目的 传统多样本块稀疏表示的图像修复过程中，容易因为匹配样本块的错误，重构出不正确的填充块，致使在边缘部分产生不连贯的现象，然而单样本匹配填充又容易产生明显的伪影，不能保证与周围结构保持连续，以致产生明显的修复痕迹，影响人的视觉感受。本文提出一种改进的基于相似匹配块组的稀疏表示方法。方法 首先采用颜色信息与余弦距离结合的方法定义图像块匹配准则，在目标邻域范围内获得结构变化趋势更相似的匹配块组；然后在稀疏重构过程中，同时考虑已知信息和估计的未知信息，利用相似块与目标块的匹配程度，对稀疏系数增加不同的权重，以此来增强筛选匹配块的能力，减少纹理模糊现象；最后根据结构稀疏度自适应地在各结构复杂度不同的区域确定样本块尺寸，减少图像修复过程中的错误传播现象。结果 实验结果表明，本文算法改善了在边缘修复过程中产生断裂或者纹理延伸的现象，不仅在主观视觉有明显的提升，其修复结果的峰值信噪比（PSNR）也提高0.5~3dB。结论 本文算法实现了对不同结构特征的彩色破损图像的修复，在结构边缘处有理想的修复效果，并且对各种形状的破损也具有良好的修复效果。
Sparsity image inpainting algorithm based on similar patch group
Lou Xingxin,Tang Xianghong,Zhang Yue(School of Communication Engineering,Hangzhou Dianzi University)
Objective In the process of traditional patch group based image inpainting, it is easy to synthetic the incorrect filling patch which causes incoherence at the edge because of the incorrect matching patches. However, traditional exemplar-based image inpainting algorithm only uses one patch to fill damaged area, which is prone to generate obvious artifacts and it is not guaranteed to remain consistency with the surrounding structure, resulting in obvious inpainting trace and affect people"s visual sense. In order to solve the above mentioned problems, an improved sparsity image inpainting algorithm based on similar patch group was proposed in this paper. Method Firstly, searching similar patch group only use color differences makes it easy to match inconsistent patch whose variation tendency is obviously different with target patch. Since the cosine distance is an effective tool for measuring the change of direction between vectors, it can better measure the proximity in the vector direction. The patch matching criterion is defined by the combination of color information with the cosine distance, it is used to obtain matching patches whose structural change trend is more similar to the target patch. In this way, error matching rate can be obviously reduced when match the similar patches. Secondly, in the process of sparse reconstruction, in order to enhance the ability to filter matched patches and decrease blur texture, we estimate the unknown pixels based on the similar patch group obtained by first step. At the same time, considering that different matching patch have different filling effects on the filling area, in order to make the whole sparse representation model more capable of filtering matching patch, we calculate the matching degree between every similar patch selected from patch group and target patch, which is used to add different weights on the sparse coefficients. Due to image always has rich and varied information, it is difficult to repair the detailed structure of the area with fixed patch size. Finally, the structural sparsity is used to reflect the structural complexity, and the patches of different sizes are adaptively used in different regions in this paper. Different from the method of fixed threshold to determine the patch size, our method repeatedly calculate the structural sparsity under the condition of decreasing the patch size until the structural complexity of the patch is reduced to the average value, which helps reduce error propagation during image inpainting. Result This paper shows several representative color images damaged artificially to different shapes for inpainting in Figure 6 to Figure 9. In addition, it also compares the repair results for the same image damaged with kinds of shape and gives a detailed view of result in Figure 10, and it compares the results with other related inpainting algorithms. The experimental results show that the proposed algorithm is more effective in subjective vision than the other related algorithms and improves the phenomenon of fracture or texture extension during the process of repairing edge structure. More experimental data details are given in Tables 1~3. It can be see not only has a significant improvement in subjective vision, but also has an improvement in peak signal-to-noise ratio(PSNR) and structural similarity index(SSIM). However, due to it needs to repeatedly calculate the patch structure sparsity and dynamically adjust the adaptive patch size, the running time has increased. Conclusion This study proposes a weighted sparsity image inpainting method based on similar patch group. The method combines the features of color information with cosine distance, so that more similar trends to the target patch can be obtained in the process of searching similar patch group. Thus, it will be more precise during patch propagation; Moreover, different sparse weights are added according to the degree of similarity, therefore the structural information is preserved as much as possible during the whole process of inpainting, and the texture blur is reduced. The combination of these two strategies can improve the efficiency of our image inpainting algorithm. The proposed inpainting algorithm in this paper achieves the inpainting of damaged color images with different structural features, has an ideal repair effect at the edge of the structure, and also has a good repair effect on damage of various shapes. We compare existing image inpainting algorithms to prove that the proposed method id more efficient than the existing color image inpainting methods. Furthermore, The improved method in this paper has an ideal repair effect on the strong edge structure and multi-structure area, which effectively improves the image restoration quality, but this method only utilizes the information within a certain range of the image, there are still some shortcomings, such as the unsatisfactory repairing result in the case of inconsistent edge structure and the inability to maintain edge curvature. Thus, our method is not suitable for all kinds of image. Future research will focus on recovery of edge curve and irregular texture breakage.