The quantification of fish feeding activity is a prerequisite for precise feeding. Deep learning has achieved notable achievements in the analysis of fish feeding activity. However, existing deep learning methods for analysing fish feeding activity still have some limitations: 1) they are all strongly supervised methods, and require labels to guide learning; 2) they can only qualitatively evaluate the feeding stage, while continuous quantification is not attained. To address these issues, an appearance-motion autoencoder network (AMA-Net) for quantifying fish feeding activity is proposed. Specifically, to avoid the tedious and time-consuming manual labelling, semi-supervised learning is adopted in the network. In the learning phase, only non-feeding samples that do not require labelling are used as training data. During the application phase, the continuous quantification of fish feeding activity is realised through the hypothesis of reconstruction error. Moreover, for higher accuracy and stability in fish feeding activity measurement, the proposed network performs modelling from both appearance and motion with the dual-stream framework to accurately quantify fish feeding activity. The experiments qualitatively and quantitatively analysed the performance of AMA-Net, which confirmed the effectiveness, accuracy and stability of AMA-Net in quantifying fish feeding activity. Further, to realise feed fish on demand, an intelligent feeding decision system is proposed, which can automatically determine whether the fish has completed feeding. The experimental results showed that the feeding decision system accurately and effectively decided the endpoints of the feeding process and provided theoretical support for production practices in aquaculture.
Read full abstract