【作者】 李静；

【导师】 陈帆；

【作者基本信息】 西南交通大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 在计算机技术的推动下,数字图像修复技术受到了越来越多的关注,该技术依据事先设定的方法并通过已知区域对缺失部分进行信息重建。本文首先介绍了图像修复技术的背景与意义,并简单地阐述了图像修复技术的研究现状。接下来,在第二章中给出了基于样本块的几种经典修复模型并详细分析了本文的对比文献算法,最后对样本块修复算法中仍存在的不足进行了总结,本文主要对存在的不足进行了深入地分析与研究。在研究基于样本块的修复方法中,针对样本间相似度函数不够准确,从而不能保持结构特性及纹理一致的问题,设计了一种基于结构张量和块结构化稀疏的图像修复算法。该方法首先利用结构张量定义样本间相似度,寻找更合适的匹配块;然后以该相似度作为块结构化稀疏项构造依据,更有效地区分结构区域,使修复顺序更加稳定,以保持结构特性;最后将结构张量特征信息作为稀疏约束项加入目标函数,提高稀疏表示准确性,保证邻域信息一致性。实验结果表明,本文算法能更好地区分样本间差异并保持结构连贯。基于结构张量和块结构化稀疏的修复算法对破损图像修复后,修复结果仍会出现结构性不高、纹理模糊的情况,同时也会出现一定程度上的错误填充。针对以上问题,本文设计了一种结合图像二维信息熵与相关系数的修复算法。该算法首先采用二维信息熵度量样本块自身的复杂程度,并在此基础上设计了新的优先权计算方法,保证结构连通性;然后引入相关系数构造匹配准则,提高匹配准度;最后对置信度项的更新方式做了改进,从而降低误差累积。实验结果表明,该算法能够在保持修复结果结构连贯性的同时降低错误填充机率。为了使本文的修复算法演示起来更加方便,采用MATLAB GUI设计了一款修复算法效果展示系统。通过该系统用户可以直接选择需要修复的图片进行修复并输出显示修复结果。更多还原

【Abstract】 With the promotion of computer technology,digital image inpainting technique,which reconstructs the missing region according to certain rules and the known information,obtains more and more attentions.Firstly,the background and significance of image inpainting are introduced,and then the research status of the technology is briefly described too.Next,the second chapter gives several classical inpainting model and analyzes the contrast algorithms.Finally,the problems still existing in the exemplar-based inpainting algorithms are summarized and deeply analyzed.Based on study the shortcomings of the current exemplar-based inpainting algorithms such as the inaccuracy of similarity measure,the broken structure and the inconsistent texture,an image inpainting algorithm based on structure tensor and block-structured sparsity is designed.Firstly,the structural tensor is combined with the color to construct the similarity between two patches,which makes it possible to find more suitable matching blocks.And then,using the structure-tensor based similarity,the block-structured sparsity is defined to better distinguish structure region.The structure continuity can be effectively maintained due to more stable fill sequence.Finally,the structure tensor feature information is added as a sparse constraint item to the objective function to improve the sparse representation accuracy,thereby ensuring the consistency of the neighborhood information.Experimental results show that the proposed method can better measure the differences between two patches and maintain the structure coherence of the repaired results.The results repaired by inpainting algorithm based on structure tensor and block-structured sparsity still have structure incoherence,texture unclearness and the mismatching.To address these problems,an image inpainting algorithm combining two-dimensional information entropy and correlation coefficient is put forward.Firstly,the algorithm measures the complexity of patch using two-dimensional information entropy.Based on this,the priority is designed to ensure the structural connectivity;then the matching criteria is constructed by the correlation coefficient to improve the matching accuracy;finally,the update of confidence is improved so as to reduce errors accumulated.Experimental results show that the algorithm can reduce the mismatching while maintaining the structure coherence of results.In order to demonstrate the image inpainting algorithm,this paper designed an image display system using MATLAB GUI.Through the system,user can directly select the damage image to repair and output the repaired results.更多还原