Compression noise estimation and reduction via patch clustering

Images compressed at low bit rates usually suffer from annoying artifacts due to coarse quantization of transform coefficients. In this paper, we propose a soft-thresholding scheme to reduce compression noise with content-based noise level estimation. In the proposed method, a compressed image is divided into multiple similar image patch groups, and the compression noise is estimated from every group respectively based on coefficient distribution in transform domain. For each group of similar patches, soft-thresholding is applied to the singular values in the singular value decomposition (SVD) of every group of similar patches. The threshold is adaptively determined based on the standard deviation of image signals and compression noise. Finally, quantization constraint is applied to estimated images to avoid over-smoothing. Extensive experimental results show that the proposed method improves the quality of compressed images obviously, and outperforms state-of-the-art denoising algorithms significantly.