Structure preserving image denoising based on low-rank reconstruction and gradient histograms

Abstract One of the main challenges of denoising approaches is preserving images details, like textures and edges, while suppressing noise. The preservation of such details is essential to ensure good quality, especially in high-resolution images. This paper presents a novel denoising method that combines a low-rank regularization of similar non-local patches with a texture preserving prior based on the histogram of gradients. A dynamic thresholding operator, deriving from the weighted nuclear norm, is also used to reconstruct groups of similar patches more accurately, by applying less shrinkage to the larger singular values. Moreover, an efficient iterative approach based on the ADMM algorithm is proposed to compute the denoised image, under low-rank and histogram preservation constraints. Experiments on two benchmark datasets of high-resolution images show that the proposed method to outperform state-of-the-art approaches, for all noise levels.