Patch-based sparse representation and low-rank approximation for imageprocessing attract much attention in recent years. The minimization of thematrix rank coupled with the Frobenius norm data fidelity can be solved by thehard thresholding filter with principle component analysis (PCA) or singularvalue decomposition (SVD). Based on this idea, we propose a patch-basedlow-rank minimization method for image denoising, which learns compactdictionaries from similar patches with PCA or SVD, and applies simple hardthresholding filters to shrink the representation coefficients. Compared torecent patch-based sparse representation methods, experiments demonstrate thatthe proposed method is not only rather rapid, but also effective for a varietyof natural images, especially for texture parts in images.
展开▼
