image and Video Processing Springer Journals

Weighted median guided filtering method for single image rain removal

Free
8 pages
Loading next page...
 
/lp/springer_journal/weighted-median-guided-filtering-method-for-single-image-rain-removal-jn2h91g4d0
Publisher
Springer International Publishing
Copyright
Copyright © 2018 by The Author(s)
Subject
Engineering; Signal,Image and Speech Processing; Image Processing and Computer Vision; Biometrics; Pattern Recognition
eISSN
1687-5281
D.O.I.
10.1186/s13640-018-0275-9
Publisher site
See Article on Publisher Site

Abstract

Because there is no temporal information available, rain removal with a single image is more challenging than that with a video. In this paper, we present a weighted median guided filtering method for rain removal with a single image. It consists of two filtering operations. Firstly, a weighted median filter is convoluted with an input rainy image to obtain a coarse rain-free image; then, guided filter is employed to obtain a refined rain-free image, where the coarse rain-free image is used as a guided image and convoluted with the input rainy image via guided filter. Experimental results show that the proposed method generated comparable results to the state-of-the-art algorithms with low computation cost. Keywords: Rain removal, Weighted median filter, Guided filter 1 Introduction layer by sparse coding, with a learned dictionary from In rainy days, the performance of outdoor vision systems the histogram of oriented gradients (HOG) features. will significantly degrade due to visibility obstruction, However, the aforementioned dictionary partition-based deformation, and blurring caused by raindrops. There- rain removal methods inevitably result in reconstructed fore, it is highly desirable to remove raindrops from rainy images with either over smooth or incomplete rain images to ensure the reliability of outdoor vision systems removal. This is caused by the inaccurate decomposi- tion of the high-frequency portion into rain components [1]. For this purpose, numerous efforts have been made in past years. One common strategy is using video sequences and non-rain components, which failed to recover the [1, 2] for rain removal. The main idea of this kind of non-rain components and faulty incorporation of the methods is to explore the redundant temporal informa- rain components into the low-frequency partition. Similar tion from multiple images. Though such kind of method methodshavealsobeenproposedin[4, 5], Kang et al. [4] works well, it heavily depends on the temporal contents proposed a method that employing bilateral filter to divide in videos and cannot be applied for the case where only a the image with rain into low-frequency portions and high- single image is available. Nevertheless, in this age of ubiq- frequency portions firstly. The rain component is then uitous smart phone usage, there is an increasing need for extracted from the high-frequency portion by using a techniques where only a single image available. Motivated sparse representation-based dictionary partition in which by this need, in this paper, we instead focus on removing the dictionary is classified using HOG in each atom where rain from a single image. the bilateral filter is used to separate the low-frequency Compared with video-based rain removal, due to the part from its high-frequency part of an input image. lack of temporal information, rain removal with a sin- Though the decomposition idea is elegant, the selection gle image is more challenging. Some single-image-based of dictionaries and parameters are heavily empirical, and rain removal methods regard the problem as a layer sep- the results are sensitive to the choice of dictionaries. aration problem. For example, Huang et al. [3]attempted Moreover, all the three dictionary learning-based frame- to separate the rain streaks from the high-frequency works [3, 5] suffer from heavy computation cost. In [6], Manu uses the L0 gradient minimization approach for *Correspondence: zhenghaoshimtap@163.com rain removal. The minimization technique can globally School of Computer Science and Engineering, Xi’an University of Technology, control how many non-zero gradients are resulted in the No. 5 Jinhua South Road, 710048 Xi’an, China Full list of author information is available at the end of the article image. In [7], Kim et al. proposed a two-stage method for © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 2 of 8 rain removal. In the first stage, rain streaks are detected state-of-the-art methods in rain removal with lesser by using a kernel regression method under an assumption computation cost. of the elliptical shape and the vertical orientation of rain Our main contributions in this paper are