Image Denoising with a Low-Rank Preprocessing before BM3D

In this paper, we attempted to improve the denoising performance of BM3D. As one of the most efficient denoising algorithms, BM3D is the fastest method while achieving remarkable denoising performance. However, since it employs fixed and separable 3D transformation, it is not adaptive enough to represent structural image areas which results in poor performance on this areas. In addition, it is greatly influenced by noise levels. To overcome these problems, in this paper, we introduced an adaptive low-rank preprocessing and backtracking procedure before BM3D. The experimental results prove that the proposed method has a great improvement on BM3D. The proposed algorithm not only produces better denoising results than BM3D in most cases, but also adapts to a variety of noise levels.