An Adaptive Richardson-Lucy Algorithm for Single Image Deblurring
Transcription
An Adaptive Richardson-Lucy Algorithm for Single Image Deblurring
Journal of Applied Science and Engineering, Vol. 16, No. 3, pp. 269-276 (2013) DOI: 10.6180/jase.2013.16.3.06 An Adaptive Richardson-Lucy Algorithm for Single Image Deblurring Using Local Extrema Filtering Jiunn-Lin Wu*, Chia-Feng Chang and Chun-Shih Chen Department of Computer Science and Engineering, National Chung Hsing University, Taichung, Taiwan 402, R.O.C. Abstract Motion Blur is one of the common artifacts in digital photographing. With the population of handheld camera and smart phone, image deblurring becomes an important problem. RichardsonLucy algorithm is well-known deconvolution algorithm. But the ringing artifacts usually appear while the estimated point spread function is not accurate. In this paper, we proposed an improved Richardson-Lucy deconvolution algorithm. Before deconvolution step, we separate the blurred image into smooth part and edge part which is called the edge map. The blurred image and edge map are then used for image deblurring. By using the proposed edge map, the ringing artifacts in the deblurred image are significantly reduced while preserving the sharp edge information. Key Words: Motion Blur, Deconvolution, Richardson-Lucy, Ringing Artifacts, Edge Map 1. Introduction Motion Blur is common artifact which produces disappointing blurred image under dim light in digital photography. It is caused by camera shaking or the target object moving during exposure time. Recently, handheld camera and smart phone are getting more and more popular; restoring a blurred image is becoming an important issue. Figure 1 shows the example of image deblurring. In general, motion blur can be separated into two types. The first one is camera shake caused the image blurred; the second one is the target object moving. A blurred image can be viewed as an unblurred image convolution with point spread function (PSF) which is the movement of camera or the movement of target object. If the point spread function is shift-invariant, the image deblurring problem can be transformed into image deconvolution. According to the point spread function is known or not, the deconvolution can be classified into non-blind deconvolution [1-4] and blind deconvolution [5-8]. Blind deconvolution is an ill-pose problem, it uses only blurred image to restore image. Therefore image*Corresponding author. E-mail: [email protected] pair deblurring method [9,10] is proposed to solve motion blur problem, but the additional image is hard to acquire. Nonblind deconvolution uses known point spread function and blurred image to restore deblurred image. Wiener filter [11] and Richardson-Lucy deconvolution [2] are well-known nonblind deconvolution algorithms which are good at image deblurring. But if the point spread function is not accurate, the deblurring result will appear undesired ringing artifacts at the smooth region around edges. To solve this problem, Wang et al. [3] proposed using different weight values for smooth region and edge region during the deblurring process to suppress ringing. But it still failed around some edges. Zhao et al. [12] suggested that using only two weight values for describing smoothness information of image is not enough. Figure 1. Example of image deblurring. 270 Jiunn-Lin Wu et al. In this paper, we proposed an adaptive image deblurring algorithm using local extrema filtering. To reduce ringing artifacts from deblurred image, the extrema-based multiscale decomposition filter is used to distinguish the edge region and smooth region, and then edge map is obtained. According to Zhao et al. suggestion, it is better that existing more than two weight values in edge map instead of only two weight values in edge map. Thus we use blurred image and edge map which contains more than two weight values are both used to deconvolution procedure. Finally the deblurred image is obtained with less ringing. The rest of the paper are organized as follows. Related works which is mentioned in Introduction are described in detail. Then according to related works our proposed method is presented. And then experiments are designed for verifying our deblurring algorithm. Finally, we conclude this paper. 