Physics prior-based contrastive learning for low-light image enhancement

Abstract

Capturing images in low-light conditions can lead to losing image content, making low-light image enhancement a practically challenging task. Various deep-learning methods have been proposed to address this challenge, demonstrating significant progress. However, existing methods still face challenges in achieving uniform brightness enhancement. These methods rely solely on normal-light images to guide the training of the enhancement network, resulting in insufficient utilization of low-light image information. We propose a novel Illumination Contrastive Learning (ICL) that employs positive and negative samples to improve contrast relationships and combines local brightness data to align luminance images closer to normal light and away from low-light areas. Existing methods that use channel attention mechanisms often neglect global channel dependencies, leading to poor color contrast in enhanced images. We address this issue by developing a Multi-scale Channel Dependency Representation Block (MCRB) that utilizes multi-scale attention to capture a wide range of channel dependencies, thereby enhancing contrast more effectively. Based on the Retinex theory, our method maximizes the use of illumination information in low-light images and integrates contrast learning into a Retinex-based framework. This integration results in a more uniform brightness distribution and improved visual effects in enhanced images. The effectiveness of our method has been validated through tests on various synthetic and natural datasets, surpassing existing state-of-the-art methods.