Long-Tailed Continual Learning For Visual Food Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Deep learning for food recognition faces challenges with new food classes and imbalanced datasets (long-tailed distribution).
• Existing methods struggle with continual learning and recognizing rare food items accurately.

Purpose of the Study:

• To develop a robust food recognition system capable of continual learning and handling long-tailed data distributions.
• To improve generalization for instance-rare food classes in real-world scenarios.

Main Methods:

• Introduced a new dataset of 186 American foods and benchmark datasets (VFN186-LT, VFN186-INSULIN, VFN186-T2D).
• Proposed a novel end-to-end framework using knowledge distillation to prevent representational misalignment during continual learning.
• Implemented an augmentation technique integrating class-activation-map (CAM) and CutMix for rare class generalization.

Main Results:

• The proposed method demonstrated significant improvements over existing approaches on multiple benchmark datasets (Food101-LT, VFN-LT, VFN186-LT, VFN186-INSULIN, VFN186-T2DM).
• The framework effectively enhances generalization for instance-rare food classes.
• Ablation studies confirmed performance gains, highlighting the method's potential.

Conclusions:

• The novel framework successfully addresses key challenges in long-tailed continual learning for food recognition.
• The proposed techniques offer a promising solution for real-world food recognition applications, especially for diverse and imbalanced food data.