Fine-grained Chinese dish recognition on CNFOOD-241: A discriminative learning approach for improving category-level discrimination

Accurate recognition of food images is essential for image-based dietary assessment, nutritional tracking, and digital health applications. However, closed-set Chinese dish recognition remains challenging because many categories in the CNFOOD-241 dataset exhibit strong visual similarity, substantial intra-class variation, and subtle inter-class differences. In this study, we present an integrated discriminative training strategy for fine-grained Chinese dish recognition using RegNetY-32GF as the student backbone. Rather than proposing a new backbone architecture, the study focuses on improving category-level discrimination under a fixed convolutional architecture by combining teacher-guided knowledge distillation, margin-based classification, triplet-based embedding regularization, and mixed-sample augmentation. Across five random seeds, the proposed full model achieved 84.37 + /- 0.29% Top-1 accuracy, 97.66 + /- 0.11% Top-5 accuracy, 83.41 + /- 0.57% macro-F1, and 84.27 + /- 0.30% weighted-F1. Paired bootstrap analysis showed a positive Top-1 improvement of 0.651 percentage points over the baseline, with a 95% confidence interval of [0.491, 0.817], and the mean class-wise Delta F1 across five seeds was significantly greater than zero according to the Wilcoxon signed-rank test (p = 0.017803). These findings suggest that the proposed framework provides a modest improvement in category-level discrimination on CNFOOD-241, with statistically supported evidence at the dataset level, although the magnitude of improvement varies across individual categories. The training code, evaluation scripts, derived result files, and trained model weights are publicly available at the project repository and GitHub Release.