부산대학교 도서관

Small object detection is an indispensable and challenging part of object detection. This paper proposes a small object detection method based on YOLOvS improved model. By adding shallow feature extraction networks in FPN layer and PAN layer, feature fusion was carried out with the first C3 layer to extract more details of small objects. The detection output part was extracted in the new fusion layer, and the detection output part of 32 times of the original network was deleted. Up-sampling is performed directly behind the SPPF layer, and the detail features are amplified and fused with the features of the previous layer. The experimental results show that the mAP@0.5 value and mAP@0.5:0.95 value of the improved model for small object detection reach 0.39 and 0.22, respectively, which are 6 and 5 percentage points higher than the original YOLOv5 algorithm. Recall rate increased from 0.34 to 0.39; The accuracy rate increased from 0.43 to 0.48; Classification loss, location loss and confidence loss decreased by 1 percentage point, 4 percentage point and 9 percentage point respectively. The improved model has higher accuracy and faster speed in small object detection.