Accurate localization of fish heads and tails in underwater images using a cascaded DWDA DETR approach

Underwater fish detection and part localization are crucial for fisheries management, conservation, and aquaculture. However, accurately identifying fish heads and tails in complex aquatic environments remains challenging. This study proposes a novel cascaded object detection approach that combines YOLOv5s with the Deformable Window Deformable Attention Detection Transformer (DWDA DETR) model to enhance the accuracy of fish head and tail detection in underwater environments. The method first detects fish targets using YOLOv5s and then precisely localizes their heads and tails with DWDA DETR. By incorporating Deformable Window Attention and Variable Window Downsampling, our model effectively handles images of varying fish sizes. Experimental results on an underwater fish dataset demonstrate that our approach achieves a mean Average Precision (mAP@50) of 94.23%, significantly outperforming existing models. The lightweight DWDA DETR model (6.6M parameters) also demonstrates robust performance in challenging scenarios, such as fish aggregations and occlusions. Our work has implications for improved fish counting, tracking, behavior analysis, and disease detection in aquatic ecosystems.