Abstract: To address the reliance on manual labor for fish detection and recognition in commercial fishing vessel electronic monitoring systems, a lightweight real-time fish detection model called YOLOv7-MRN was proposed in which the backbone network of YOLOv7 was replaced with the MobileNetv3 backbone network to reduce computational complexity. Additionally, receptive field modules were incorporated to enhance the network’s feature extraction capabilities. The neck feature fusion network RFB to suppress irrelevant weights was redesign by introducing a normalization-based attention mechanism module NAM. The test of the HNY768 offshore fishing vessel electronic monitoring video fishery dataset revealed that YOLOv7-MRN achieved mAP@0.5 of 86.5%, with only 9.8% of the computational load compared to the original model. The inference speed of the model was improved by 121.69% on GPU and 219.09% on CPU. In comparison to other models, YOLOv7-MRN exhibited superior performance in practical fish detection, particularly in strong sunlight conditions. These findings indicate that the YOLOv7-MRN model proposed here can be deployed in electronic fishing vessel monitoring systems with reduced computational resource consumption to accomplish fish detection tasks.
梅海彬, 黄政, 袁红春. 基于YOLOv7模型改进的轻量级鱼类目标检测方法[J]. 大连海洋大学学报, 2023, 38(6): 1032-1043.
MEI Haibin, HUANG Zheng, YUAN Hongchun. A lightweight fish object detection method improved based on the YOLOv7 model. Journal of Dalian Ocean University, 2023, 38(6): 1032-1043.
2023, Vol. 38 
