An improved YOLOv8 model for fish classification and disease detection

Abstract

Fish classification and disease detection are crucial for sustainable aquaculture, necessitating accurate and efficient vision models. This study introduces FISH-YOLOV8, an enhanced YOLOv8 variant, incorporating: (1) SPD-Conv for optimized feature extraction and reduced computational load; (2) BiFormer Attention for enhanced small object detection and occlusion management; (3) dynamic IoU-threshold NMS to minimize false positives. This Article states that, evaluated on 15,162 images, FISH-YOLOV8 attains a mAP@50 of 0.990 and a mAP@50:95 of 0.859, outperforming baseline YOLOv8 and advanced models such as YOLOv11, at 45 fps, supports effective real-time aquaculture monitoring.