Lightweight marine biodetection model based on improved YOLOv10

Abstract

In IoT-enabled marine biology, real-time monitoring of marine organisms faces challenges due to blurred images and complex underwater backgrounds, which hinder feature extraction and lead to missed detections. Addressing these issues, the lightweight YOLOv10-AD model introduces AKVanillaNet, a novel backbone optimized for the distinct shapes of marine organisms, improving detection accuracy while minimizing parameters and computational cost. Additionally, the model incorporates the DysnakeConv module within the C2f structure to enhance feature extraction, along with the Powerful-IOU (PIOU) loss function for better data fitting. Testing on URPC dataset shows that YOLOv10-AD achieves an mAP of 85.7%, with a parameter count of 2.45 M, 6.2 GFLOPs, a model size of 5.0 M, and a frame rate of 156 FPS. Compared to the baseline, YOLOv10-AD improves mAP by 5.7% and FPS by 25.8%, while reducing parameters, computational load, and model size by 9.3%, 24.4%, and 9.1%, respectively. This IoT-compatible model enables precise, real-time classification of marine organisms across various lighting conditions, making it a valuable framework for intelligent grading applications in marine biology and advancing IoT-based environmental monitoring.