2. Related Works In general, a blurred image can be modeled as (1) where I is blurred image; L is latent unblurred image; k is point spread function; n is noise and Ä is convolution operator. According to point spread function is known or not, image deconvolution can be separated into nonblind deconvolution and blind deconvolution. Richardson-Lucy algorithm [2] is a well-known iterative deconvolution method for image deblurring. A more clearly deblurred image is generated for each time of iteration. After several iterations, a sharp and unblurred result image can be obtained. According to Bayes’ theorem, (1) can be transform into spread function is not accurate enough, the ringing artifacts will appear in the result image. Figure 2 shows that if the point spread function is not precise enough, then the ringing artifacts will be generated in the result image. Therefore, to reduce ringing artifacts is an important issue for image deblurring. Several methods have been proposed to solve this problem. Shan et al. [13] suggested that the smooth region in the blurred image is also smooth in deblurred image. Therefore, Shan included a constraint about the smooth region called local prior during the restoration process. The formulation of local prior is defined as (4) where i is pixel position in image, L is latent unblurred image, I is blurred image s is the standard deviation and W is smooth region in image. Figure 3 shows the corresponding smooth region of original image. For this reason Wang et al. proposed improved Richardson-Lucy algorithm using local prior. It includes an adaptive weight map called edge map during deblurring process. Before deconvolution, the blurred image is used to extract the edge map which indicates where the smooth or edge region it. And then the edge map and blurred Figure 2. Example of the ringing artifacts. (a) Original image. (b) Blurred image. (c) The result of Richardson-Lucy algorithm. (2) Then the formulation of Richardson-Lucy algorithm can be defined as (3) where * is correlation coefficient operator, Ä is convolution operator, t is the tth of iteration, Lt is current iteration deblurred image, Lt+1 is the next iteration deblurred image, k is PSF and I is blurred image. If the point Figure 3. Example of edge map. (a) Blurred Image. (b) The corresponding smooth region of blurred image. Adaptive Richardson-Lucy Algorithm for Single Image Deblurring image are used as inputs for each time of iteration. When the ringing is appear during deconvolution process, the edge map will reduce the effect of ringing by assigning the smooth region where ringing appeared large weight to update it. Then the ringing artifacts can be efficiently reduced from result. From Figure 4, we can notice that Wang’s method reduces most ringing artifacts, but the ringing artifacts around the region where between two edges is not reduced successfully. In next section, we will discuss this problem, and then proposed our deblurring algorithm. 3. Proposed Method In this section, we will propose an adaptive Richardson-Lucy deconvolution algorithm using local extrema filtering. Observe on the result which is deblurred by Wang et al, the region between two edges is not fully described by Wang’s edge map. Consequently, the ringing artifacts still exist in deconvolution result. To get an appropriate edge map is an important issue for obtaining a good deconvolution result. As a result, there are two problems for extracting edge map, first is well description edge map about the smooth and edge region; the other is only two weight values in edge map is not enough. Extrema-based multiscale decomposition (EMD) [14] is good at separating detail and mean layer. Compare with existing edge-preserving image decomposition algorithms, EMD can extract the detail from input image Figure 4. Comparison of deblurring results. (a) The result using Richardson-Lucy. (b) The result using Wang’s method. (c) The corresponding magnitude of (a). (d) The corresponding magnitude of (b). 