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

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning is crucial for food classification due to data scarcity.
• Traditional methods struggle with limited datasets, increasing collection and annotation costs.
• Vision Transformer (ViT) models offer potential but require adaptation for few-shot scenarios.

Purpose of the Study:

• To introduce FSF-ViT, a novel Vision Transformer-based model for Few-Shot Food (FSF) classification.
• To reduce data collection and annotation expenses in food image recognition.
• To enhance the accuracy and efficiency of food classification systems.

Main Methods:

• Developed FSF-ViT by integrating image augmentation and adaptive global-local feature fusion.
• Trained the model using limited food image data.
• Evaluated performance on a custom Food-30 dataset and three benchmark datasets.

Main Results:

• FSF-ViT achieved a top classification accuracy of 95.1% on the test set.
• Demonstrated significant accuracy improvements over the standard ViT model (12.8% to 15.1% increase).
• Validated model effectiveness through classification result visualization.

Conclusions:

• FSF-ViT offers a low-cost, high-accuracy solution for few-shot food classification.
• The model provides efficient technical support for smart device-based dietary recording.
• Advances dietary management and health monitoring through improved food recognition technology.