Object detection in low-light conditions based on DBS-YOLOv8

Object detection algorithms struggle with several challenges in low-light conditions, including blurry and dim targets, unclear imaging, and significant loss of detail. These challenges often result in false positives, missed detections, and inaccurate localization. This paper introduces a novel method, DBS-YOLOv8, for enhancing object detection in such environments. First, lightweight deformable convolutions are introduced into the backbone feature extraction network to predict sampling offsets and modulation scales, enhancing the network’s feature extraction capabilities under complex low-light backgrounds. Then, the spatial pyramid pooling structure of the backbone network is improved to retain feature information while increasing the receptive field, thus improving the model’s computational efficiency and accuracy. Furthermore, the BiFormerBlock module is added to the neck feature fusion network, enhancing the detail processing capability for low-light images. This adjustment allows the network to better focus on regions of interest, reducing the probability of missed or misjudged targets due to insufficient light. Finally, the bounding box regression loss function is optimized to improve the average accuracy of detecting multiple overlapping objects. Experimental results show that the proposed algorithm improves mAP50 by 3.7%, mAP50-95 by 6.7%, precision by 2.4%, and recall by 4.1% compared to the YOLOv8s algorithm.