Food-101 Image Classification

Author: Sevendi Eldrige Rifki Poluan

A deep learning project for food image classification using transfer learning with EfficientNetB7.

Project Overview

This project focuses on image classification methods utilizing transfer learning techniques with the EfficientNetB7 model pre-trained on ImageNet. The model employs a hybrid approach where some layers are frozen while others are fine-tuned for weight updates during training.

Dataset

• Name: Food-101
• Classes: 101 food categories
• Total Images: 101,000 images
• Train/Test Split: 750 training images + 250 manually reviewed test images per class
• Image Resolution: Maximum side length of 512 pixels (rescaled by dataset authors)
• Dataset Source: ETH Food-101 Dataset

Requirements

• Python 3.7+
• TensorFlow 2.x
• scikit-learn
• NumPy
• Matplotlib

Installation

1. Clone the repository:

git clone <repository-url>
cd food-101-classification

2. Create a virtual environment:

python -m venv .venv

3. Activate the virtual environment:

# On macOS/Linux
source .venv/bin/activate

# On Windows
.venv\Scripts\activate

4. Install dependencies:

pip install -r requirements.txt

3. Download the Food-101 dataset and extract it to the project root directory:

food-101-classification/
├── food-101/
│   └── images/
│       ├── apple_pie/
│       ├── baby_back_ribs/
│       └── ...

Usage

Training

Run the main training script:

python food-101.py

Jupyter Notebook

For interactive exploration and experimentation:

jupyter notebook Food-101.ipynb

Model Architecture

• Base Model: EfficientNetB7 (pre-trained on ImageNet)
• Input Size: 224×224 pixels
• Batch Size: 16
• Training Strategy: Fine-tuning with layer freezing/unfreezing

Results

Sample Food Image

Confusion Matrix

Model Performance: Achieves up to 70% accuracy on the Food-101 test dataset with the basic EfficientNetB7 architecture.

References

• Food-101 Dataset: Learning a Deep Convolutional Network for Image Classification

  • Official dataset: ETH Food-101 | 101 Food Categories & 101,000 Images

• EfficientNet: Mingxing Tan & Quoc V. Le, EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

• Transfer Learning: Jason Yosinski et al., How transferable are features in deep neural networks?

• TensorFlow & Keras: Keras Documentation