摘要
现有基于组稀疏表示的图像恢复方法大多利用非局部自相似先验特性,相似的小块聚类成组,对每组系数施加稀疏度,从而有效保留图像纹理信息。然而,这些方法只对组中每个单独的块施加了简单稀疏性,而忽略了其他有益的图像属性。基于此,提出了一种基于低秩模型和残差模型的图像去噪算法,不仅利用了每组相似块的稀疏性和低秩性,还利用残差学习方法来自动估计图像块的真实稀疏表示。实验结果表明所提算法充分考虑了块之间的关系,将块的相关性和特异性结合,有效实现了图像去噪,从而得到高质量的恢复图像。同时,实验还表明所提算法的峰值信噪比平均增益比块匹配三维协同滤波(Block-Matching and 3D Filtering,BM3D)算法提高0.34 dB,比非局部集中稀疏表示(Non-Local Centralized Sparse Representation,NCSR)提高0.48 dB,比低秩正则联合稀疏(Low-Rank Regularized Joint Sparsity,LRJS)提高0.2 dB,比低秩引导的组稀疏表示(Low-Rankness Guided Group Sparse Representation,LGSR)和GSR_SRLR提高0.04 dB,且平均结构相似性值达到次高,足以证明其优于许多流行或先进的去噪算法。
Most existing group sparse representations based image restoration methods utilize the non-local self-similarity prior property to cluster similar small blocks into groups and apply sparsity to each group of coefficients,which effectively preserves image texture information.However,these methods only apply simple sparsity to each individual block in the group,but ignore other beneficial image attributes.Based on this,an image-denoising algorithm based on a low-rank model and a residual model is proposed.It not only utilizes the sparsity and low rank of each group of similar blocks,but also uses residual learning methods to automatically estimate the true sparse representation of image blocks.The experimental results show that the proposed algorithm fully considers the relationship between blocks,com‐bines the correlation and specificity of blocks and then effectively performs image denoising to obtain high-quality restored images.The experimental results also show that the PSNR average gain of the proposed algorithm was 0.34 dB higher than BM3D,0.48 dB higher than NCSR,0.2 dB higher than LRJS,0.04 dB higher than LGSR and GSR-SRLR,and the average SSIM value reached the second highest,which is sufficient to prove that it is superior to many popular or state-of-the-art denoising algorithms.
作者
杨雅兰
胡红萍
杨正民
YANG Yalan;HU Hongping;YANG Zhengmin(School of Mathematics,North University of China,Taiyuan 030051,China;School of Information and Innovation Industry,Shanxi University of Electronic Science and Technology,Linfen 041099,China)
出处
《测试技术学报》
2025年第5期548-557,共10页
Journal of Test and Measurement Technology
基金
山西省基础研究计划资助项目(20210302123019,202103021224195,202103021224212,202103021223189)
山西省回国留学人员科研项目(2021-108)。
关键词
图像去噪
稀疏表示
非局部自相似
交替最小化
image denoising
sparse representation
nonlocal self-similarity
alternating minimization
