CIFAR10-CompressAI

Description

CIFAR10-CompressAI is a project that implements a convolutional autoencoder for compressing and reconstructing images from the CIFAR-10 dataset. The autoencoder is trained using a combination of perceptual loss and Mean Squared Error (MSE) loss, providing efficient compression while preserving the quality of reconstructed images.

Features

• Efficient Compression: Utilizes a convolutional autoencoder to significantly reduce the size of CIFAR-10 images.
• High-Quality Reconstructions: Combines perceptual loss and MSE loss to maintain the visual quality of reconstructed images.
• Data Augmentation: Implements advanced techniques to enhance the model's robustness and performance.
• GPU Support: Optimized for training on GPUs using TensorFlow, accelerating the training process.
• Modularity: Organized codebase with modular components, facilitating easy contributions and extensions.

Compression Metrics and Results

Compression Performance

• Original Size: 61,440 bytes per image
• Compressed Size: 5,120 bytes per image
• Compression Ratio: 12.00

Explanation of Results

The compression ratio of 12.00 indicates that each image is compressed to one-twelfth of its original size. This substantial reduction in size demonstrates the effectiveness of the convolutional autoencoder in minimizing storage requirements without compromising the quality of the images.

Why This Model Excels:

• Balanced Loss Functions: By leveraging both perceptual loss and MSE loss, the model ensures that reconstructed images retain essential visual features and textures, providing a balance between compression efficiency and image fidelity.
• Advanced Architecture: The convolutional layers in the autoencoder are adept at capturing spatial hierarchies and patterns in images, enabling effective compression.
• Data Augmentation: Enhancing the training data with augmentation techniques makes the model more robust and improves its generalization capabilities.
• Optimized Training: Utilizing GPU acceleration with TensorFlow significantly speeds up the training process, allowing for faster iterations and model improvements.

Overall, CIFAR10-CompressAI offers a powerful solution for image compression tasks, achieving high compression ratios while maintaining the quality of the original images.

Project Structure

CIFAR10-CompressAI/
├── data/                  # Folder for data
├── models/                # Storage for trained models and images
├── notebooks/             # Jupyter notebooks for exploration
├── src/                   # Source code
│   ├── data_preprocessing.py
│   ├── models.py
│   ├── train.py
│   ├── evaluate.py
├── .gitignore             # Files and folders to ignore by Git
├── README.md              # Project documentation
├── requirements.txt       # Project dependencies
├── LICENSE                # Project license
├── CONTRIBUTING.md        # Contribution guide

Installation

1. Clone the Repository

   git clone https://github.com/pierridotite/CIFAR10-CompressAI.git
   cd CIFAR10-CompressAI

2. Create a Virtual Environment (Optional but Recommended)

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

3. Install Dependencies

   pip install -r requirements.txt

Usage

Training the Model

To train the autoencoder, run:

python src/train.py

Evaluating the Model

To evaluate and compare the original and reconstructed images, run:

python src/evaluate.py

Contribution

Contributions are welcome! Please refer to the CONTRIBUTING.md file for more details.

License

This project is licensed under the MIT License. See the LICENSE file for more details.

Examples

Training

Compression Comparison

Advanced Usage

You can explore the notebooks in the notebooks/ folder for additional analyses and visualizations.

---

Thank you for using CIFAR10-CompressAI! Feel free to contribute and share this project with the community.