Research on tuna purse seine fishing target detection and recognition method for Unmanned Aerial Vehicle aerial images based on knowledge distillation

Abstract

Abstract Tuna purse seine fleets face high costs in locating fish aggregations, whether free-swimming or around fish aggregating devices (FADs), therefore efficient detection is crucial. While traditional methods relied on visual cues like bird activity, modern technologies such as helicopters and radar have improved search capacity. However, due to high costs and safety concerns, the industry is shifting toward unmanned aerial vehicles (UAVs). This study proposes an optimized YOLOv6-based deep learning model for real-time marine object detection, using model compression and knowledge distillation (KD). Four YOLOv6 variants (YOLOv6l/m/s/t) were evaluated, with YOLOv6t-student/distillation (Model 13) achieving the best balance—0.722 mAP0.5 at just 24.89 GFLOPs. Field tests in the Pacific Ocean showed 94.81% FAD detection accuracy, including perfect detection (19/19) in defocused conditions. KD analysis found that isomorphic teacher model (YOLOv6t) performed best, while complex model (YOLOv6l) showed reduced mAP0.5 (0.589) but achieved higher recall (0.800). This framework offers a cost-effective UAV alternative to helicopters and supports future integration of additional sensors (e.g. radar, sonar) to achieve more intelligent and efficient fish detection capabilities.