Marine-YOLO: improved YOLO for marine organism detection

Underwater marine organism detection is crucial in marine resource monitoring and ecological research. However, it faces challenges such as low image quality, scale differences, and uneven sample distribution. To address the above issues, we proposed an improved marine organism detection model, Marine-YOLO in this paper. Firstly, a Bi-Level Routing Attention (BRA) mechanism is introduced to dynamically select key feature regions and enhance context awareness. Secondly, a multi-scale feature enhancement module (MFE) is designed, which combines multi branch convolution to adaptively capture detailed features of objects at different distances. Finally, a composite loss function WIoU- NWD that integrates Wise-IoU and Normalized Wasserstein Distance (NWD) is proposed to balance the regression weights of different scales with dynamic gradient allocation strategy. Comparative experiments were conducted on an underwater marine organism detection dataset with 6 different object detection models. The experimental results showed that the proposed Marine YOLO achieved the highest mAP@50 of 84.2% and is 3.7% higher than the baseline method YOLOv8-s.