271 while preserving edge mean and contrast properly. Its idea is getting the local extrema such that maxima or minima, and then getting the mean layer by averaging maximum layer and minima layer, the mean layer is also called smooth layer. Then we extract the detail layer by subtracting smooth layer from input image. As mentioned before the edge map must have good description at smooth region, thus we improved EMD algorithm for purpose. First EMD is used for separating the detail layer and mean layer, and then the contour is extracted from mean layer by calculate the gradient of input image, called smooth map. For further edge map information, the gradient of detail layer is also used for describing the detail part called detail map. Finally, Combining the detail map into smooth map, then the result map is the proposed edge map. The flowchart of our improved EMD method is shown in Figure 6. In Figure 7(a), we can observe that the smooth map contains most contours, so we can extract the contours which are strong edges (Figure 7(c)) from smooth map. And the detail map (Figure 7(b)) contains most texture information, thus we extract further detail information (Figure 7(d)) from detail map. But the detail map also contains undesired noise information, so a simple threshold method is used for filtering out the noise. Compare proposed edge map with the edge map extracted by Wang’s method in Figure 8; the region between two strong edges is clear defined in our proposed edge map. Wang’s edge map only separate the blurred image into smooth region and edge region, it is too strict to define Figure 5. Overview of extrema-based multiscale decomposition algorithm [12]. 272 Jiunn-Lin Wu et al. Figure 6. Flowchart of our improved EMD method. the edge map using only two weight values. For example, some regions contain ringing artifacts or noise may be classified as an edge region. As a result, the ringing artifacts will still exist in result image. On the other hand, our edge map defines the smooth region and edge region more flexible, so the deblurring result can be reduced more ringing artifacts. After obtained a reasonable edge map, blurred image and edge map are both used as inputs for deconvolution. During the deconvolution process, the edge map is used as the local prior of Richardson-Lucy algorithm, and then the cost function Epl (L) = -log(pl (L)) can be defined as Figure 7. Intermediate results of our improved EMD. (a) Smooth map. (b) Detail map. (c) The result of gradient of smooth map and then use threshold method. (d) The gradient of detail map. Figure 8. Comparison between our method and others. (a) Proposed edge map. (b) Wang’s edge map. (5) where ÑEpl (Lt) is defined as where L is latent unblurred image, I is blurred image, M is edge map which describes the smoothness information of the image, ¶x is the partial derivative operator in x direction and ¶y is partial derivative operator in y direction. Combine (5) with Richardson-Lucy algorithm, it can be defined as (6) (7) where * is correlation coefficient operator, ¶x* is conjugate matrix of ¶x, ¶y* is conjugate matrix of ¶y, t is the number of iteration, Lt is current deblurred image, Lt+1 is the deblurred image of next iteration, k is PSF, I is blurred image, M is edge map and l is the predefined parameter. The comparison of the proposed method with Wang’s approach is shown in Figure 9; we observe that the ringing artifacts within red rectangle are sig- Adaptive Richardson-Lucy Algorithm for Single Image Deblurring nificantly reduced in our result. In the other word, our deblurred image has better visual result than Wang’s method with similar PSNR. It is obvious that the ringing artifacts can be efficiently reduced by using the proposed edge map in the deconvolution step. The flowchart of our proposed method is shown in Figure 10. 4. Experiments and Discussions The aim of the proposed method is to reduce the ringing artifacts in the deblurred images. Therefore, in this section, we will use several well-known natural images which failed in Richardson-Lucy algorithm, and the synthesis images to demonstrate the effectiveness of the proposed method. For all test images, at first the traditional Richardson-Lucy algorithm is used as a failed example with ringing artifacts. Then several deconvolution methods with edge map are used to recover blurred image and reduce ringing artifacts. Then we will compare our result with the traditional Richardson-Lucy result and other 273 methods in detail. For the first experiment, we use the blurred image “old man” which its size is 800 ´ 532 and PSF which its size is 27 ´ 27 to verify our method. There are many wrinkles on the old man face and hands, so we can observe these two parts to verify our method. Figure 11 shows the