FDMNet: A Multi-Task Network for Joint Detection and Segmentation of Three Fish Diseases.

Area of Science:

• Aquaculture
• Computer Vision
• Deep Learning

Background:

• Fish diseases cause significant economic losses in aquaculture.
• Current deep learning models often have limitations in detecting single fish disease types or performing single tasks.

Purpose of the Study:

• To develop an advanced deep learning network, FDMNet, capable of simultaneous fish disease detection and lesion segmentation.
• To address the limitations of existing single-task models in aquaculture.

Main Methods:

• FDMNet is a multi-task learning network built on the YOLOv8 framework, incorporating a semantic segmentation branch with multi-scale perception.
• Utilizes the C2DF dynamic feature fusion module to prevent information loss during feature fusion across scales.
• Employs uncertainty-based loss weighting and PCGrad to manage conflicting gradients between detection and segmentation tasks.

Main Results:

• FDMNet achieved 97.0% mAP50 for detection and 85.7% mIoU for segmentation on a custom dataset of three fish diseases.
• Demonstrated a 2.5% improvement in detection mAP50 and a 5.4% improvement in segmentation mIoU compared to the YOLO-FD baseline.
• Showcased competitive accuracy in both detection and segmentation tasks.

Conclusions:

• FDMNet effectively performs simultaneous fish disease detection and segmentation.
• The proposed methods enhance model stability and performance, offering practical utility for aquaculture disease management.