Abstract
With the advancement of industrial aquaculture, intelligent fish feeding has become pivotal in reducing feed and labor costs while enhancing fish welfare. Computer vision, as a non-invasive and efficient approach, has made significant strides in this domain. However, current research still faces three major issues: qualitative labels lead to models that produce only qualitative outputs; redundant information in images causes interference; and the high complexity of models hinders real-time application. To address these challenges, this study innovatively proposes the quantification of fish feeding behaviors through satiety experiments, enabling the generation of quantitative data labels. A two-stage recognition network is then designed to eliminate redundant information and enhance model performance. This network utilizes pose detection to extract key features, while a graph convolutional network (GCN) effectively models the topological relationships between fish posture and distribution, achieving a satiety classification accuracy of 98.1%. Furthermore, to reduce model complexity, lightweight RepSELAN and SPPSF modules were developed, resulting in a 31.4% reduction in parameters and a 26.2% decrease in computational load, with only a 0.11% decrease in mAP(B) and a 0.95% increase in mAP(P). Compared with existing methods, this approach outperforms conventional models in both accuracy and efficiency, providing a novel and efficient model foundation for developing intelligent feeding strategies.