Improved YOLOv4 Network for Small Target Detection

In the process of target detection of UAV power inspection, there are some challenges and difficulties owing to the multi-scale targets and object occlusion. The improved YOLOv4 for small target detection is proposed in this article, so as to raise the accuracy of detection results. This paper designs a four-scale feature extraction network, a $$104 \times 104$$ feature extraction channel is added on the original YOLOv4 model. For the purpose of further improving the computational efficiency and the ability of feature information extraction, the SPP network is added to the three-layer features of $$26 \times 26$$ , $$52 \times 52$$ and $$104 \times 104$$ . K-means++ and transfer learning are introduced to improve the training speed and detection accuracy. The experimental result demonstrates that by comparing with SSD and the original YOLOv4 algorithms, the algorithm pointed out in this paper considerably increases the detection accuracy, especially for small target detection.