Fish Target Detection Method Based on EPSA-CenterNet2

At present, small target detection and target detection under complex backgrounds are still a major difficulty in the field of image target detection. However, fish image detection scenes often contain complex backgrounds such as water grass and reef, and fish form is small. In order to overcome the problem which is low accuracy of detection of small fish targets in complex backgrounds, In this paper, a Center Point Network 2 with Efficient Pyramid Split Attention (EPSA-CenterNet2) was proposed. The Network incorporated an Efficient Pyramid Split Attention Network (EPSANet) into CenterNet2 to improve small target detection accuracy in complex environments. In this paper, 149 images of the oplegnathus punctatus were used as a dataset to train EPSA-CenterNet2 and four other mainstream target detection networks. The experimental results showed that EPSA-CenterNet2 was superior to CenterNet2, YOLOv3, YOLOv5 and SSD in the average accuracy including AP and AP50, and the number of missed targets in small target images was less. Therefore, EPSA-Centernet2 can detect fish image targets in complex backgrounds more accurately.