as follows: steaks. Then, rain steaks are removed by using non-local (1) To our best knowledge, our work is the first one to mean filtering in the second stage. Though the method is apply the weighted median filter for rain removal effective for images with a simple structure, some desir- with a single image. Without any image priors (e.g., able details in images with complex structures are usually the relationship between the input and desirable removed. In [8], Zheng et al. separated the low-frequency output images), our method just takes rain steaks as part of input image using guided filter, and experiments image noise, which makes single-image-based show that the results are better than those of using bilat- applications applicable in real-world scenarios. eral filter. Apart from the abovementioned filters, the (2) The novelty of our method attributes to the use of weighted median filter [9] is also a better alternation for weighted median filtering images to preserve the median filter to effectively filter images while not geometrical details in rain-removed image via guided strongly blurring edges. A recent work of [10] exploits filter. the Gaussian mixture models to separate the rain streaks, achieving the state-of-the-art performance, however, still The remainder of this paper is organized as follows. with slightly smooth background. Section 2 describes the proposed method in detail. Basedonwhatwasmentionedabove,inanattempt to Section 3 provides experimental results on both synthetic preserve more complex structures in the rain-removed rain images and real rain images. Section 4 discusses some images, in this paper, we present a weighted median issues about the proposed method. Finally, the paper is guided filtering method for rain removal with a sin- concluded in Section 5. gle image. It consists of two main operations. Firstly, a weighted median filter is convoluted with an input 2 Proposed method rainy image to obtain a coarse rain-free image. Then, 2.1 Overview of the proposed method the coarse rain-free image is used as a guided image Figure 1 shows the framework of the proposed method. It and convoluted with the input rainy image via guided consists of two main steps: firstly, the input rainy image filter to obtain the final rain-free image. Unlike the is filtered using the weighted median filter [9], where the aforementioned methods, the proposed method does rain steaks will be excluded and the most basic informa- not rely on other image processing modules for pre- tion will be retained; then, the weighted median filtered or post-processing, which avoids the possible vulnera- image is used as a guide image and convoluted with bility of these techniques when processing images of the input rainy image to obtain a texture/edge preserved complex structures. Experimental results show that the rain-free image via guided filter. Details of each step are proposed method generates comparable outputs to the elaborated below. Fig. 1 Framework of the proposed method. It consists of two computation stages: firstly, to filter the input rainy image using the weighted median filter; then, the weighted median filtered image is used as a guide image and convoluted with the input rainy image to get the final rain-free image Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 3 of 8 2.2 Remove rain steaks via weighted median filtering image. Formally, given f (x, y) as the value of the centered Median filter has been widely used in image denoising pixel (x, y) of a local window R(x, y), then all pixels in the due to its well smoothing effect on noise with long tail window can be expressed as probability distribution as well as better preserving func- R(x, y) =|f (x + k, y + r)|k, r =−1, 0, 1| (1) tion on image details. However, the filtering window size To compute the average value of all pixels in the window has an important effect on the denoising performance of the traditional median filter. A small size of the filtering R(x, y) window, a better detail preserved with a lower denosing 1 1 performance; while a large size of the filtering window, a R(x, y) = f (x + k, y + r) (2) averaged high denosing performance with a poor detail preserved. k=−1 r=−1 To address this problem, the weighted median filter was Let Z , Z is the max and min pixel value in the max min proposed [9]. The main idea of the weighted median local window R(x, y), then for the pixel value f (x, y) of the filter is to replace the current pixel with the weighted centered pixel (x, y), if f (x, y) = Z ,or f (x, y) = Z ,or max min median of neighboring pixels within a local window, as |f (x, y) − R(x, y) | > d , then the pixel (x,y) will be averaged x,y showninFig. 