Abstract
In order to enhance the fine-grained recognition of fish species, this paper proposes a lightweight object detection model YOLOv8n-DFG. The model accurately identifies six deep-sea fish species including Flatfin sailfish, Striped marlin, Atlantic bluefin tuna, Bigeye tuna, Longtail tuna, and Yellowfin tuna, meeting the requirements for real-time detection and fishing monitoring. Firstly, by introducing FasterNetBlock and EMA attention mechanism into the YOLOv8 network structure to improve C2f and obtain the C2f-FE module, this model enhances feature extraction accuracy and operational efficiency. Subsequently, it combines BiFPN structure with C2f-FE module to construct a fast and lightweight neck network structure that achieves multi-scale feature fusion. Additionally, Dysample dynamic upsampling module is introduced along with porting of Adown downsampling module from YOLOv9 to optimize feature pyramid sampling method named as YOLOv8-FG. Finally using large-sized YOLOv8s-FG as teacher network and small-sized YOLOv8n-FG as student network based on CWD loss intermediate layer feature distillation method constructs the final model YOLOv8n-DFG. Experimental results on a dataset containing six morphologically similar fish species demonstrate the effectiveness of these improvements and distillation effects are significant. Compared to YOLOv8n, precision has increased by 7.8%, recall by 3.3%, mAP@50 by 5.6%, while FlOPs decreased by 42% with a reduction in model size of 58%. The results indicate that our proposed YOLOv8n-DFG demonstrates exceptional accuracy and real-time performance, effectively fulfilling the requirements for real-time fine-grained fish recognition.