Multiclass Food Classification, project description.

The project is dedicated to the development of a computer vision model capable of automatically classifying food images into 101 categories. The final product is a Telegram bot. A user uploads a photo of a dish, and the system returns its name along with additional information, including characteristics and recommendations.

Several methods were used throughout the project:

• A simple CNN model (accuracy - 44%)

• CNN model based on the EfficientNet architecture (accuracy - 58%)

• Vision Transformers (accuracy - 75%)

• Improved Vision Transformers (accuracy - 86%)

However, the truly significant results with Transformers were achieved thanks to:

• Advanced augmentation, which creates complex training data.

• A partially unfrozen ViT for efficient data training

Setup Instructions

1. Prerequisites

• Python 3.12+
• Telegram Bot Token from @BotFather
• Api key from USDA

2. Installation

1. Clone the project with GIT LFS

2. Create a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

3. Configuration

Create .env file in the project root like .env.example

4. Run

python main.py