Accurate assessment of fish appetite intensity is essential for achieving precise feeding in aquaculture. To address the current issues of low accuracy and poor real-time performance in fish appetite intensity evaluation, this study proposes a fish appetite intensity assessment model based on an improved MobileNetV2 and transfer learning. First, we enhance the MobileNetV2 architecture by introducing the CBAM attention mechanism module after the skip connections in inverted residual blocks, thereby significantly improving the model's ability to capture key feature information. Next, we collect surface video data of fish feeding behavior in actual aquaculture environments. Through keyframe extraction and redundancy removal using hash-based differential threshold methods, we construct a real-world fish appetite intensity dataset. Validation experiments on this dataset show that after 200 iterations, the optimized model outperforms other models, with a memory footprint of only 9.3 MB, achieving 92.75% accuracy, 92.92% recall, 92.65% precision, and 92.70% F1-score. The proposed model offers high assessment accuracy with a compact memory footprint, providing significant technical support for intelligent and precise aquaculture management.