One stage anchor-based Semi-Supervised Learning for Fish Disease Detection and Classification

To overcome the prohibitive annotation cost of monitoring diseases in intensive aquaculture, we propose SSFDD, a semi-supervised one-stage detector for fish disease. Using only 10% labeled data, it achieves 81% AP50 and 32% AP50:95 at 3.2 ms inference speed. This is enabled by three key components: (1) A dense anchor sampling strategy, combining RetinaNet's architecture with YOLOv5's anchor design, which improves detection of small lesions; (2) A confidence-guided pseudo-label filter that reduces teacher-student bias through dual thresholds; (3) An epoch adaptation mechanism that automatically adjusts training length based on label scarcity, stabilizing convergence without manual tuning. Experiments show SSFDD achieves competitive performance versus YOLOv5/v8/v11 baselines, which require fully labeled data. This approach enhances robustness while reducing annotation costs, offering a practical solution for real-world fish disease detection.