<p>The accurate classification of fish feeding intensity serves as a fundamental prerequisite and key to realizing precision feeding in aquaculture. Aiming at the issue that current methods for quantifying fish school feeding behavior are susceptible to external conditions and exhibit low reliability, this study proposed a fish feeding intensity classification model based on an optical flow algorithm and improved deep residual shrinkage network. First, the Gunnar Farneback dense optical flow algorithm was used to quantify the movement state of fish during feeding accurately; subsequently, the Squeeze-and-Excitation Networks (SENet) attention mechanism was added to the network to focus the model on key features and enhance its important feature learning ability; finally, the soft threshold function was applied to reinforce effective information and suppress redundancy, thus achieving accurate fish feeding intensity classification. To assess the effectiveness of the proposed model, its performance was analyzed under different activation functions and model depths, and it was compared with classical Convolutional Neural Network (CNN) architectures such as Residual Networks (ResNet), Visual Geometry Group (VGG), ShuffleNet, and Inception. The experimental results indicate that the proposed model achieves a final evaluation accuracy, average precision, average recall, and <i>F</i>1-score of 97.53%, 97.53%, 97.46%, and 97.49%, respectively, all of which outperform the corresponding metrics of the aforementioned classical CNNs. Additionally, the model also has a significant advantage in terms of parameter quantity. In conclusion, the model proposed in this study can accurately evaluate fish feeding intensity in the complex environment of a Recirculating Aquaculture System (RAS), balancing performance and parameter efficiency well and laying a solid foundation for its subsequent practical use in production.</p>