FoodVision AI 🍕🔬

An end-to-end image classification application that identifies 101 different food categories and estimates nutritional information using Deep Learning.

🚀 Project Overview

FoodVision AI leverages a fine-tuned MobileNetV2 architecture to provide high-accuracy food recognition. It features a modern React/Next.js frontend and a high-performance Flask backend, allowing users to upload images and receive instant predictions with confidence scores and caloric data.

Key Features

• Deep Learning Model: Fine-tuned MobileNetV2 achieving ~76% Validation Accuracy.
• Real-time Inference: Fast image processing and prediction pipeline.
• Calorie Mapping: Comprehensive database mapping 101 food classes to estimated caloric values.
• Responsive UI: A sleek, side-by-side dashboard for seamless user experience.

---

🏗️ Technical Architecture

🧠 The Model (Kaggle)

• Base Model: MobileNetV2 (Pre-trained on ImageNet).
• Technique: Transfer Learning followed by selective Fine-Tuning.
• Optimization: * Unfrozen the top 54 layers for domain-specific feature extraction.
  • Adam Optimizer with a reduced learning rate ($1 \times 10^{-4}$) for stable convergence.
  • Data Augmentation (Random Rotation, Flip) to prevent overfitting.

⚙️ Backend (Flask)

• Language: Python
• Package Manager: uv
• Core Libraries: TensorFlow, NumPy, Pillow, Flask-CORS.
• Functionality: Handles image preprocessing (MobileNetV2 spec), model inference, and metadata lookup.

🎨 Frontend (React/TypeScript)

• Framework: Next.js / React
• Styling: Tailwind CSS
• Icons: Lucide-React / Heroicons
• Functionality: Async file uploads, real-time preview, and dynamic result rendering.

---

🛠️ Installation & Setup

Prerequisites

• Python 3.9+
• Node.js & npm
• uv (Python package manager)

1. Backend Setup

cd Backend
# Create pyproject.toml and install dependencies
uv sync
# Ensure model.h5 and classes.txt are in the Backend folder
uv run main.py

2. Frontend Setup

# Navigate to frontend directory
npm install
npm run dev

---

📊 Model Evaluation (v0.1)

The current version of the model utilizes a fine-tuned MobileNetV2 architecture. Detailed metrics and training visualizations are stored in /evaluation/model-version0.1/.

Training Metrics

• Final Validation Accuracy: 75.9%
• Final Training Accuracy: 80.1%
• Loss: 0.71

Performance Visuals

Accuracy & Loss Curves	Confusion Matrix

Analysis: The model shows strong convergence. The slight gap between training and validation accuracy (~4%) indicates a well-regularized model that generalizes well to unseen data.

---

📁 Project Structure

FoodVision/
├── Backend/
│   ├── main.py            # Flask Server & Inference Logic
│   ├── model.h5           # Trained Keras Model
│   ├── classes.txt        # 101 Food Labels
│   └── pyproject.toml     # UV Dependencies
├── Frontend/
│   ├── src/
│   │   └── app/page.tsx   # React Dashboard
│   └── tailwind.config.js # Styling Configuration
└── README.md

👨‍💻 Author

Sakib Mansuri Software Engineering Student at Seneca Polytechnic GitHub Profile