Learning-based low-rank denoising

Abstract The denoising of 2D images through low-rank methods is a relevant topic in digital image processing. This paper proposes novel method that trains learning network to predict the optimal thresholds singular value decomposition involved images. To improve results, we apply block-matching algorithm and classify each 3D block according four parameters, which increase specificity for prediction thresholds. Our outperforms state-of-the-art denoising; furthermore, it general with respect type noise provides an upper bound accuracy Singular Value Decomposition.