Image Augmentation with TensorFlow

📌 Project Overview

This project demonstrates the impact of image augmentation techniques on model performance by training a neural network on the MNIST dataset. Key comparisons include model accuracy and generalization with/without augmentation.

Dataset: MNIST.
Goal: Evaluate how augmentation improves robustness and reduces overfitting in general Image classification tasks.

---

🚀 Key Features

1. Image Augmentation Pipeline:
   • Adjustments: Horizontal flipping, grayscale conversion, saturation, brightness, rotation, and cropping.
   • Real-time augmentation using TensorFlow’s tf.image module.
2. Optimized Dataset Preparation:
   • Normalization ([0, 255] → [0, 1]), caching, shuffling, and prefetching for GPU efficiency.
3. Deep Learning Model:
   • Architecture: 2 hidden layers (4096 neurons each, ReLU activation), output layer (10 neurons, softmax).
   • Trained separately on augmented vs. raw data for performance comparison.

---

🔍 Findings

• Augmented Model:
  • Accuracy: 94.2% (train) vs. 95.8% (test)
  • Runtime: 3s/epoch | Memory: 4GB (NVIDIA GPU).
• Baseline (No Augmentation):
  • Accuracy: 99.1% (train) vs. 94.4% (test)
  • Runtime: 3s/epoch | Memory: 3.8GB (NVIDIA GPU).
• Conclusion:
  • Augmentation improved test generalization by 1.4% while adding minimal computational overhead.

---

🛠 System Requirements

Dependencies

• Python 3.8+
• Libraries: tensorflow, tensorflow-datasets, matplotlib, Pillow
• Hardware: GPU with cuDNN support (recommended)

---

📄 License

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