Small-Object Detection Based on ITS-YOLOv5s-P2

In small object detection tasks, the instances of small objects are numerous, multi-scale, densely distributed, and easily occluded between objects, often resulting in missed and false detections, making it difficult to improve detection accuracy. A new small object detection algorithm called ITS-YOLOv5s-P2 is proposed to address these issues. Firstly, a P2 detection layer is added to YOLOv5s, introducing high-resolution feature maps and rich detail information to improve the overall recognition performance of the model. Secondly, an improved adaptive anchor box strategy (IAAS) is proposed, allowing the model to learn better and fit the real bounding boxes, which is beneficial for multi-scale object recognition. Then, a tiny object detection layer (TODL) is designed, incorporating texture information from shallow regions of the backbone network into the prediction head, enhancing the model's ability to perceive details of small objects. Finally, Soft-NMS is used instead of NMS for non-maximum suppression, effectively mitigating the issue of object occlusion that leads to missed detections caused by the aggressive filtering of traditional NMS. Experiments show that the ITS-YOLOv5s-P2 algorithm effectively improves the detection accuracy of small objects. On the public dataset VisDrone2019, the parameters and the number of floating-point operations of this algorithm increase by only 0.2M and 7.3G compared to YOLOv5s, while mAP@O.5 and mAP@[.5:.95] increase by 11.6% and 9.1%.