Edge-Deployable Fish Feeding-State Quantification and Recognition via Frame-Pair Motion Encoding and...

Area of Science:

• Aquaculture technology
• Computer vision
• Animal behavior analysis

Background:

• Accurate feeding-state monitoring is crucial for efficient aquaculture management, reducing waste, and ensuring fish welfare.
• Current vision-based methods face limitations due to subjective labeling and high computational costs, hindering practical application.

Purpose of the Study:

• To develop an objective, edge-deployable framework for real-time feeding-state quantification and recognition in aquaculture.
• To enable timely, data-driven feeding decisions through automated motion analysis.

Main Methods:

• Integration of frame-pair dense optical-flow encoding with a lightweight neural network (EfficientFeedingNet).
• Utilizing an optical-flow-derived motion-intensity signal (V-Value) for automatic delineation of feeding intervals.
• Construction of a perception-based dataset (Perceptual Dataset) with reproducible labels.

Main Results:

• Models trained on the Perceptual Dataset achieved over 90% test accuracy, outperforming those trained on observer-labeled data.
• The EfficientFeedingNet achieved 96.53% test accuracy and operated at 143.24 fps on edge hardware (Jetson Orin NX).
• The framework demonstrated a practical basis for real-time, motion-driven feeding-state quantification.

Conclusions:

• The proposed framework offers a practical solution for objective, real-time feeding-state monitoring in aquaculture.
• EfficientFeedingNet's lightweight design facilitates edge deployment, supporting precision aquaculture practices.
• This technology can significantly improve feeding management, reduce feed waste, and enhance fish welfare.