1, where the current pixel I is replaced takenasanoisepixel.Here, d x, y) is a threshold which is with the weighted median of its neighboring pixels within determined by a local window I . This filter has the following special characteristics: 1 1 d x, y) = [ f (x + k, y + r) − R(x, y) ] ( averaged (1) The filtering kernel is not separable. r=−1 k=−1 (2) It cannot be approximated by interpolation or (3) down-sampling. (2) Determining the filtering window size. In order to (3) There is no iterative solution. combine the advantage of both window size of a filter, For the reasons mentioned above, the weighted median the size of the filtering window is determined accord- filter can effectively remove noises from an noised images ing to the number of noise pixels in a local window while not strongly blurring the edges of image structure. R(x, y). Given the number of noise pixels in a local window This is why we employ the weighted median filter for rain R(x, y) as Num(R), then the size of the filtering window is streaks removal in this work. The filter used here con- determined as sists of the following three operation steps: (1) rain streaks 3 × 3, Num(R) ∈{1, 2, 3} detection, (2) determining the filtering window size, and Size(R) = 5 × 5, Num(R) ∈{4, 5, 6} (4) (3) noise filtering. Details of each operation are described 7 × 7, Num(R) ∈{7, 8, 9} as followings: (1) Rain streaks detection. This operation will provided (3) Noise filtering. Formally, given a pixel p in an image basis for rainy image pixel classification. To determine the I, and a local window R(p) of radius r centered at p,for noised pixels, a 3 × 3 window is used to slide over the each pixel q ∈ R(p), we define weighted median filter as Fig. 2 Results with different filtering method. a Input rainy images. b Filtering results using the weighted median filter. c Filtering results using guided filter Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 4 of 8 all pixels in R(p) as{(I(q), w )}. Then, by sorting all pix- pq els values in an ascending order, we can get a weighted median filtered image I(p ),where k n p = minks.t. w ≥ w (6) pq pq q=1 q=1 In order to accelerate weighted median filtering pro- cessing, ideas proposed in [9], including the joint- histogram and median tracking strategy using balance counting box for dynamically finding median, are used in this work. Figure 2 shows part results using weighted median filtering for rain removal. It can be seen that all rain steaks in the input rainy image are well removed after filtered with the weighted median filter, as shown in Fig. 2b. 2.3 Recover texture details from weighted median filtered image using guided filter As can be seen from Fig. 2b, though rain steaks in input rainy images are well removed, part regions in the input image are also over smoothed, where some edge and tex- ture details are removed, for example, the wave lines and textures in the left image of the first row and edges and textures of grass in the left image of the second row are all removed, so those regions look very unnaturally. Aiming at this problem, guided filter [11]isemployedin this section, where we used the weighted median filter- ing image as the guide image, and used input image as the input of guided filter. The main idea of guided filter is to filter input images by considering the content of the guidance image. Formally, given a guidance image I, a guided filtering output image Fig. 3 Comparison of different rain removal methods on synthetic guided I of input image I can be defined as in in rainy images. The top row is the input rainy images, follows in sequence are the ground true, results with Li’s method [10], Manu’s guided I (x) = a I (x) + b , ∀i ∈ ω (7) method [6],Luo’smethod[5], Huang’s method [3], and our method k k k in where ω is a window centered at the pixel x, a and b are k k k assumed to be constant in ω , and determined as w = G(f (p), f (q)) (5) pq i∈ω I p −μ p n i ω k k k where w is a weight based on the affinity of pixel p and q in pq a = (8) σ + ε in corresponding feature map f. f (p) and f (q) are features at pixels p and q in f,respectively. g is a typical influ- ence function between neighboring pixels, which can be b = p − a μ (9) k k k in Gaussian exp−||f (p) − f (q)|| or other forms. Let I(q) denote the value of a pixel at the location q in where μ and σ are the average value and the vari- image I, n = (2r + 1) denote the number of pixels in ance of the input image I in window ω ,respectively. in R(p), we can describe the values and weight elements for n is the number of pixels in window ω . p is the k k Table 1 Quantitative comparison of different methods on synthetic rainy images Li’s method [10] Manu’smethod[6] Luo’s method [5] Huang’s method [3] Our method PSNR 31.93 ± 0.01 24.05 ± 0.01 33.34 ± 0.01 33.98 ± 0.01 35.46 ± 0.01 SSIM 0.81 ± 0.01 0.72 ± 0.01 0.80 ± 0.01 0.83 ± 0.01 0.90 ± 0.01 Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 5 of 8 is why we employed guided filter as a tool for image texture recover. Figure 2c shows the recovered images with guided fil- ter. As can be seen that the removed image texture and edge details by the weighted median filter are well recov- ered, for example, as shown in Fig. 2c, the wave lines and textures in the left image of the first row and edges and textures of the grass in the left image of the second row are very clear to observe, and those regions look very naturally compared with that in Fig. 2b. 3 Experimental results 3.1 Experiment setting To evaluate the proposed method, extensive tests have been done using MATLAB R2015 on a PC with a 2.60 GHz Intel Dual Core Processor and 4G RAM. To the best of our knowledge, there is no standard rainy image data set available for bench marking currently. Hence, we collected totally 100 natural/synthetic rainy images from the Inter- net and also from the test image data set released from [10]. We compare the proposed method with several state- of-the-art methods, including Li’s method [10], Manu’s method [6], Luo’s method [5], and Huang’s method [3], the results of these methods are generated in MATLAB with suggested parameter setting by the authors. For quantitative evaluation of different methods on syn- thetic rainy images, because the ground-truth images are available, we employ the indexes of peak signal-to-noise ratio (PSNR) [12] and Structure Similarity Index (SSIM) Fig. 4 Comparison of different rain removal methods on real rainy images. The top row is the input rainy images, follows in sequence [13] on the luminance channel as evaluation measures. are the results with Li’s method [10], Manu’s method [6], Luo’s For quantitative evaluation of different methods on nat- method [5], Huang’s method [3], and our method ural rain images, because the ground-truth images are usually unavailable, PSNR and SSIM are unable to be used as evaluation measures; we employ an overall qual- ity index proposed in [14] as evaluation measure. This mean of the guided image in window ω . ε is the regu- indicator estimates the average visibility enhancement larization parameter which is used to control the struc- obtained by the restoration method. The higher the value tural similarity. The larger ε is, the smoother the output of the overall quality index is, the better the enhanced will be. visibility is. Edges in an image after using guided filter will change differently. For step edges, it is still step edges after using 3.2 Evaluation on synthetic rainy images guided filter, but their ranges become smaller, which Figure 3 shows part experiment results on synthetic rainy means that the step edges become smoother after guided images [10]. It can be seen that though Li’s method [6] filter. For ridge edges, if the ridge edges with small size are works well on rain removal, the contrast of the output unaffected by the other edges, their variances are close to images is obviously decreased, and details in the dark part 0, then the ridge edges will disappear and tend to the back- of the output images are difficult to observe, as shown in ground; for valley edges will become larger than the input. the third row in Fig. 3, for example, the feet of the old man From what was mentioned above, we can see that the in the third row in Fig. 3b are not clear to be seen. For guided filter has well preserving ability on image edges. Manu’s method [10], as shown in the fourth row in Fig. 3, Therefore, it can be used for image texture recovery. This Table 2 Quantitative comparison of different methods on real rainy images Li’s method [10] Manu’s method [6] Luo’s method [5] Huang’s method [3] Our method 0.2854 ± 0.01 0.1665 ± 0.01 0.3287 ± 0.01 0.1823 ± 0.01 0.3925 ± 0.01 Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 6 of 8 it works well on rain removal. However, the whole image 4 Discussion is over smoothed, the output image looks unnatural, and In this paper, we present a weighted median guided fil- part texture information is lost. For example, because of tering method for rain removal with a single image. To the lost of the textures information in the output images, several mountains become one in the first image in the fourth row, and clouds in the third image in the fourth row are also disappeared. For Luo’s method [5]andHuang’s method [3], it can be seen that there are many rain streaks remained in the output images, as shown in the fifth row and sixth row in Fig. 3. Whereas, with respect to the afore- mentioned methods, the proposed method performs well on these synthetic images, as shown in the seventh row in Fig. 3, and is able to successfully remove majority of the rain streaks while maintaining most details of the original images and achieved a good visual effect, which are very close to the ground true (the second row in Fig. 3). Table 1 shows quantitative evaluation results of differ- ent methods on synthetic rainy images. It can be clearly observed that theproposedmethodisabletoachieve bet- ter quantitative performance compared to all of other four mentioned methods in terms of both of PSNR and SSIM, and this is consistent with what we observed in Fig. 3. 