data_augmentation_for_CNN

This project demonstrates image data augmentation techniques using TensorFlow/Keras to enhance CNN training on a car classification dataset. It applies random flips, contrast adjustments, and explores pipeline optimizations to reduce overfitting and improve model generalization.

Project Description

This project demonstrates the practical application of data augmentation techniques to improve model generalization and reduce overfitting in deep learning. Using the modified cars196 dataset from Kaggle's Computer Vision course, we implement a full training pipeline that includes image preprocessing, real-time augmentation, model building, training, and evaluation. The notebook is structured to be both educational and reproducible, suitable for learners and practitioners in computer vision.

Project Structure

data_augmentation_CNN/
├── data_augmentation_CNN.ipynb  # Main Jupyter notebook
├── README.md                     # Project documentation
├── LICENSE                       # MIT License
└── requirements.txt              # Dependencies

Key Concepts Covered

• Data Augmentation in CNNs: Using random transformations (contrast, flips, rotation) to artificially expand the training dataset
• TensorFlow Data Pipeline: Efficient data loading, caching, and prefetching with tf.data
• Model Integration: Building augmentation directly into the Keras model as layers
• Comparative Analysis: Training and evaluating two CNN architectures with different augmentation strategies
• Performance Visualization: Plotting training/validation loss and accuracy curves

Technical Implementation

1. Data Loading & Preprocessing

• Loads images from directory structure using image_dataset_from_directory
• Converts images to tf.float32 normalized range [0, 1]
• Implements data caching and prefetching for optimal GPU utilization

2. Augmentation Techniques

The notebook implements several augmentation methods:

• RandomContrast: Adjusts image contrast by a specified factor
• RandomFlip: Horizontal and vertical image flipping
• RandomRotation: Image rotation within specified range
• Additional techniques are commented for experimentation

3. CNN Architectures

Model 1: Standard CNN with Moderate Augmentation

• Augmentation: Contrast(0.1), Horizontal Flip, Rotation(0.1)
• Convolutional blocks with 64, 128, 256 filters (kernel_size=3)
• Activation: ReLU, Padding: 'same'
• Dense head: 8 neurons → 1 output (sigmoid)

Model 2: Alternative CNN with Strong Augmentation

• Augmentation: Contrast(0.5), Vertical Flip, Rotation(0.5)
• Convolutional blocks with 64, 128, 256 filters (kernel_size=2)
• Activation: SELU, Padding: 'valid'
• Dense head: 6 neurons → 1 output (sigmoid)

4. Training Configuration

• Optimizer: Adam with epsilon=0.01
• Loss: Binary Crossentropy
• Metric: Binary Accuracy
• Epochs: 40 (with early stopping potential)
• Batch Size: 64

Results & Evaluation

The notebook includes comprehensive evaluation:

• Training and validation loss plots for both models
• Accuracy progression throughout training
• Best validation loss and accuracy metrics comparison
• Visual examples of augmented images

Getting Started

Prerequisites

• Python 3.8+
• TensorFlow 2.8+
• Jupyter Notebook

Installation

# Clone the repository
git clone https://github.com/esosetrov/data_augmentation_CNN
cd data_augmentation_CNN

# Install dependencies
pip install -r requirements.txt

# Launch Jupyter Notebook
jupyter notebook data_augmentation_CNN.ipynb

Dataset Setup

The notebook uses the Kaggle Cars vs Trucks dataset. You can:

1. Use the provided Kaggle dataset paths in the notebook
2. Replace with your own image dataset by modifying the directory paths
3. Ensure your dataset follows the structure: train/class1/, train/class2/

Customization

You can easily modify:

• Augmentation parameters: Adjust factor values in augmentation layers
• Model architecture: Change filter sizes, activation functions, or add/remove layers
• Training parameters: Modify optimizer, learning rate, or number of epochs
• Dataset: Replace with your own image classification dataset

Expected Outcomes

After running the notebook, you will:

1. Understand how data augmentation affects CNN training
2. Have a reusable pipeline for image classification tasks
3. Be able to compare different augmentation strategies
4. Gain insights into model generalization and overfitting prevention

References

1. Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on Image Data Augmentation for Deep Learning.
2. Simard, P. Y., Steinkraus, D., & Platt, J. C. (2003). Best practices for convolutional neural networks applied to visual document analysis.
3. LeCun, Y., et al. Comparison of learning algorithms for handwritten digit recognition.

Contributing

Feel free to fork this repository, submit issues, or create pull requests with improvements or additional augmentation techniques.

License

This project is intended for educational purposes. The dataset may have its own usage restrictions.

Note

This notebook is designed for educational and research purposes. Real-world applications may require additional considerations and validation.