Low-Light Image Enhancement via Unsupervised Illumination Estimation

Abstract

Retinex theory-based methods have been popularly studied in the field of low-light image enhancement these years. Supported by Retinex theory, a low-light image can be decomposed into an illumination layer and a reflectance layer, which are then utilized to produce enhancement results. Traditional Retinex-based methods typically perform an optimization to achieve the decomposition, where handcraft priors are incorporated to guarantee fine textures and structures. However, they are limited by the possible low generalization of the priors and the long runtime. To solve the above problems, this paper implements the Retinex decomposition as a lightweight convolutional neural network, which takes the low-light images as inputs and is responsible for estimating their illumination components. The reflectance components can be easily obtained and considered as the enhanced results. The model is trained in an unsupervised manner, thus getting rid of the dependence on paired training data. Experiments on a series of benchmarks demonstrate the effectiveness and superiority of our method when compared to some state-of-the-art methods.