undesired result of Richardson-Lucy algorithm. And then we compare our result and edge map with Wang’s method in Figure 12. We can observe that compare with Wang’s method, our edge map provides more edge information. Therefore we can reduce more ringing artifact while preserving the edge. Figure 13 is a magnitude view of our result and others. Compare with traditional Richardson-Lucy algorithm, our proposed method and Wang’s method can efficient reduce the ringing artifacts, but our method has better image vision than Wang’s. Next experiment we Figure 10. The flowchart of our proposed method. Figure 9. Comparison of deconvolution result (a) Original image (512 ´ 512). (b) PSF (21 ´ 21). (c) Wang’ method with 40 times iteration (PSNR is 27.16). (d) Our proposed method with 40 times iteration (PSNR is 26.72). Figure 11. The result of traditional Richardson-Lucy deconvolution algorithm using “Old man”. (a) Blurred image and PSF. (b) Deblurred result. 274 Jiunn-Lin Wu et al. use blurred image “flower” which its size is 701 ´ 494 and bigger PSF which its size is 35 ´ 35 to verify our method. The “flower” has more texture then “old man”. Figure 14 is the result of Richardson-Lucy algorithm. We can observe that Figure 15(a) has less ringing artifacts, but some region is failed to recover sharp edge. In contrast, our result in Figure 15(c) has sharper result. From Figure 16, we can see that our method has the best image vision of the three. Next, we want use an image which contains a lot of narrow detailed to verify the robustness of proposed method. So we use he blurred image “doll” which its size is 800 ´ 558 and bigger PSF which its size is 41 ´ 41 to verify our method. Figure 17 is the result of Richardson-Lucy algorithm. From Figure 18, the detail part is well preserved, and the ringing artifacts are reduced successfully. Figure 14. The result of traditional Richardson-Lucy deconvolution algorithm using “Flowers”. (a) Blurred image and PSF. (b) Deblurred result. Figure 12. Comparison results with other methods using “Old man” (a) Wang’s result. (b) Wang’s edge map. (c) Our proposed method. (d) Our corresponding edge map. Figure 15. Comparison results with other methods using “Flowers” (a) Wang’s result. (b) Wang’s edge map. (c) Our proposed method. (d) Our corresponding edge map. Figure 13. Close-ups which is extracted from Figures 11(b) and 12(a), (c). (a) Richardson-Lucy algorithm. (b) Wang’s method. (c) Our proposed method. Figure 16. Close-ups which is extracted from Figures 14(b) and 15(a), (c). (a) Richardson-Lucy algorithm. (b) Wang’s method. (c) Our proposed method. Adaptive Richardson-Lucy Algorithm for Single Image Deblurring The close-ups show proposed method reduces more ringing artifacts where near the contour of the doll’s face than other deconvolution methods. The processing time of deblurring using all natural images is show in Table 1. Finally, we use the synthesis images shown in Figure 20 to verify our proposed method. The synthesis images include, “Lena”, “Airplane” and “Fruit”. The image size Figure 17. The result of traditional Richardson-Lucy deconvolution algorithm using “Doll”. (a) Blurred image and PSF. (b) Deblurred result. 275 of burred “Lena”, burred “Airplane” and burred “Fruit” are 490 ´ 490, 486 ´ 486 and 490 ´ 458 respectively. The size of corresponding PSF are 23 ´ 23, 19 ´ 19 and 23 ´ 23 respectively. The result comparisons of the proposed method with Wang’s algorithm are shown in Figure 20. It is obvious that the proposed method has better visual result than Wang’s approach. Table 2 shows that the proposed method has similar PSNR with Wang’s method. The processing times for deblurring examples is shown in Table 2. We implemented our method with Visual C++. Net language. The testing environment was as PC running Window 7 64 bit version with AMD Phenom II X4 945 3.4 GHz. 4 GB Ram. The processing time of the proposed method is too long to be waiting for. The FFT (Fast Fourier transform) which is mentioned in many existence deconvolution methods can be used to solve this problem. Table 1. The comparison of processing time between the proposed method and other approaches Processing time (second) Image Proposed method Wang’s method Richardson-Lucy Old man Flower Doll 610.5927 556.0106 816.6494 178.457 200.890 342.269 139.32 170.90 0302.936 Figure 18. Comparison results with other methods using “Doll” (a) Wang’s result. (b) Wang’s edge map. (c) Our proposed method. (d) Our corresponding edge map. Figure 19. Close-ups which is extracted from Figures 17(b) and 18(a), (c). (a) Richardson-Lucy algorithm. (b) Wang’s method. (c) Our proposed method. Figure 20. Deblurring result of synthesis image. First row: blurred image and its PSF. Second row: the result of proposed method. Third row: the result of Wang’s method. (a) “Lena”. (b) “Airplane”. (c) “Fruit”. 