Low-light Image Enhancement Based on Contrastive Learning with Generative Adversarial Network

The low-light image enhancement task aims to transform low-light images into normal-light ones while preserving structure and incorporating target appearance. Large domain appearance gaps challenge traditional networks. This paper proposes a novel method: first use traditional light enhancement to create transition-state images, then generate normal-light images via a generator and contrast learning. Contrast learning boosts detail/texture recovery amid structure/content focus. The method mitigates insufficient light, noise, and color bias, enabling efficient low-light to normal-light conversion.