Low-light image enhancement integrated semantic aware guidance

Abstract

Simply increasing brightness is not an optimal solution for low-light image enhancement, as excessive adjustment often leads to overexposure, amplified noise, and color distortion. To address this issue, we propose a dual-branch semantic-aware guided enhancement network. One branch focuses on natural brightness adjustment, while the other conducts coarse semantic segmentation to guide region-specific enhancement. Indoor scenes are segmented into foreground and background, whereas outdoor scenes are divided into sky, foreground, and ground, with each region enhanced using tailored strategies. The backbone employs lightweight inverted residual convolutional blocks with attention mechanisms, and spatial-positional encoding is incorporated to inject absolute positional cues, thereby improving the understanding of image structures and spatial relationships. Extensive experiments on both the Outdoor-Synthetic dataset (synthesized from CamVid and Cityscapes) and the Indoor-LLRGDB_Real low-light dataset demonstrate that our method consistently surpasses state-of-the-art approaches in both qualitative and quantitative evaluations, achieving 33.448/0.977/0.141 (PSNR/SSIM/LPIPS) on Outdoor-Synthetic dataset and 17.939/0.993/0.259 on Indoor-LLRGDB_Real dataset. Furthermore, no-reference image quality assessments confirm the naturalness and realism of the enhanced results. Our code and corresponding database can be obtained at https://github.com/zhangjunchang2023/LLIE-SAG.