3.3 Evaluation on real rainy images Figure 4 shows part results of this experiment. It is obvi- ously the result with the proposed method which out- performs most of the other state-of-the-art methods in terms of both the effectiveness of removing rain streaks and the visual quality of recovered images. For example, Manu’s method [6] removes many image details and tex- tures from the input image and that makes the image unnatural, as shown in the third row in Fig. 4.Incon- trast, Li’s method [10] (the second row in Fig. 4.) and our method (the last row in Fig. 4.) are able to remove most of the rain streaks and meanwhile produce lesser artifacts on the recovered images. But the results with Li’s method [10] look a little dark and some details are not clear to see, asshowninthe secondrowinFig. 4, whereas that with our method has the sharpest appearance of the trees and leaves, as well as clear people faces, and there no afore- mentioned issues appear, as shown in the last row in Fig. 4. For Luo’s method [5]and Huang’smethod[3], as shown in the fourth row and fifth row in Fig. 4, there are not only many rain streaks that remained in the output images, but also the background is blur which lead difficult to observe image details. Table 2 shows the quantitative evaluation results of dif- ferent methods on real rainy images in terms of the overall quality index. It can be seen that the average overall qual- ity index value of the proposed method are higher than that of other methods, which indicates that the proposed Fig. 5 Comparison of different noising filters for removing raining. The top row is the input rainy images, follows in sequence are the method has better ability in recovering the visibility of results with Gaussian filter, median filter, bilateral filter, guided filter, rainy images compared to other methods. This confirms the weighted median filter, and the ground true our observations on Fig. 4. Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 7 of 8 Table 3 Running time (seconds) for different methods on a 321× 481 image Li’s method [10] Manu’s method [6] Luo’s method [5] Huang’s method [3] Our method 12.74 ± 0.01 31.40 ± 0.01 61.35 ± 0.01 26.86 ± 0.01 1.22 ± 0.01 our best knowledge, our work is the first one to apply 5Conclusions the weighted median filter for rain removal with a single In this paper, we present a simple and effective weighted image. Compared to most of the existing methods, our median guided filter for rain removal with a single image. method does not rely on other image processing modules Different from most existing methods, the proposed for pre- or post-processing, just takes rain steaks as image method does not rely on other image processing mod- noise, which avoids the possible vulnerability of these ules for pre- or post-processing, which avoids the likely techniques when processing images with complex struc- vulnerability of these techniques when processing images tures and makes single-image-based applications applica- with complex structures. Experimental results show that ble in real world. the proposed method produces comparable outputs to the To verify whether the weighted median filter employed state-of-the-art algorithms with low computation cost. in this method can be replaced by other de-noising filters, Abbreviations such as Gaussian filter, median filter, bilateral filter, and HOG: Histogram of oriented gradients; PSNR: Peak signal-to-noise ratio; SSIM: guided filter, we compare their abilities in rain removal Structure similarity index with that of the weighted median filter in this section. Acknowledgements Part experimental results are shown in Fig. 5.FromFig. 