ASNet : Attention-guided structure-aware network for low-light image enhancement

Low-light image enhancement (LLIE) aims to optimize the visibility and perceptual quality of images described under poor lighting conditions by increasing brightness and restoring fine details. This letter introduces a LLIE method called ASNet, which combines attention guidance with structure-aware mechanisms to perform brightness adjustment. Specifically, we design a Pixel Adjustment Module (PAM), whose core is a Channel-Guided Attention Module (CGAM). It leverages the global contextual information of the input feature maps to guide the network’s attention, enhance key semantic channels, and improve detail preservation in LLIE tasks. CGAM is embedded into the feature fusion stage, significantly enhancing the network’s ability to perceive illumination distribution and respond to structural edges. Additionally, we design a structure-aware illumination adjustment loss to encourage the network to learn natural and structurally consistent illumination mappings. Extensive experiments on four benchmark datasets confirm that ASNet achieves superior performance compared to existing advanced methods in both visual and numerical assessments.