In view of the problems of the existing fish classification network models, such as poor anti-interference ability, high computational resource consumption, and difficulty in field deployment, this study proposed a lightweight fish intelligent classification and identification model based on the improved EfficientNetV2 model. By introducing Hybrid Dilated Convolution and Coordinate Attention modules, this model improves the model structure of EfficientNetV2, increases the receptive field, improves the global attention of the model to the fine-grained features of the target, and enhances the anti-interference ability of the model. After training, the model was evaluated by comparative ablation experiments, and the results showed that the accuracy of the EfficientNetV2 - HDCA model proposed in this study on the verification set was 97.01 %, which was 3.8 percentage points higher than that the accuracy before improvement. The number of parameters in the improved EfficientNetV2 - HDCA model is 22.06MB, which is 0.45MB higher than that before improvement. In order to visually demonstrate the effectiveness of the EfficientNetV2-HDCA model proposed in this study, the Grad-CAM thermal experiment was also passed. The experimental results show that the model can extract the features of key parts of fish more comprehensively than before, and has a certain anti-interference ability.