Learning the Difference of Few-Shot Food Data Using Multivariate Knowledge-Guided Variational Autoencoder.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Food image recognition is crucial for dietary monitoring, promoting healthy lifestyles, and preventing diseases like diabetes and obesity.
• Current methods struggle with few-shot learning due to limited annotated data.
• Few-shot food recognition requires robust methods that can generalize from minimal examples.

Purpose of the Study:

• To develop a novel variational generative method for effective few-shot food recognition.
• To address the limitations of existing methods in data-scarce environments.
• To improve the accuracy and reliability of food recognition systems with limited training samples.

Main Methods:

• Introduction of the Multivariate Knowledge-guided Variational AutoEncoder (MK-VAE).
• Leveraging handcrafted features and semantic embeddings as multivariate prior knowledge.
• Utilizing a feature distillation module for enhanced feature learning and a variational autoencoder for feature generation with boosted latent representations.

Main Results:

• MK-VAE significantly outperforms state-of-the-art few-shot food recognition methods.
• Demonstrated superior performance in both 5-way 1-shot and 5-way 5-shot settings.
• Validated effectiveness on benchmark datasets: Food-101, VIREO Food-172, and UECFood-256.

Conclusions:

• The proposed MK-VAE method offers a powerful solution for few-shot food recognition.
• This advancement can enhance dietary monitoring and disease prevention applications.
• MK-VAE shows strong potential for real-world applications with limited food image data.