Image denoising by bounded block matching and 3D filtering

The block-matching with 3D transform domain collaborative filtering (BM3D) achieves very good performance in image denoising. However, BM3D becomes ineffective when an image is heavily contaminated by noise. This is because it allows block-matching to search out of the region where a template block is located, resulting in poor matching. To address this, this paper proposes a bounded BM3D scheme. The novelty of our bounded BM3D is two-folded. First, our scheme partitions an image into multiple regions, and identifies the boundaries between regions. And we restrict block matching search within the region of the template block. Second, to prevent important geometric features such as edges from being removed by collaborative filtering in BM3D, we do partial block matching for different block coherent segments which belong to different regions. Compared to BM3D, the proposed bounded BM3D achieves better visual performance, as well as 0.23–1.33 dB PSNR increase for heavily noisy images.