Deep Learning-Based Fish Detection Using Above-Water Infrared Camera for Deep-Sea Aquaculture: A Comparison Study

Area of Science:

• Marine Biology
• Computer Vision
• Aquaculture Technology

Background:

• Deep-sea aquaculture requires reliable fish monitoring for safety and efficiency.
• Automated fish detection systems are essential for long-term data collection in marine environments.

Purpose of the Study:

• To evaluate deep learning models for automated fish detection in deep-sea aquaculture.
• To compare the performance of different backbone networks and improvement modules within the Faster R-CNN framework.
• To investigate the impact of learning rates, feature extraction layers, and data augmentation on detection accuracy.

Main Methods:

• An infrared camera was used on a deep-sea net cage to collect fish images.
• A labeled fish dataset was created for training and testing object detection models.
• Faster R-CNN was employed as the base object detection framework, with variations in backbone networks (e.g., EfficientNetB0, VGG16) and feature pyramid network (FPN) modules.
• Experiments included varying learning rates, feature extraction layers, and data augmentation strategies.

Main Results:

• Faster R-CNN with EfficientNetB0 backbone and FPN module achieved a competitive AP50 of 0.85 with significantly shorter detection times.
• The highest AP50 of 0.86 was obtained using VGG16 with all improvement modules and data augmentation.
• The study confirmed the effectiveness of deep learning for fish detection in aquaculture settings.

Conclusions:

• Deep learning-based object detection methods are effective for automated fish monitoring in deep-sea aquaculture.
• The EfficientNetB0 backbone with FPN offers a strong balance between speed and accuracy for fish detection.
• Further network improvements and data augmentation strategies can enhance detection performance.