【作者】 张晴；

【导师】 林家骏；

【作者基本信息】 华东理工大学 ， 控制科学与工程， 2012， 博士

【摘要】 数字图像修复是指计算机按照某种特定的方式对图像中丢失信息的区域进行自动填充,并要求其修复痕迹不为人眼所察觉的一种半自动化的图像处理技术。图像修复其本质是通过不完全信息重构出完全信息,这一直是计算机视觉和人工智能领域一个梦寐以求的目标,同时也是许多其他图像处理技术需要处理的问题。研究数字图像修复技术对于缺乏一定先验知识但又能有效估计信号情况的计算机视觉和人工智能问题具有重要的理论研究意义及广阔的应用前景。本文研究工作围绕基于样本的图像修复技术展开,对基于样本的图像修复技术的基本理论和现有方法进行了阐述和总结。从算法修复效率和修复质量这两方面展开研究工作,针对现有方法中存在的不足之处,分别从贪婪填充策略和全局建模优化这两个角度提出了一些新的算法和解决方案,并从理论和实验的角度分析和验证了新算法的有效性和普遍适用性。主要的创新工作体现在以下几个方面：1)提出了基于局部平均灰度熵的快速图像修复算法,在不影响修复质量的前提下,加快传统图像修复算法的计算机执行速度。分析现有图像修复算法的复杂度可知,在每次迭代计算中,样本块与待修复块的一一匹配操作是图像修复算法耗时的主要原因,因此样本块集合的大小,即样本搜索采样范围对计算运行时间至关重要。本文根据局部平均灰度值的一维熵来度量待修复区域周围已知信息的复杂程度,在每次迭代计算中根据此邻域复杂度来自动选择搜索采样范围大小以提高修复速度。2)提出了基于色差分析和特征统计的图像修复算法,较现有算法而言能较好连接破损边缘和保持结构完整性。分析了现有的基于样本的图像修复算法所采用的修复优先权方法的不足之处以及其对图像修复质量的影响,根据局部色差分析,提出了两种基于特征统计的修复优先权方法：基于距离统计的修复优先权方法和基于相似性统计的修复优先权方法。实验证明,提出的方法较经典的基于等照度线的修复优先权方法而言能更好得分辨图像的结构成分,修复图像的主要结构信息、连接边缘,之后再修复纹理成分,获得令人满意的修复结果。3)提出了动态加权匹配的图像修复算法,更好得利用了图像的已知信息,提高了图像修复质量。分析了现有的最优匹配块选择方法存在的问题：使用单个样本块的填充方法未能充分利用图像的已知信息；根据距离确定加权系数的加权合成方法可能会模糊修复结果。同时,为了更好地保持填入信息与原有信息的一致性,引入一致性和一阶梯度信息作为约束条件。4)提出了仅仅旋转位于边缘样本块的算法,以平衡修复质量的提高和计算代价的降低之间的冲突矛盾。为了更充分利用图像的已知信息,采用旋转的方法增加样本空间以提高最终的修复质量,但旋转全体样本块必定大大增加计算开销。为了兼顾计算速度,考虑到自然图像的一些特性,提出只旋转位于边缘的样本块以增加样本空间,并用实验证明所提算法的合理性和优越性。5)提出了改进的基于置信传播的图像修复算法,更好衡量块与块之间的相似性,提高最终的修复质量。分析了采用贪婪填充策略进行图像修复的不足之处,以及现有的采用图像全局能量建模的置信传播算法进行图像修复存在的两个问题：仅用欧几里得距离度量图像块与块之间的近似程度不足以衡量视觉上的相似性：能量函数的平滑项和数据项的权重一致性不足以表达已知邻域信息的重要性。本文从这两个问题着手,改进了基于置信传播的图像修复算法,并用有环置信传播方法求解其最优近似解。更多还原

【Abstract】 Image inpainting, which is a semi-automatic image reconstruction technology, is an art of modifying an image or video in a form that is not easily detectable by an ordinary observer, and has become a fundamental research area in image processing. Although stating the image inpainting problem is very simple, the task that actually tries to successfully solve it is far from being a trivial thing to achieve. It is essentially to reconstruct the missing information from the known part, which is not only the desired aim of image processing and computer vision but also the problem with which the image processing would be involved.The study on the image inpainting is of significance in both the theoretical research and practical applications such as special effect, image editing and compression, super-resolution image reconstruction and so on.This dissertation mainly focuses on the exemplar-based image inpainting which is also known as image completion, explores and sums up some existing theory and schemes, and pro-poses several novel exemplar-based inpainting algorithms based upon greedy filling-in strategy and global optimization to overcome certain limitations of existing inpainting methods. A large number of experiments are used to validate the effectiveness and universality of our proposed algorithms. The main innovation is as follows:1) We propose a novel rapid image inpainting scheme that could cut down the compu-tation time without reducing the quality of inpainted results. In each iteration, the matching operation between the target patches and the source patch one by one is the primary reason of the high computation cost; as a result, the size of target patch set, namely the area where the target patches are from, is quite important. This paper determines adaptively the area where the algorithms samples target patches in accordance with the entropy of local average distribution of pixel values around the missing region.2) We analyze the influence of the current isophote-driven filling-in order on the perfor-mance and propose two novel inpainting priority methods according to the statistics of color distribution:distance-based and similarity-based. The experiments show that our proposed methods can identify the structures, inpaint the main missing structure, connect the broken edges in a better way and achieve impressive results.3) We propose a novel algorithm to generate the best matching patch to fill in the cor-responding missing position of source patch, which can better utilize the known information and improve the quality of inpainted results. Existing representative methods have some prob-lems because they merely use a single target patch as the best matching patch or they simply reconstruct target patch with a weighted similarity function. 4) To make better use of the known information, we propose a novel scheme to extend the sample space through rotating existing target patches. However, the rotation step would make the computation cost intolerable because the original sample space is so large. To take into account both inpainting performance and computation cost, we propose to rotate the tar-get patches on the edges under the considerations of several characteristics of natural images. Moreover, we validate the feasibility and superiority of our proposed algorithm.5) We improve the inpainting algorithm proposed by Komodakis and use the loopy belief propagation method to get the optimal solution. Current algorithms which minimize a global energy of the MRF have certain problems, for example, the Euclidean distance is not enough to measure the similarity of two patches and the same weight coefficients of the two terms in the energy function do not show their respective signification. The experiments show that our proposed algorithms enhance the quality of inpainted results.更多还原