Low-Light Image Enhancement via Implicit Priors Regularized Illumination Optimization

Abstract

Low-light image enhancement is a very challenging problem due to insufficient or uneven illumination, complicated noise and low contrast. Retinex-based methods have shown to be effective in separating the illumination from the reflectance with well-designed priors. However, the commonly used hand-crafted priors may not model the piecewise smoothness of the illumination. In this article, we propose a Retinex-based variational framework, which imposes an implicit prior on the illumination component. By formulating decomposition problems as an implicit prior regularized model, the regularized illumination term can be inferred by an adaptable mapping instead of using hand-crafted priors, which makes our model extremely versatile. In addition, an adaptive regularizer and the sparsity-enforcing regularization term are carefully designed, responsible for artifact-alleviation and noise suppression in realistic enhanced results. In order to accomplish an efficient numerical implementation, we propose a plug-and-play inspired algorithm to alternatively update the sought image and the illumination. Experimental results on four public datasets show the effectiveness of our method, which significantly outperforms the state-of-the-art methods in terms of visual quality and quantitative comparisons.