<p>The significant impact of low-light environments on target detection performance has become a key issue in computer vision. Limited by factors such as image feature degradation, noise interference and contrast reduction, conventional detectors’ performance degrades dramatically under low-light conditions. Existing solutions usually decouple image enhancement and target detection as independent tasks, yet this paradigm struggles to achieve stable detection in complex low-light environments. Therefore, this paper proposes a feature-aware detection framework integrating multi-scale edge enhancement and light-guided attention mechanisms, built on YOLOv8 with three key modules: the multi-scale edge enhancement module (MEEM) for enhancing edge features in low-contrast scenes, the light-guided attention module (IGAB) for achieving regionally adaptive feature modulation, and the partial convolution module (PConv) for optimizing computational efficiency. Unlike traditional methods using image enhancement as preprocessing, the proposed framework integrates enhancement directly into feature extraction, enabling adaptive learning of low-light scene representations. Experimental results show our method achieves 49.7% mean average precision (mAP) on the ExDark dataset, exceeding YOLOv8 baseline by 6.0%. Additionally, the proposed method demonstrates superior detection robustness across complex low-light scenes at 170.3 FPS, providing efficient and reliable support for visual perception systems in real low-light environments.</p>