Group Sparse Weighted Log-Sum Regularized Model for Hyperspectral Image Denoising
摘要
During signal sampling and digital imaging, hyperspectral images (HSI) are inevitably contaminated by mixed noises. Recent studies have integrated the non-local self similarity (NSS) prior into HSI denoising. Although the NSS based methods have achieved satisfactory denoising effects, they heavily rely on block matching (BM) operations to search for non-local similar patches. However, the number of non-local similar patches for each patch in HSI is limited, which will result in the estimation bias. To address these issues, this article proposes a new HSI denoising model based on the non-local weighted log-sum penalty (NWLP), which automatically assigns weights according to NSS information and adaptively adjusts the convexity of the objective function. Considering the strong correlation between sparse coefficients, our proposed method integrates spatial non-local similarity with global spectral group sparsity. We propose an efficient, closed-form solution for the weighted log-regularized shrinkage problem and subsequently solve the proposed NWLP model based on the alternating minimization framework.Extensive tests on simulated (WDC, Pavia;