FoodVision: Automated Food Detection Using YOLOv8

Project Overview

FoodVision is a deep learning-based food detection system that utilizes YOLOv8 to identify and classify various food items in images. The system is capable of detecting 55 different food classes with a focus on fruits and vegetables, making it useful for dietary monitoring and nutritional analysis.

Features

• Real-time food detection using YOLOv8
• Support for 55 different food classes
• Calorie estimation per 100g of detected food items
• Web interface using Streamlit
• Support for both image upload and camera capture
• Bounding box visualization with confidence scores

Model Architecture

• Base model: YOLOv8n (nano version)
• Input size: 640x640 pixels
• Batch size: 32
• Learning rate: 3e-4
• Training epochs: 45

Performance Metrics

• mAP50: ~0.8 (80% accuracy at 50% IoU)
• Precision: ~0.8
• Recall: ~0.75

Installation

1. Clone the repository:

git clone [email protected]:2302660/aai3001_final_project.git
cd aai3001_final_project

2. Install the required dependencies:

pip install -r requirements.txt

Usage

1. Run the Streamlit application:

streamlit run Sapp.py

2. Use the web interface to:
   • Upload images or capture them using your camera
   • View detected food items with bounding boxes
   • See confidence scores and calorie information

Project Structure

.
├── Model.ipynb         # Notebook for model training and evaluation
├── cal.py              # Core calorie calculation and detection functions
├── Sapp.py             # Streamlit web application
├── best.pt             # Trained model weights (not included in repo)
└── README.md           # Project documentation

Supported Food Classes

The model can detect 55 different food items including:

• Green foods: asparagus, avocados, broccoli, cabbage, etc.
• White/Beige foods: banana, cauliflower, garlic, mushroom, etc.
• Purple/Red foods: beetroot, blackberry, cherry, eggplant, etc.
• Orange/Yellow foods: apricot, carrot, corn, mango, etc.

Live Demo

You can try out the live demo at:

• Hugging Face Space
• Project Files

Team Members

• Brian Tham
• Hong Ziyang
• Javier Si Zhao Hong
• Timothy Zoe Delaya

Course Information

AAI3001 Deep Learning and Computer Vision, Trimester 1, 2024 Singapore Institute of Technology

Future Work

• Expand the dataset to include more food categories.
• Implement portion size estimation.
• Compare uploaded food images with dietary recommendations.