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.
- Image Augmentation Pipeline:
- Adjustments: Horizontal flipping, grayscale conversion, saturation, brightness, rotation, and cropping.
- Real-time augmentation using TensorFlow’s
tf.imagemodule.
- Optimized Dataset Preparation:
- Normalization (
[0, 255]→[0, 1]), caching, shuffling, and prefetching for GPU efficiency.
- Normalization (
- 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.
- 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.
- Python 3.8+
- Libraries:
tensorflow,tensorflow-datasets,matplotlib,Pillow - Hardware: GPU with cuDNN support (recommended)
This project is licensed under the MIT License. See the LICENSE file for details.