Low-Light Environment Object Detection Based on YOLO

In recent years, object detection has advanced significantly, yet low-light detection remains challenging. Enhancers are commonly used for preprocessing, but they often amplify noise, reducing detection accuracy. To address this, we propose VD-YOLOv5, which mitigates noise interference while leveraging enhancer-detector synergy. A novel Channel Denoising Module using side window filtering removes noise from enhanced images. We introduce an Extremum-Sensitive Spatial Pyramid Pooling to handle noise-induced extremes and design a Gradient-Attention Module based on gradient pathway planning for better low-level feature extraction. Experiments show that VD-YOLOv5 achieves a mAP0.5 of 70.7% on ExDark, outperforming YOLOv5 by 6.2%.