FSF-ViT: Image augmentation and adaptive global-local feature fusion for Few-Shot Food classification

We proposed FSF-ViT, a Vision Transformer (ViT)-based model integrating image augmentation and adaptive global-local feature fusion, for Few-Shot Food (FSF) classification. The proposed method focused on training with limited food images to reduce data collection and annotation costs. This approach achieved the highest classification accuracy of 95.1% on the test set. Compared to the ViT model, FSF-ViT improved average accuracy by 12.8%, 15.1%, 4.6%, and 8.3% on our constructed Food-30 and three benchmark datasets, respectively. Furthermore, this study visualized the classification results and verified the validity of FSF-ViT. This study provided low-cost and efficient technical support for rapid online dietary recording using smart devices, advancing the development of dietary management and health. (The Food-30 dataset and implementation code: https://github.com/HZAI-ZJNU/FSF-ViT; dataset DOI: 10.5281/zenodo.15619141).