Feishe Chen (current student), Lixin Shen, Yuesheng Xu, and Xueying Zeng
Abstract: This paper presents the Moreau envelope viewpoint for the L1/TV image denoising model. The main algorithmic difficulty for the numerical treatment of the L1/TV model lies in the non-differentiability of both the fidelity and regularization terms of the model. To overcome this difficulty, we propose five modified L1/TV models by replacing one or two the non-differentiable functions in the L1/TV model with their corresponding Moreau envelopes. We prove that several existing approaches for the L1/TV model essentially solve some of the modified models, but not the original L1/TV model. Algorithms for the L1/TV model and its five variants are proposed under a unified framework based on the proximity operator. Depending upon whether we smoothen the regularization term or not, two different types of proximity algorithms are presented. The convergence rates of both types of the algorithms are improved significantly by exploring either the strategy of Gauss-Seidel iteration, or the FISTA, or the both. We compare the performance of various modified L1/TV models for the problem of impulsive noise removal, and make recommendations based on our numerical experiments for using these models in applications.
Journal: Inverse Probl. Imaging 8 (2014), no. 1, 53–77.
DOI: http://dx.doi.org/10.3934/ipi.2014.8.53