Fish feeding behavior recognition via lightweight two stage network and satiety experiments

Area of Science:

• Aquaculture
• Computer Vision
• Artificial Intelligence

Background:

• Industrial aquaculture demands efficient and cost-effective feeding strategies.
• Computer vision offers a non-invasive method for monitoring fish, but current methods lack quantitative data and efficiency.
• Existing models struggle with redundant image data and high computational complexity, limiting real-time applications.

Purpose of the Study:

• To develop a quantitative approach for assessing fish feeding behavior and satiety.
• To design an efficient computer vision model for intelligent fish feeding systems.
• To address limitations of current models, including qualitative outputs, data redundancy, and complexity.

Main Methods:

• Quantification of fish feeding behaviors through satiety experiments to generate quantitative labels.
• A two-stage recognition network employing pose detection and graph convolutional networks (GCN) to model fish posture and distribution.
• Development of lightweight RepSELAN and SPPSF modules to reduce model complexity.

Main Results:

• Achieved 98.1% accuracy in satiety classification.
• Reduced model parameters by 31.4% and computational load by 26.2%.
• Maintained high performance with minimal decrease in mAP(B) and an increase in mAP(P).

Conclusions:

• The proposed method provides a novel and efficient foundation for intelligent fish feeding strategies.
• The quantitative approach and optimized model significantly outperform conventional methods in accuracy and efficiency.
• This advancement supports reduced costs and enhanced fish welfare in industrial aquaculture.