Small Target Detection in Soccer Scenes Based on YOLOv8

A small target detection algorithm HPDR-YOLOv8 based on YOLOv8n proposes to solve the challenges of large deformation and small size in soccer video images. Firstly, the original network neck adopts the HS- PAN structure to achieve more efficient feature extraction. Secondly, a P2 small target detection layer is added to improve the detection accuracy of small targets effectively. At the same time, the lightweight module DCSPELAN is designed to alleviate the defect of increasing calculation due to the increase of network depth in the P2 layer. Finally, reparameterized convolution is introduced, ordinary convolution is replaced by RepConv convolution, and convolution is shared. Compared with the original detection head, the parameter number and calculation amount reduce. Experiments are carried out on the public data set SoccerDB and the self-built data set MD-Soccer. The experimental results show that compared with YOLOv8n, the HPDR-YOLOv8 algorithm improves the indices of mAP 50 and mAP50:95 by 2% and 1.4%,3.7% and 2.8%, respectively, enhancing the detection effect of the model.