Adversarially Regularized Low-Light Image Enhancement

William Y. Wang, Lisa Liu, Pingping Cai

Abstract: The task of low-light image enhancement aims to generate clear images from their poorly visible counterparts taken under low-light conditions. While contemporary approaches leverage deep learning algorithms to enhance low-light images, the effectiveness of many of them heavily hinges on the availability of large amounts of normal images and their low-light counterparts to facilitate the training process. Regrettably, it is very challenging to acquire a sufficient number of such paired training images with good diversity in real-world settings. To address this issue, we present a novel approach that employs an adversarial attack process to maximize the utility of the available training data, thereby improving the network’s performance while requiring far fewer images. Our key insight involves intentionally degrading the input image to create a deliberately worse version, effectively serving as an adversarial sample to the network. Moreover, we propose a novel low-light image enhancement network with specific multi-path convolution blocks, which preserve both global and localized features, resulting in better reconstruction quality. The experimental results validate that the proposed approach achieves promising low-light image enhancement quality by surpassing the performance of many previous state-of-the-art methods.