基于字典学习和非局部算子的无伪影图像重构-模式识别与智能系统专业论文.docxVIP

摘要I 摘要 I 摘要 本文主要研究了图像重构的相关问题，包括图像插值、图像超分辨以及图像 去块。本文针对图像超分辨问题和图像去块问题做了三个方面的工作，包括基于 非局部残差增强的图像插值方法，基于字典学习和非局部总变差的图像超分辨方 法以及基于字典学习和非局部总变差的高压缩率图像去块。 传统的图像插值通常只利用局部信息进行图像插值，然而在插值过程中，高 频信息的丢失通常导致图像过平滑现象和模糊效应。针对上述问题，提出了基于 非局部残差增强图像插值方法。在该方法中，引入非局部的思想，利用非局部信 息来补偿丢失的高频信息。本方法可以获得良好的差值结果，不仅补偿了高频信 息获得了清晰的超分辨结果，而且得到更自然的图像和视觉效果。 针对传统字典学习的缺点，提出了一种基于字典学习和非局部总变差的图像 超分辨算法。该方法中，首先利用输入的低分辨图像和低分辨字典求得稀疏系数， 其次利用高分辨字典和稀疏系数重构出高分辨图像，最后利用非局部总变差进行 去振铃效应并对高分辨图像的边缘进行补偿。实验结果表明，本文的方法优于一 些经典的图像超分辨方法。 基于工作二的算法框架，对高压缩率的图像进行去块效应。该工作利用一个 通用的去块字典和非局部总变差实现图像的无伪影图像去块，首先利用去块字典 除去大部分的块效应，其次利用非局部总变差对图像的边缘进行去块效应，以获 得边缘清晰的重构图像。实验结果表明，本文的方法可以得到较好的去块效应结 果。 本论文得到国家自然科学基金(No的资助。 关键词：图像插值 图像超分辨 字典学习 稀疏表示 图像去块 非局部总变差 ABS ABSTRACT III IVTo IV Towards Artifact-free Image Reconstruction with Dictionary Learning and Nonlocal Operators ABSTRACT This thesis mainly deals with the image reconstruction problems including interpolation, super-resolution (SR), and de-blocking. It aims to reconstruct artifact-free reconstruction results from degraded images. The main contributions of this thesis include: first, image interpolation based on nonlocal residual enhancement; second, super-resolution reconstruction based on dictionary learning and nonlocal total variation; third, compression artifact removal using de-blocking dictionary and nonlocal total variation. In the traditional image interpolation, only local neighborhood similarity was usually used to enlarge the image. However, high frequency information is lost, which makes images smooth and blur. Thus, we provided image interpolation based on nonlocal residual enhancement. In the proposed method, the nonlocal information was utilized to enhance the high frequency texture. The proposed algorithm exhibits good performance, not only in effectively recovering the high frequency information but also successfully achieving natural-looking results. Furthermore, we proposed a new image super resolution method in natural images by applying the dictionary to image super-re