Small object detection algorithm based on improved YOLOv7 multi-layer attention fusion

Small object detection, as a key and difficult problem in computer vision, has the characteristics of small scale, low contrast, and severe occlusion. To solve the existing difficulties in small object detection, a small object detection algorithm based on multi-layer attention fusion mechanism is proposed. Attention weights are extracted and fused in different depth feature layers, and the fused attention weights are used as reference to guide the detection head for detection; Changing the detection layer of YOLOv7 network to a detection head for small objects to enhance the network's extraction of rich detailed information from small objects and improve the phenomenon of missed and false detections of small objects; In response to the limitations of IOU loss in the field of small object detection, the NWD loss function is introduced to replace the IOU loss in YOLOv7, which improves the detection accuracy of the model. We constructed a dataset of over 20000 images, including over 70000 instances of birds and small objects under natural light. The experimental results on this dataset showed that compared with the benchmark model YOLOv7, the algorithm's MAP0.5 improved by 21.7 percentage points. At the same time, the experimental results on the public dataset AI-TOD showed that the algorithm's MAP0.5 improved by 8.1 percentage points.