Modified BM3D algorithm for image denoising using nonlocal centralization prior

• DOI: 10.1016/j.sigpro.2014.08.014
• Publication Date: 2014-08-21T09:35:58Z
• Authors: Hua Zhong; Ke Ma; Yang Zhou
• Keywords: Shrinkage; Similarity (geometry)

Block matching and 3D collaborative filtering (BM3D) has shown great power for image denoising. For grouped image blocks, this letter proposes to remove the 1D transform inter-blocks and introduce the nonlocal centralization prior to better utilize both local sparsity of wavelet coefficients and nonlocal similarity of grouped blocks. Three nonlocal shrinkage functions are developed under different norm restrictions of wavelet coefficients intra- and inter-blocks. Such shrinkage functions are verified to be competitive or better than the BM3D algorithm and other state-of-the-art denoising methods. We remove the 1D transform across grouped blocks and introduce the nonlocal centralization prior.Both local sparsity of wavelet coefficients and nonlocal similarity can be better utilized.Nonlocal shrinkage functions are developed under different norm restrictions.Such shrinkage functions are verified to be competitive or better than the BM3D algorithm.