YOLO-PRA: a multi-scale feature fusion and adaptive optimization network for underwater seahorse health detection

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Curled-tail posture is an important visual cue for identifying seahorses affected by bacterial enteritis from underwater images in seahorse aquaculture. Therefore, accurate recognition of tail posture is essential for underwater seahorse health detection. However, research on this task is limited by the lack of task-specific datasets, while accurate detection is further challenged by degraded image details in underwater environments, the small and slender tail region, and class imbalance between healthy and diseased samples. To address both the dataset limitation and detection challenges, this study constructed an underwater seahorse health detection dataset and proposed YOLO-PRA, a multi-scale feature fusion and adaptive optimization network. Specifically, the PWC3k2 module was introduced to strengthen local detail extraction, the RepLWGANet module was integrated to improve multi-scale feature fusion and target-to-background feature discrimination under degraded underwater imaging conditions, and Adaptive Threshold Focal Loss, named ATFLoss, was adopted to improve hard-sample learning under category imbalance. Experiments comparing different methods and ablation studies were performed using the self-constructed dataset. The results showed that YOLO-PRA achieved 80.0% Precision, 77.2% Recall, an mAP@0.5 of 83.3%, and an mAP@0.5:0.95 of 66.8%. Evaluations on real underwater imaging-condition subsets showed that YOLO-PRA outperformed YOLOv11n under illumination variation and blur-induced detail loss. Noise-augmented experiments further indicated relatively stable detection performance under synthetic noise, blur, and exposure disturbances. This study provides a task-specific dataset and an engineering-oriented detection algorithm for underwater seahorse health assessment.