CNN-Based Food Image Classification

Minho Song

Overview

This project focuses on using Convolutional Neural Networks (CNNs) to classify food images from the Food-11 dataset. The project involves experimenting with different CNN architectures, evaluating their performance using F-1 scores, and applying transfer learning with ResNet50.

Dataset

• Source: Kaggle - Food-11 Image Dataset
• Description: The dataset contains 16,643 food images categorized into 11 major food categories. For this project, a random extraction of 3,000 images was performed due to computational constraints.

Project Structure

• food_classification.ipynb: Main Jupyter notebook containing the entire workflow, including data preprocessing, model building, and evaluation.
• requirements.txt: List of dependencies required to run the notebook.
• .gitignore: Excludes unnecessary files from the repository.

Key Features

• Evaluation Metric: F-1 score was chosen as the evaluation metric to balance precision and recall, which is crucial for minimizing both false positives and false negatives in food categorization.
• CNN Architectures: Several CNN architectures were explored, including basic CNN models, ResNet-style models, and transfer learning with ResNet50.
• Cross-Validation: Nested cross-validation with stratified k-fold was used to ensure robust model evaluation.
• Visualization: Comprehensive visualizations of model performance, including training history and ROC curves.

How to Use

1. Clone the repository:

   git clone https://github.com/minhosong88/convolution_neural_network.git

2. Install dependencies:

pip install -r requirements.txt

3. Run the notebook: Open food_classification.ipynb in Jupyter Notebook and execute the cells to reproduce the results.

Results

The project found that simpler CNN architectures outperformed more complex models like ResNet in this specific task, indicating that model complexity does not always translate to better performance.

References

• Moro, S., Rita, P., and Cortez, P. (2012). Bank Marketing. UCI Machine Learning Repository. https://doi.org/10.24432/C5K306
• Kaggle - Food-11 Image Dataset. https://www.kaggle.com/datasets/trolukovich/food11-image-dataset