Adaptive BM3D Algorithm for Image Denoising Using Coefficient of Variation

Abstract

Block matching 3D (BM3D) algorithm has shown powerful image denoising capability. This is achieved by block-matching, filtering and aggregating the three-dimensional arrays generated from noisy images. However, high computational cost, inadequate recovery of edge information, etc., limit its application. In this paper, we propose to reduce its high computational cost by an adaptive algorithm based on pre-classification using coefficient of variation. After pre-classification, we obtain two block subsets with different local structural information. In the subset with complex changes, called structural region, size-adaptive reference block matching is adopted for its blocks. In the subset with uniform variation, called flat region, the original size-fixed reference block matching procedure is applied. The adaptive algorithm can significantly reduce the traversal range of the BM3D algorithm for matching, and increase the similarity of the reference block size and the target block (the block to be processed) size if they are similar. This will lead to better removal of noise with lower computational cost. Experimental results show that computational cost of the adaptive algorithm is significantly reduced with close denoising performance to the original BM3D algorithm.