5, The authors thank the editor and anonymous reviewers for their helpful it can be seen although all de-noising filters can remove comments and valuable suggestions. most rain streaks, Gaussian filter (shown in the second Funding row in Fig. 5) also smoothed some details and that make This work was supported in part by a grant from the National Natural Science the image looks a little blur, the results both with median Foundation of China (nos. 61202198, 61401355, and 41601353), a grant from the China Scholarship Council (no. 201608610048), the Key Laboratory filter (shown in the third row in Fig. 5) remained too Foundation of Shaanxi Education Department (no. 14JS072), and the Nature many rain steaks. Though the result with bilateral filter Science Foundation of Science Department of PeiLin count at Xi’an (no. and that with the guide filter look better than that with GX1619). Gaussian filter and that with median filter, the results look Availability of data and materials a little blur compared to that with the weighted median We can provide the data. filter. It is obvious that the weighted median filter suc- Authors’ contributions cessfully removes most rain streaks while preserving most All authors take part in the discussion of the work described in this paper. The non-rain image details in these test cases. The main rea- author ZS wrote the first version of the paper. The author LY did part son is that these de-noising methods are mainly designed experiments of the paper. ZM, JB, FY, and CZ revised the paper in different versions, respectively. The contributions of the proposed work are mainly in for removing Gaussian noise with known standard devi- two aspects: (1) To our best knowledge, our work is the first one to apply the ation. However, it is not easy to model the rain streaks weighted median filter for rain removal with a single image. Without any to be removed as Gaussian noise because of the very image priors (e.g., the relationship between the input and desirable output images), our method just takes rain steaks as image noise, which makes different characteristics between them. The main reason single-image based applications applicable in real-world scenarios. (2) The lead to this might be that the most de-noising filters are novelty of our method attributes to the use of weighted median filtering mainly designed for removing Gaussian noise with known images to preserve geometrical details in rain-removed image via guided filter. All authors read and approved the final manuscript. standard deviation, and it is not easy to model the rain streaks to be removed as Gaussian noise because of the Authors’ information very different characteristics between them. For what was Zhenghao Shi received the BS degree in Material Science and Engineering from Dalian Jiaotong University, Dalian, China, in 1995, the MS degree in mentioned above, we select the weighted median filter as computer application technology from Xi’an University of Technology, Xi’an, our choice in this work for rain removing. China, in 2000, and the Ph.D degree in computer architecture from Xi’an We also compare the running time of the proposed Institute of Microelectronics, Xi’an, China, in 2005. In 2000, he joined Xi’an University of Technology, Xi’an, China. From 2000 to 2005, he was an assistant method with that of Manu’s method [6], Li’s method [10], professor in the Department of Computer Science and Engineering at Xi’an Luo’s method [5], and Huang’s methhod [3]onimages University of Technology. From 2006 up to now, he is an associate professor of with a size of 321× 481, as shown in Table 3.Itcan be the Department of Computer Science and Engineering at the same University. Since December 2016, he joined IDBE laboratory, University of North Carolina seen that it only take 1.22 s to obtain the final rain removal at Chapel Hill as a visiting associate professor. During the period of 2006 to image, whereas that with method [10] is 12.74 s, with 2007, of 2008 to 2009, he was on leave with the Department of Computer method [6] is 31.40 s, with Luo’s method [5] is 61.35 s, Science and Engineering at Nagoya Institute of Technology, Nagoya, Japan, for image research as a postdoctral researcher, respectively. From 2007 to and that with Huang’s method [3] is 26.86 s. It is obvious 2008, he was a research associate in the Kurt Rossmann Laboratories for that our method outperforms all of other state-of-the-art Radiologic Image Research, the Department of Radiology, the Division of methods in terms of running time. Biological Sciences, the University of Chicago. His research interests include Zhenghao et al. EURASIP Journal on Image and Video Processing (2018) 2018:35 Page 8 of 8 neural networks for image processing and pattern recognition, 7. JH Kim, C Lee, JY Sim, CS Kim, in 20th International Conference on Image computer-aided diagnosis, and image processing suggested by the human Processing: Sep 15, 2013 - Sep 18, 2013; Melbourne, Australia. Single-image visual systems. He is a member of the IEEE, also a member of ACM. Yaowei Li is deraining using an adaptive nonlocal means filter (IEEE, Australia, 2013), currently study for his master degree in computer science and technology in pp. 914–917 Xi’an University of Technology, Xi’an, China. Her