Research on Underwater Small Target Detection Algorithm Based on Improved YOLOv3

Underwater target detection has important research significance and value in the fields of fish exploration technology, fishery resources research, fishery prediction, aquaculture and so on. In this paper, an improved YOLOv3-ST model is proposed to improve the detection accuracy of underwater small targets. In this model, a linear scaling K-means clustering algorithm is designed to adapt to multi-scale feature map detection. At the same time, the 8-fold down-sampling scale feature map of the Darknet-53 network is removed, and the splice-conv module is used for up-sampling and feature fusion, finally the 4-fold down-sampling scale feature map is output as the third detection layer. On this basis, in order to further accurately detect targets with inconspicuous features such as waterweeds, two kinds of attention mechanisms are embedded in the YOLOv3-ST model. The experimental results show that the YOLOv3-ST model with the attentional mechanism of SEnet can effectively improve the underwater small target detection accuracy. The detection accuracy of echinus is 91.80%, and the detection accuracy of waterweeds is 22.33% higher than that of the original yolov3 model, while the average detection accuracy of all categories is increased by 12.13%.