基于样本和稀疏表示的图像修复方法研究

发布时间：2018-06-07 03:36

本文选题：图像修复 + 样本　；参考：《西北大学》2016年博士论文

【摘要】：图像修复是图像处理和模式识别领域中非常重要的分支,近年来已经引起越来越多研究人员的关注。其基本思想是根据破损图像中的有效信息,对破损区域中的缺损信息进行有效估计,使修复之后的图像在整体上更加协调,并且使不熟悉原始图像的人觉察不到修复痕迹。目前,图像修复技术在老照片和珍贵文献资料的修复、文物保护、机器人视觉等各个领域发挥了越来越重要的作用。因此,对图像修复方法进行广泛深入研究具有非常重要的现实意义。本文首先对图像修复的相关方法以及稀疏表示的相关理论进行研究。在此基础上,针对现有方法中存在的修复顺序不合理问题、错误匹配问题以及贪婪性问题,对大尺度破损的图像修复展开深入研究。论文的主要工作和创新点如下：(1)提出一种基于结构稀疏度的图像修复方法。针对传统的图像修复方法中存在优先权迅速降低导致修复顺序不合理的问题,利用块结构稀疏度和邻域像素差异度对优先权进行定义,使修复顺序更加合理。其次,针对修复过程中存在错误匹配以及错误累积的问题,利用块间距离和块内距离对匹配规则进行定义,可以有效避免图像块的错误匹配,并避免错误不断累积,进而提高修复效果。(2)提出一种基于样本和稀疏表示的混合修复方法。针对传统方法存在块的错误匹配以及错误不断累积的问题,将基于样本的修复方法与基于稀疏表示的修复方法进行融合,利用各自方法的优点,使它们互为补充。如果没有发生错误匹配,则用基于样本的方法进行修复,可以保持纹理的多样性和完整性;否则用基于稀疏表示的方法进行修复,可以及时纠正匹配错误,防止错误不断累积。该方法可以使修复图像符合人类主观视觉要求。(3)提出一种基于主成分分析(Principal Component Analysis, PCA)分类的快速图像修复方法。针对传统的图像修复方法中需要全局遍历搜索匹配样本块,导致比较耗时的问题,利用PCA方法将图像块分为平滑块、边缘块和纹理块三类。对于平滑块,采用基于DCT字典的稀疏表示方法进行修复,不需要全局遍历搜索样本块；对于边缘块,将其搜索区域设置为其周围邻域,缩小搜索区域；对于纹理块,为了保证纹理的多样性,仍采取全局遍历搜索。该方法可以有效减少匹配样本块的搜索时间,因而可以提高图像的修复效率。(4)提出一种基于形态成分分析(Morphological Component Analysis, MCA)的边缘提取方法。从图像修复的角度出发,边缘提取的根本目的是提取对象的主要边缘轮廓,并且尽可能避免由复杂纹理细节造成的孤立和琐碎的边缘。基于以上考虑,首先利用MCA方法把图像进行分解,得到平滑层和纹理层,然后在平滑层上利用Otsu算法估计自适应阈值,最后根据非极大值抑制算法对图像的边缘进行提取。该方法可以避免过多复杂纹理对边缘图像的影响,使边缘图像只保留对象的主要轮廓。(5)提出一种基于边缘引导和非局部均值的修复方法。针对传统的图像修复方法不能很好地保持对象轮廓的连续性和完整性的问题,首先利用基于MCA的边缘提取方法提取边缘图像,并对破损的边缘进行修复。并针对非局部均值的修复方法容易导致纹理细节模糊的问题,提出一种基于非局部均值的自适应方法。然后,在已修复边缘的引导下,利用非局部均值的自适应修复方法分别对破损图像的边缘区域和其余区域进行修复。该方法可以有效保护对象轮廓的连续性,提高图像修复效果。
[Abstract]:Image restoration is a very important branch in the field of image processing and pattern recognition. In recent years, more and more researchers have attracted more and more attention. The basic idea is to effectively estimate the defect information in damaged area according to the effective information in damaged image, so that the image after repair is more coordinated and unfamiliar. At present, image restoration technology has played a more and more important role in the restoration of old photos and precious literature, cultural relics protection, robot vision and so on. Therefore, it is of great practical significance to carry out extensive and in-depth research on image restoration methods. In this paper, the main work and innovation of this paper are as follows: (1) the main work and innovation of this paper are as follows: (1) a kind of knot based on the knot is put forward. In view of the problem that the priority of the traditional image restoration is rapidly reduced, the priority is defined by the block sparsity and the neighborhood pixel difference degree, which makes the repair order more reasonable. Secondly, there are error matching and error accumulation in the repair process. With the use of inter block distance and intra block distance to define the matching rules, it can effectively avoid error matching of image blocks, avoid continuous accumulation of errors and improve the effect of repair. (2) a hybrid restoration method based on sample and sparse representation is proposed. The problem is to combine the repair method based on the sample and the repair method based on the sparse representation, and make use of the advantages of each method to complement each other. If there is no error matching, the sample based method is used to repair it, and the diversity and integrity of the texture can be maintained; otherwise, the method based on sparse representation can be used to repair it. In time, matching errors can be corrected in time to prevent errors from accumulating. This method can make the restored image conform to the human subjective vision requirements. (3) a fast image restoration method based on the Principal Component Analysis (PCA) classification is proposed. Block, resulting in more time-consuming problems, the image block is divided into three categories: smooth block, edge block and texture block by using the PCA method. For smooth blocks, the sparse representation method based on DCT dictionary is used to repair, without global traversal search sample blocks; for edge blocks, the search area is set to its neighborhood, and the search area is narrowed. Texture block, in order to ensure the variety of texture, still take global traversal search. This method can effectively reduce the search time of the matching sample block, and thus can improve the efficiency of image restoration. (4) a method of edge extraction based on Morphological Component Analysis (MCA) is proposed. The fundamental purpose of edge extraction is to extract the main edge contour of the object, and avoid the isolated and trivial edges caused by the details of the complex texture as much as possible. Based on the above consideration, the image is decomposed by the MCA method, and the smooth layer and texture layer are obtained. Then the adaptive threshold is estimated by Otsu algorithm on the smooth layer, and the final root is obtained. The edge of the image is extracted according to the non maximum value suppression algorithm. This method can avoid the influence of too many complex textures on the edge image, so that the edge image can only retain the main contour of the object. (5) a restoration method based on edge guidance and non local mean is proposed. The problem of the continuity and integrity of the profile is to extract the edge images by using the edge extraction method based on MCA and repair the damaged edges. A self-adaptive method based on the non local mean is proposed for the problem that the non local mean restoration method can easily lead to the blurred texture details. At the same time, the adaptive repair method of non local mean is used to repair the edge region and the rest of the damaged image respectively. This method can effectively protect the continuity of the object contour and improve the image restoration effect.
【学位授予单位】：西北大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】