Pelagic fishery is still faced with the problems of low automation and intelligence level. As the traditional method of catch statistics is counting by people, which wastes time and energy, automatic catch statistics becomes one of hot spots in the study of pelagic fisheries. To solve this problem in the pelagic tuna fishery, this paper presents an automatic counting method based on DP-YOLO (DCNv2-PConv-YOLO) model combined with dynamic detection gate algorithm. Due to the lightweight nature of YOLOv7-tiny, it is treated as a baseline model. Using deformable convolution DCNv2 can obtain more shape features. To alleviate the speed decrease caused by adopting DCNv2 and ensure that it can work on edge devices, partial convolution PConv is used to replace regular convolution to improve detection speed, reduce model computation and performance demands of the hardware. Dynamic detection gate algorithm is designed to avoid repeated counting. Meanwhile, a new miscounting error index called ECE (Error Counting Error) is proposed to evaluate the counting method. To verify the influence of the module on YOLOv7-tiny, ablation test results show that DP-YOLO not only reduces 3.3% of parameters, 23.7% of calculation time and 2.1% of calculation time, but also increases the average accuracy by 5.3%. The test results of automatic catch statistics show that the identification accuracy of this method is 95.8%, the counting accuracy achieves 97.9%, and ECE is only 2.1%, which is respectively 45.8% and 25% higher than the existing counting algorithms of YOLOv5s+Deepsort and YOLOv7-tiny+Deepsort. This algorithm provides a novel method of automatic counting catch for pelagic fishery. Therefore, the study can meet the catch statistics requirements of tuna longline fishery.
