Small Object Detection Algorithm in Drone Aerial Images Based on Improved YOLOv5

Abstract

In response to the low accuracy of conventional object detection algorithms in small object detection during drone aerial missions, this paper proposes an improved YOLOv5 algorithm for small object detection. Firstly, we introduced a 160x160 small object detection head and removed the 20x20 large object detection head, adopting a strategy of fusing feature maps of P2, P3, and P4 scales to enhance small object detection performance. Secondly, by eliminating the P5 scale network layers in the backbone network, we strengthened the high-order semantic information of the P4 layer. Finally, we utilized adaptive spatial feature fusion technology to replace the direct concatenation method in the Neck network for merging different scale feature maps, optimizing the fusion effect of feature maps of different scales. The experimental results on the VisDrone2019 small object dataset show that, compared to the original YOLOv5, the proposed algorithm in this paper achieved significant improvements of 7.3%, 7.1%, and 9.2% in Precision, Recall, and [email protected], respectively. Furthermore, the model's parameter quantity was reduced by 40%, indicating that the improved model presented in this paper has superior performance in drone small object detection tasks and is highly practical.