Dennis-K-Koros/Lightweight-Neural-Networks-for-Kenyan-Local-Cuisine-Recognition-on-Mobile-Devices

NutriScan is an Android application using a lightweight MobileNetV2 neural network, optimized with TensorFlow Lite, for real-time, offline recognition of 12 Kenyan local cuisines. The project solves the challenge of cultural data bias and cloud dependency, providing fast and accurate on-device nutritional information in low-resource environment # Lightweight Neural Networks for Kenyan Local Cuisine Recognition on Mobile Devices **NutriScan** - Real-time Kenyan cuisine recognition powered by on-device machine learning Features • Getting Started • Dataset • Contributing --- ## Table of Contents - Overview - Project Achievement - Features - Technical Specifications - Prerequisites - Getting Started - Repository Structure - Dataset and Training - Usage - Future Works - Contributing - License - Acknowledgments --- ## Overview **NutriScan** is a lightweight, mobile-optimized system for real-time recognition of **12 Kenyan local cuisines**. Built with a fine-tuned **MobileNetV2** architecture and optimized via **TensorFlow Lite (TFLite)**, this solution enables fast, **offline inference** on standard Android devices without requiring internet connectivity. This project addresses the critical need for accurate, accessible dietary assessment in low-resource environments where network connectivity is often unreliable and where global food recognition systems fail to recognize culturally specific dishes. ## Project Summary & Achievement ### Problem Statement - **Cloud Dependency:** Most food recognition solutions require constant internet connectivity, making them impractical in areas with unreliable networks - **Cultural Bias:** Global food recognition datasets overwhelmingly focus on Western cuisines, leaving African and specifically Kenyan dishes underrepresented - **Accessibility:** Limited availability of nutrition tracking tools that understand local dietary patterns ### Our Solution - ✅ **Offline-First Design:** Complete on-device inference with no internet required - ✅ **Culturally Relevant:** Trained specifically on Kenyan local cuisines - ✅ **Lightweight & Fast:** 2.5 MB model with ~350ms inference time - ✅ **High Accuracy:** Achieved 78% Top-1 accuracy on validation set - ✅ **Accessible:** Runs on devices with Android 8.0+ and 4GB RAM ## Features - **12 Kenyan Dish Recognition:** Ide …