Low-light image enhancement and denoising via dual-constrained Retinex model

Abstract

For low-light image enhancement, Retinex-based models have been acknowledged as a representative technique, yet they frequently amplify noise hidden in dark image. There have been some attempts to noise reduction. Because they assume a uniform level of noise, the background seems blurry. In this paper, a dual constraint is developed to perform Retinex decomposition and de-noising. First, a patch-aware low-rank model (PALR) is proposed to reject noise. In contrast to the global noise intensity assumption, PALR can measure noise intensity for each image patch, and flexibly control the regularization extent to achieve a trade-off between noise removal and details preservation. Second, noise looks like image background texture instead of image structure. Relative total variation (RTV) is used to simulate the visual difference and works well in enlarging image gradients for highlighting details. By imposing PALR constraint on clean scene image, enforcing visual difference constraint on the illumination and reflectance, a dual-constrained Retinex decomposition model (DCRD) is proposed, which can remove noise during decomposition. DCRD can be solved by an alternating optimization algorithm. Experiments on commonly tested low-light image datasets demonstrate the competing performance of our proposed model in comparison with the state-of-the-art methods.