Enhancing Thai food image classification accuracy via hybrid sampling transfer learning

Thai cuisine is globally recognized for its diversity and visual richness, motivating the development of intelligent food recognition systems. This study proposes a simple hybrid sampling strategy to address class imbalance in a dataset of 10,782 images covering 22 Thai single-dish categories. Under-sampling, over-sampling, SMOTE, and hybrid sampling combined with data augmentation were systematically evaluated using seven transfer learning models: ResNet50, MobileNet, InceptionV3, EfficientNetB0, DenseNet121, NASNetMobile, and Xception. Experimental results demonstrate that increasing balanced data volume significantly improves classification performance. The Hybrid-1500 configuration achieves the most stable results, with EfficientNetB0 attaining 99.12% accuracy, 99.13% precision, 99.12% recall, and 99.12% F1-score. Moreover, EfficientNetB0 maintains competitive computational efficiency compared to larger architectures. The findings highlight the importance of balanced sampling and sufficient data scale in fine-grained food image classification and provide practical insights for handling imbalanced visual datasets in real-world applications. Furthermore, the novelty of this work is centered on its controlled hybrid sampling framework and the systematic evaluation of data balancing strategies across seven transfer learning models.