Seeing What's on the Plate: Composition-Aware Fine-Grained Food Recognition for Dietary Analysis

Area of Science:

• Computer Vision
• Artificial Intelligence
• Nutritional Science

Background:

• Accurate visual food characterization is crucial for dietary assessment and health analysis.
• Fine-grained food recognition faces challenges due to high intra-class variation and subtle inter-class differences.

Purpose of the Study:

• To develop a fine-grained food classification framework enhancing spatial relation modeling and key-region awareness.
• To improve the discriminative feature representation for non-invasive food composition analysis.

Main Methods:

• Proposed a framework enhancing spatial relation modeling and key-region awareness.
• Introduced a lightweight multi-branch fusion strategy for heterogeneous feature integration.
• Designed a token-aware subcenter-based classification head for robust classification.

Main Results:

• Achieved 82.28% accuracy on FoodX-251 and 82.64% on UEC Food-256.
• Demonstrated improved sensitivity to composition-related visual cues and suppressed background interference.
• Showcased stable recognition across diverse acquisition conditions.

Conclusions:

• The framework provides reliable food characterization for dietary interpretation.
• Enables practical image-based dietary analysis under real-world conditions.
• Supports dietary monitoring and health-oriented food analysis applications.