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

Deploying sophisticated deep learning models on mobile devices for dietary assessment remains challenging due to computational and model-size constraints. We address these challenges by developing and validating a mobile application for real-time Chinese dish classification and nutrient estimation, leveraging model compression. Building upon a previously developed ensemble achieving 86.28% top-1 accuracy on the CNFOOD-241 dataset, we implemented a two-stage joint compression strategy combining palettization (4-bit weight quantization) and structured pruning. This reduced the model size from 3.22 GB to 456 MB (85.9% reduction) while maintaining inference times under 2 s on iOS. The resulting iOS application incorporated dual nutrient estimation algorithms for 24 nutritional components, accessible user interfaces, and local data persistence with intelligent caching. Real-world validation with 25 participants demonstrated 91.67% classification accuracy (22 of 24 dishes); on the standardized CNFOOD-241 benchmark, the compressed model retained 98.1% of the uncompressed ensemble's top-1 accuracy. User experience evaluation showed 84% overall satisfaction, 100% positive ratings for navigation, and 56% recommendation likelihood. Profiling on iPhone 14 and MacBook Air confirmed efficient resource utilization, with peak memory of 369.7 MB and 505.4 MB, respectively, and nominal thermal states throughout. The weighted nutrient estimation method yielded numerically lower mean absolute error for 13 of 24 nutrients, with the largest reductions in copper (68.1%), dietary fiber (24.2%), and selenium (24.2%); however, after Holm correction, these differences were not statistically significant, and the only significant difference favored the non-weighted method (manganese). This work presents a pipeline for compressing, deploying, and pilot-validating a previously published Chinese dish classification ensemble, packaged as a practical iOS dietary assessment application.