DiffLight: Integrating Content and Detail for Low-Light Image Enhancement

Abstract

The Low Light Image Enhancement (LLIE) task has been a hotspot in low-level computer vision research. The camera sensor can only capture a small amount of ambient light signal in low-light condition, resulting in significant noise black pseudo artifacts in images, which not only degrade visual quality but also affect the performance of down-stream visual tasks. However, current methods often produce overly smoothed and distorted results, or introduce strong noise artifacts. Moreover, for recent UHD high-definition low-light images, due to GPU memory limitations, LLIE must be conducted in patches, leading to block artifacts. Faced with these challenges, we propose a dual-branch pipeline called DiffLight. Specifically, it consists of the Denoising Enhancement (DE) branch and the Detail Preservation (DP) branch. The DE-branch adopts a combination of DiffIR and LEDNet to reduce noise and enhance brightness, while the DP-branch utilizes a novel Light Full-Former (LFF) method, which comprises 20 Full-Attention (LFA) modules to preserve full-scale image details. To tackle block artifacts, we further introduce Progressive Patch Fusion (PPF) for patch fusion. Experimental results demonstrate that our approach is high-ranked in the CVPR2024 NTIRE Low Light Enhancement challenge and produced state-of-the (SOTA) results on other datasets.