R-YOLOv5s: Improved YOLOv5s for Object Detection in Low-Light Environments

In response to the challenge of low detection accuracy exhibited by mainstream object detection models in low-light environments, this paper proposes a novel detection model named R-YOLOv5s. The model incorporates several key enhancements to address this issue. First, the SCI image enhancement algorithm is designed to preserve more target features and details. Next, a newly lightweight RepVIT backbone network is built to extract more image features; the global attention mechanism (GAM) is introduced to generate multiscale features that are more readily discernible, thereby enhancing the efficiency of feature capture. To significantly enhance the efficiency and precision of prediction box regression, a specialized loss function called SIoU loss is constructed. Results from experiments conducted on the ExDark dataset indicate notable improvements over the baseline model, with precision (P) increasing by 10%, recall (R) by 11%, and the mean average precision (mAP 0.5) by 13%. The newly devised R-YOLOv5s Model achieves higher detection accuracy in low-light environments, showcasing its effectiveness in addressing the challenges posed by such conditions.