A method for fusing attention mechanism-based ResNet and improved ConvNeXt for analyzing fish feeding behavior

Abstract

Accurately identifying fish feeding behavior in complex environments is crucial for optimizing feed management, improving feed utilization efficiency, and reducing aquaculture costs. Complex real-world environments, such as variations in water quality, lighting conditions, and background interference, make it difficult to distinguish feeding states. To address this issue, based on the fusion of attention mechanism-enhanced ResNet and an improved ConvNeXt (ResNet–MoVIT–ConvNeXt), a fish feeding intensity recognition method is proposed. A multi-scenario data augmentation method is designed to simulate complex fish feeding environments replicating real-world complex scenarios. The dual-branch model, combining ResNet and improved ConvNeXt, extracts local features from fish school images. The MobileViT module is then used for multi-level feature fusion, effectively capturing feeding behavior features for accurate feeding recognition. Finally, a multi-factor dynamic feeding strategy is provided, which combines fish biomass, water quality, and feeding states to reduce feed waste. This method introduces the MobileViT module into each stage of the ResNet and improved ConvNeXt networks. The proposed method is evaluated on real-world fish school datasets, achieving an overall accuracy of 99.19% and 98.5% for the medium state, surpassing existing comparative methods.