276 Jiunn-Lin Wu et al. Table 2. The comparison between the proposed method and Wang’s method Processing time (second) PSNR (dB) Image Lena Airplane Fruit Proposed method Wang’s method Proposed method Wang’s method 28.9105 27.7854 30.7137 29.2214 27.7699 30.8389 320.7099 298.3994 298.7241 77.227 61.108 71.791 To reduce time complexity of the proposed algorithm will be the main task for future work. Our method shares common limitation with others deconvolution method. Huge PSF, bright spot, severe noise, and nonuniform blur will cause the deblurring failed. To solve these problems in the proposed method will be interesting future work. 5. Conclusions In this paper, we proposed an adaptive RichardsonLucy deconvolution algorithm using local extrema filtering. First, the local extrema filtering called EMD is used to extract the edge map of the image which contains well definition on smooth regions, and then the edge map and blurred image are both used as inputs for deblurring algorithm. With the effort of edge map, the ringing artifact can be reduced for each time of iteration. Finally, a shaper deblurred image without ringing can be obtained. The experiments show that our proposed method can reduce the ringing artifacts effectively while preserving the edge of the image. Acknowledgements This research is partially supported by National Science Council Grant NSC 101-2221-E-005-088. References [1] Levin, A., Fergus, R., Durand, F. and Freeman, W. T., “Image and Depth from a Conventional Camera with a Coded Aperture,” ACM Trans. Graph., Vol. 26, No. 3, p. 70 (2007). doi: 10.1145/1276377.1276464 [2] Richardson, W. H., “Bayesian-Based Iterative Method of Image Restoration,” JOSA, Vol. 62, No. 1, pp. 55-59 (1972). doi: 10.1364/JOSA.62.000055 [3] Wang, Y., Feng, H., Xu, Z., Li, Q. and Dai, C., “An Improved Richardson-Lucy Algorithm Based on Local Prior,” Optics & Laser Technology, Vol. 42, No. 5, pp. 845-849 (2010). doi: 10.1016/j.optlastec.2010.01.001 [4] Yuan, L., Sun, J., Quan, L. and Shum, H., “Progressive Inter-Scale and Intra-Scale Non-Blind Image Deconvolution,” ACM Trans. Graph., Vol. 27, No. 3 (2008). doi: 10.1145/1360612.1360673 [5] Levin, A., Weiss, Y., Durand, F. and Freeman, W. T., “Understanding and Evaluating Blind Deconvolution Algorithms,” Proc. CVPR, pp. 1964-1971 (2009). doi: 10.1109/CVPRW.2009.5206815 [6] Kundur, D. and Hatzinakos, D., “Blind Image Deconvolution,” IEEE Signal Processing Magazine, Vol. 13, No. 3, pp. 43-64 (1996). doi: 10.1109/79.489268 [7] Fergus, R., Singh, B., Hertzmann, A., Roweis, S. T. and Freeman, W. T., “Removing Camera Shake from a Single Photograph,” ACM Trans. Graph., Vol. 25, No. 3, pp. 787-794 (2006). doi: 10.1145/1141911.1141956 [8] Xu, L. and Jia, J., “Two-Phase Kernel Estimation for Robust Motion Deblurring,” Lecture Notes in Computer Science, Vol. 6311, Computer Vision - ECCV, pp. 157-170 (2010). doi: 10.1007/978-3-642-155499_12 [9] Yuan, L., Sun, J., Quan, L. and Shum, H., “Image Deblurring with Blurred/Noisy Image Pairs,” ACM Trans. Graph., Vol. 26, No. 3 (2007). doi: 10.1145/ 1276377.1276379 [10] Zhuo, S., Guo, D. and Sim, T., “Robust Flash Deblurring,” in Proc. CVPR, pp. 2440-2447 (2010). doi: 10.1109/CVPR.2010.5539941 [11] Gonzalez, R. C. and Woods, R. E., Digital Image Processing, 2nd ed., Prentice Hall (2002). [12] Zhao, J. F., Feng, H. J., Xu, Z. H. and Li, Q., “An Improved Image Restoration Approach Using Adaptive Local Constraint,” Optik, Vol. 123, pp. 982-985 (2012). doi: 10.1016/j.ijleo.2011.07.014 [13] Shan, Q., Jia, J. and Agarwala, A., “High-Quality Motion Deblurring from a Single Image,” ACM Trans. Graph., Vol. 27, No. 3 (2008). doi: 10.1145/1360612. 1360672 [14] Subr, K., Soler, C. and Durand, F., “Edge-Preserving Multiscale Image Decomposition Based on Local Extrema,” in SIGGRAPH Asia, Singapore (2008). Manuscript Received: Feb. 27, 2013 Accepted: Jun. 28, 2013