Attention-Enhanced YOLOv8 for Real-Time Milkfish Disease Detection with Synthetic Data Augmentation

Abstract

Milkfish (Chanos chanos) aquaculture faces persistent challenges due to disease outbreaks that are difficult to detect at early stages using traditional manual monitoring methods. Varied underwater lighting conditions, turbidity, and fish movement further complicate visual assessment and may reduce detection reliability. This study proposes an automated disease detection framework that combines YOLOv8m object detection and the Convolutional Block Attention Module (CBAM) with Cycle GAN-based synthetic data augmentation. Cycle GAN is employed to generate additional diseased samples from unpaired image domains, addressing dataset imbalance and improving training diversity. CBAM is integrated into the YOLOv8m feature extraction pipeline to enhance spatial and channel attention while preserving real-time inference capability. Multiple experimental training configurations were evaluated, including raw-image training, traditional augmentation techniques, Cycle GAN-augmented datasets, and attention-enhanced detection models. Performance was assessed using precision, recall, mAP@0.50, and mAP@0.50–0.95 metrics. The proposed YOLOv8m with CBAM and Cycle GAN framework achieved the best overall performance, with a precision of 0.940, a recall of 0.910, mAP@0.50 of 0.945, and mAP@0.50–0.95 of 0.725. These results indicate that the combined use of GAN-based augmentation and attention mechanisms significantly improves detection performance under a realistic aquatic environment.