Identification of Fish Hunger Degree with Deformable Attention Transformer

Abstract

Feeding is a critical process in aquaculture, as it has a direct impact on the quantity and quality of fish. With advances in convolutional neural network (CNN) and vision transformer (ViT), intelligent feeding has been widely adopted in aquaculture, as the real-time monitoring of fish behavior can lead to better feeding decisions. However, existing models still have the problem of insufficient accuracy in the fish behavior-recognition task. In this study, the largemouth bass (Micropterus salmoides) was selected as the research subject, and three categories (weakly, moderately, and strongly hungry) were defined. We applied the deformable attention to the vision transformer (DeformAtt-ViT) to identify the fish hunger degree. The deformable attention module was extremely powerful in feature extraction because it improved the fixed geometric structure of the receptive fields with data-dependent sparse attention, thereby guiding the model to focus on more important regions. In the experiment, the proposed DeformAtt-ViT was compared with the state-of-the-art transformers. Among them, DeformAtt-ViT achieved optimal performance in terms of accuracy, F1-score, recall, and precision at 95.50%, 94.13%, 95.87%, and 92.45%, respectively. Moreover, a comparative evaluation between DeformAtt-ViT and CNNs was conducted, and DeformAtt-ViT still dominated the others. We further visualized the important pixels that contributed the most to the classification result, enabling the interpretability of the model. As a prerequisite for determining the feed time, the proposed DeformAtt-ViT could identify the aggregation level of the fish and then trigger the feeding machine to be turned on. Also, the feeding machine will stop working when the aggregation disappears. Conclusively, this study was of great significance, as it explored the field of intelligent feeding in aquaculture, enabling precise feeding at a proper time.