Abstract: The images acquired in complex low-light environments are prone to problems such as low contrast and loss of detail information, and the results of using existing target detection methods are not satisfactory. In this paper, the YOLOv8 algorithm is improved for the special characteristics of low-light environment detection to improve the reliability of detection in low-light environment. Firstly, the backbone feature extraction part of this network adopts C3_ResBlock to improve the multi-scale and weak feature extraction ability in low-light detection; subsequently, the hollow space pyramid structure SE_ASPP is introduced to extract the information of the complex scene by using different expansion rates to maintain the computational volume while improving the training effect; finally, the adaptive fusion of SKNet and GAM attention mechanism. SKNet contains adaptive induction field mechanism, which can select more important and effective space for multi-scale feature extraction and fusion, and GAM can adjust the importance degree of each channel to improve the feature extraction and characterization ability of the network model. Numerical experiments show that compared with YOLOv8, the proposed YOLO-RSG algorithm in this paper improves the mAP by 3.60% in the ExDark dataset, which can effectively improve the performance of target detection in low illumination scenes and has good stability and applicability, and can better meet the needs of target detection in low-light scenes.

Key words: Object detection, Low-light scenes, Multi-scale features, Attention mechanism