Abstract Deep convolutional neural networks have become widely adopted for food image classification, but their deployment on resource constrained devices remains challenging due to computational overhead, model redundancy, and increased architectural complexity caused by non contributing feature channels. Existing pruning approaches mainly emphasize parameter reduction, with limited focus on identifying and eliminating layer wise redundant feature responses while preserving semantic interpretability. In addition, Explainable Artificial Intelligence (XAI) techniques are largely confined to post hoc visualization and are rarely utilized for model optimization and structured compression. To address these limitations, this work proposes a Grad-CAM (Gradient-weighted Class Activation Mapping) inspired channel importance criterion for structured pruning of our SE-DenseNet121 model for food classification. The proposed framework identifies redundant feature responses, prunes non performing channels, reduces model cost and complexity, and retains only the most informative representations while preserving Dense Block connectivity. Using this importance score, 160 of 896 transition layer channels (17.9%) were removed. The pruned model was further optimized using Knowledge Distillation and cross entropy fine tuning. Experimental results show that the proposed framework reduces the parameter count by 8.39% and MACs by 5.15%, while achieving an accuracy of 98.13% (98.61% with test-time augmentation). Across five independent runs, the model attains a mean accuracy of 98.56% ± 0.07%, which is adopted as the primary reported metric. McNemar’s test shows a statistically significant but modest divergence between the pruned and teacher models, confined to 1.9% of the test images. Applied zero-shot to 3,045 images from three independent public datasets, the pruned model reaches 86.21% accuracy against the teacher’s 87.78%. These results demonstrate that explainability guided pruning improves computational efficiency while maintaining comparable recognition accuracy and interpretability for food image recognition. Funding Open access funding provided by SRM University-AP. No external funding was received for this research. The authors acknowledge institutional support