【作者】 张辉；

【导师】 苏杨；

【作者基本信息】 武汉理工大学 ， 信息与通信工程， 2017， 硕士

【摘要】 在互联网大数据时代下,图像作为信息数据传递与共享的有效载体,已经成为人们在信息交流中不可或缺的一部分。然而图像在传输过程中,往往会受到噪声的干扰,这种干扰将对图像的视觉效果以及图像的后期处理造成严重影响。因而,减少图像的噪声干扰显得尤为重要。尽管人们提出了大量图像去噪的算法,但该问题的研究仍有待深入。近年来,基于图像自相似性,并引入稀疏表示和低秩(低秩可以看成稀疏表示的矩阵形式)理论,在图像去噪方面取得了较大的突破,成为当前研究的热点。本文主要以图像自相似性为基础,通过对当前优秀的稀疏表示和低秩的非局部图像去噪算法进行分析,进行了优化和改进。具体工作如下:(1)分析传统图像去噪和经典非局部图像去噪理论和方法,以实验为基础展开对非局部图像去噪的研究,最终验证了非局部算法在图像去噪上的优势。(2)引入稀疏表示相关理论,重点分析了基于稀疏表示的NCSR非局部图像去噪算法。在明确NCSR算法中局部字典的学习以及稀疏系数非局部估计采用欧氏距离时没有考虑图像的亮度与结构信息的缺点的基础上,改进了自相似性度量的距离定义,即通过引入结构相似度(SSIM)将图像的结构以及亮度考虑进来。相关实验表明,改进后的算法不仅在PSNR和SSIM有所提高,而且视觉效果上也有一定的改善。(3)引入低秩相关理论,重点分析了基于低秩的SAIST非局部图像去噪算法,并在此基础上提出改进。a)考虑到相似块聚合时受到噪声干扰而影响聚合效果,引入了DCT预滤波处理方式,有效的解决了聚合时噪声的干扰;b)针对样本均值因没有考虑图像的相关性而影响奇异值阈值,将图像块之间相关性考虑进来,提出了基于加权的样本均值计算,使得奇异值阈值更具有适应性。相关实验表明,基于改进算法处理后的噪声图像在客观指标与主观评价上都有一定提高。更多还原

【Abstract】 In the era of large Internet data,the image has become an indispensable part of the exchange of information as an effective carrier of information data transfer and sharing.However,In the process of image transmission,the image signal often subject to noise interference,this interference will have serious impact on image visual effects and image post-processing.Therefore,it is very important to reduce the noise interference of the image.Although there are a large number of image denoising algorithms have been proposed,the research of this problem still needs to be deep investigated.In recent years,there are a great breakthrough in the image denoising based on image self-similarity and the introduction of sparse representation and low rank(low rank can be regarded as matrix form of sparse representation)theory,it becomes the current research focus.This thesis is based on image self-similarity,Analyzing the current excellent sparse representation and low rank non-local image denoising algorithm,optimizing and improving it.The special works in concrete are mainly as follows:(1)Analyze the traditional image denoising and classical non-local image denoising theory and method.Based on the experiment,research on non-local image denoising.Finally,verified the advantage of non-local algorithm in image denoising.(2)The sparse representation of the relevant theory is introduced,and the NCSR non-local image denoising algorithm based on sparse representation is analyzed emphatically.On the basis of clear shortcomings that the brightness and the structure information of the image without considering the Euclidean distance in the NCSR algorithm,improved the distance definition of the self-similarity measure,that is,by introducing the similarity of the structure(SSIM)takes into account the structure and brightness of the image.The related experiments show that the improved algorithm not only improves the PSNR and SSIM,but also improves the visual effect.(3)Introduce the low rank correlation theory,and analyzes emphatically the SAIST non-local image denoising algorithm based on low rank and improve it.a)Considering the noise interference caused by the interference of similar blocks,the DCT pre-filtering method is introduced,which effectively solves the interference of noise at the time of polymerization.b)As the singular value threshold is affected due to the sample mean is not considered for the correlation of the imageand,propose the weighted average of the samples with the correlation between the image blocks is taken into account,which makes the singular value threshold be more adaptable.The related experiments show that the noise image based on the improved algorithm has a certain improvement in objective and subjective evaluation.更多还原