Development and Validation of a Mobile Application for Real-Time Chinese Dish Classification and Nutrient Estimation

Area of Science:

• Artificial Intelligence
• Mobile Computing
• Nutritional Science

Background:

• Deploying deep learning models on mobile devices for dietary assessment is hindered by computational and size constraints.
• Previous work established a Chinese dish classification ensemble with 86.28% top-1 accuracy on the CNFOOD-241 dataset.

Purpose of the Study:

• To develop and validate a mobile application for real-time Chinese dish classification and nutrient estimation using model compression.
• To reduce model size and maintain inference speed for practical mobile deployment.

Main Methods:

• Implemented a two-stage joint compression strategy: palettization (4-bit weight quantization) and structured pruning.
• Developed an iOS application featuring dual nutrient estimation algorithms, user-friendly interfaces, and local data persistence.
• Conducted real-world validation with 25 participants and benchmark testing on the CNFOOD-241 dataset.

Main Results:

• Reduced model size by 85.9% (3.22 GB to 456 MB) with inference times under 2s on iOS.
• Achieved 91.67% classification accuracy in real-world tests and retained 98.1% of the original accuracy on the CNFOOD-241 benchmark.
• User experience evaluation reported 84% overall satisfaction and 100% positive ratings for navigation.

Conclusions:

• The developed pipeline successfully compresses, deploys, and validates a deep learning model for a practical iOS dietary assessment application.
• The compressed model demonstrates efficient resource utilization and maintains high performance for Chinese dish classification and nutrient estimation.
• The study highlights the feasibility of deploying sophisticated AI models for mobile health applications with significant size and computational reductions.