A Classification Algorithm of Fish Feeding Behavior for Automatic Bait Feeding Control

Abstract

In aquaculture, automatic feeding control of fish can effectively improve production efficiency. However, most automatic feeding devices are open-loop control because of lack of information on fish feeding behavior, which could lead to a large amount of bait waste. To address this problem a novel video classification algorithm based on an inter-frame relationship Bayesian estimation network(IRBEN) is proposed in this paper, which provided prior knowledge for automatic feeding control of fish. The IRBEN first employs a VAE encoder to convert the frames of a video clip into multivariate Gaussian distributions(MGDs). Then, two fully connected networks, one is trained on the MGDs associated with the fish eating video clips and the other on the MGDs associated with the fish noneating video clips, are employed to predict the MGDs of the frame after an interval from the MGD of the current frame. The classification is conducted by finding the fully connected network achieving smaller KL distance between the predicted MGD and the actual MGD. The experimental results show that the IRBEN achieves the classification accuracy of 97.5%.