Research on image denoising using patch-based singular value decomposition

This paper presents an efficient patch-based image denoising scheme by using singular value decomposition (SVD). In this scheme, similar image patches are simply grouped together from a noisy image. For a better sparse representation of these similar patches, we firstly utilize the 2D SVD to reveal the essential features of each individual patch and then the 1D SVD to exploit the correlation between similar patches. By doing so, the image features can be well preserved when attenuating noise by the shrinkage of transform coefficients. In order to improve denoising performance, the proposed scheme is employed once again. Besides that the similar patch grouping and the basis calculation of 2D SVD are performed from the first-stage estimated image, a fixed orthogonal transform instead of 1D SVD is applied to remove the redundancy shared by similar patches. Experimental results show that the proposed two-stage denoising scheme achieves more competitive performance than the state-of-the-art denoising algorithms.