Long-Tailed Food Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Food Science

Background:

• Image-based dietary assessment relies on accurate food classification.
• Real-world food consumption exhibits a long-tail distribution, causing severe class imbalance in datasets.
• Existing long-tailed classification methods are not optimized for the complexities of food image data, such as inter-class similarity and intra-class diversity.

Purpose of the Study:

• To introduce novel benchmark datasets (Food101-LT, VFN-LT) for long-tailed food classification.
• To propose a new two-phase framework to address class imbalance in food image classification.
• To evaluate the proposed framework against state-of-the-art methods for long-tailed classification.

Main Methods:

• Development of two new datasets: Food101-LT and VFN-LT, with VFN-LT reflecting real-world long-tailed food distribution.
• Implementation of a two-phase framework: (1) undersampling head classes with knowledge distillation and (2) oversampling tail classes using visually aware data augmentation.
• Comparative analysis against existing state-of-the-art long-tailed classification techniques.

Main Results:

• The proposed two-phase framework achieved superior performance on both Food101-LT and VFN-LT datasets.
• Demonstrated effectiveness in mitigating the challenges posed by long-tailed data distribution in food classification.
• Outperformed existing state-of-the-art methods for long-tailed classification on the introduced food datasets.

Conclusions:

• The novel framework effectively addresses the class imbalance problem in long-tailed food classification.
• The introduced datasets serve as valuable benchmarks for future research in this domain.
• The proposed method shows significant potential for application in real-life dietary assessment and related fields.