research interests focus on 8. X Zheng, Y Liao, W Guo, X Fu, X Ding, in Neural Information Processing: image processing. Changqing Zhang received his B. S. and M. S. degrees from 20th International Conference, ICONIP 2013: November 3-7, 2013;Daegu, the College of Computer Science, Sichuan University, in 2005 and 2008, Korea, ed. by M Lee, et al. Single-image-based rain and snow removal respectively, and the Ph.D. degree in Computer Science from Tianjin University using multi-guided filter (IEEE, Korea, 2013), pp. 258–265 in 2016. He is an Assistant Professor at the School of Computer Science and 9. Q Zhang, L Xu, J Jia, in Proceedings of the IEEE Conference on Computer Technology, Tianjin University. His current research interests include machine Vision and Pattern Recognition: June 24-27, 2014; Ohio, USA. 100+ times learning, data mining, and computer vision. Minghua Zhao received her Ph.D faster weighted median filter (WMF) (IEEE, USA, 2014), pp. 2830–2837 degree in computer science from Sichuan University, Chengdu, China, in 2006. 10. Y Li, X Guo, J Lu, RT Tan, MS Brown, in IEEE Conference on Computer Vision After that, she joined Xi’an University of Technology, Xi’an, China. Currently, and Pattern Recognition: June 26th - July 1st, 2016; LAS VEGAS, USA.Rain she is an associate professor of the Department of Computer Science and streak removal using layer priors (IEEE, USA, 2016), pp. 2736–2744 Engineering at the same University. Her research interests include image 11. Kea He, Guided image filtering. IEEE Trans. Pattern Anal. Mach. Intell. 35, processing and pattern recognition. Yaning Feng received her B. S. degree 1397–1409 (2013) from Shannxi University of Technology in 1995, received her M. S. degree from 12. Q Huynh-Thu, M Ghanbari, Scope of validity of PSNR in image/video Xian University of Technology in 2004, and received her Ph.D. degree from quality assessment. Electron. Lett. 44, 800–801 (2008) Nogoya institute of Technology in 2008. She is an Assistant Professor of Xian 13. Z Wang, HR Sheikh, AC Bovik, EP Simoncelli, Image quality assessment: University of Technology. Her current research interests focus on computer from error visibility to structural similarity. IEEE Trans. Image Process. 13, vision. Bo Jiang received his Doctoral degree in electronic circuit and system 600–612 (2004) from Chinese Academy of Sciences, Shanghai, China. He is currently the 14. Z Wang, AC Bovik, Universal image quality index. IEEE Signal Process. Lett. associate professor at Northwest University, Shaanxi, China, in 2014. His 9, 81–84 (2001) research interests include aviation remote sensing and image processing. Ethics approval and consent to participate Not Applicable. Consent for publication Not Applicable. Competing interests The authors declare that they have no competing interests. Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Author details School of Computer Science and Engineering, Xi’an University of Technology, No. 5 Jinhua South Road, 710048 Xi’an, China. School of Information Engineering, Northwestern University, No. 1, Xuefu Avenue, Guodu Education and Technology Industrial District, Changan district, 710127 Xi’an, China. School of Computer Science and Technology, Tianjing University, No. 135 Yaguan Road, 300350 Tianjin, China. Received: 25 February 2018 Accepted: 2 May 2018 References 1. AK Tripathi, S Mukhopadhyay, Removal of rain from videos: a review. SIViP. 8, 1421–1430 (2014) 2. Q Zhu, J Yuan, L Shao, in Proceedings of the 2015 IEEE International Conference on Robotics and Biomimetics: December 6-9, 2015; Zhuhai, China, ed. by Ning Xi, YWST Shigeki Sugano, and Q Huang. The current challenges and prospects of rain detection and removal from videos (IEEE Robotics and Automation Society, China, 2015), pp. 843–846 3. DA Huang, YC Frank Wang, LW Kang, CW Lin, Self-learning based image decomposition with applications to single image denoising. IEEE Trans. Multimed. 16, 83–93 (2014) 4. LW Kang, CW Lin, YH Fu, Automatic single-image-based rain streaks removal via image decomposition. IEEE Trans. Image Process. 21, 1742–1755 (2012) 5. Y Luo, Y Xu, H Ji, in Proceedings of the IEEE International Conference on Computer Vision: December 13-16, 2015; Santiago, Chile. Removing rain from a single image via discriminative sparse coding (IEEE, Chile, 2015), pp. 3397–3405 6. BN Manu, in Proceeding of 7th International Conference on Information Technology and Electrical Engineering: October 29-30, 2015; Chiang Mai, Thailand. Rain removal from still images using l0 gradient minimization technique (IEEE, Thailand, 2015), pp. 263–268

Journal

EURASIP Journal on Image and Video ProcessingSpringer Journals

Published: May 28, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off