FoodSnap AI: An Attention-Augmented MobileNetV2 Framework for Real-Time Food Recognition, Nutritional Estimation, & Personalized Dietary Recommendations on Mobile Devices

Unhealthy dietary habits are a primary driver of lifestyle diseases such as obesity and type-2 diabetes glob- ally. Existing calorie-tracking apps require manual food logging, which is tedious and often abandoned. FoodSnap AI addresses this by enabling users to photograph a meal and receive instant food identification and nutritional details. The system employs an Attention-Augmented MobileNetV2 (AA-MobileNetV2) model trained on 153,000 images across 151 categories, including a novel Kerala cuisine dataset of 2,000 images. Built with Flutter, Django REST Frame- work, and TensorFlow Lite, the app supports offline inference, OCR-based packaged food scanning, and personalized meal recommendations. AA-MobileNetV2 achieves 93.2% top- 1 and 98.7% top-5 accuracy with a 162 ms inference time and an 11.4 MB model size, outperforming standard mobile baselines. A user study with 20 participants recorded 90% overall satisfaction. FoodSnap AI demonstrates that accurate, offline-capable dietary monitoring is achievable on everyday smartphones.