🧩 MNIST Denoising Autoencoder (PyTorch)

This project implements a Convolutional Denoising Autoencoder using PyTorch for the classic MNIST digit dataset.
The model learns to remove noise from images, reconstructing clear digits from corrupted inputs.

It’s a practical example of Deep Learning for Computer Vision, showcasing the use of Autoencoders for denoising tasks.

🚀 Features

• Convolutional Autoencoder built in PyTorch.
• Noise injection with Gaussian noise.
• Image denoising with clean reconstructions.
• Training pipeline with MSE loss + PSNR metric.
• Visualization of Original / Noisy / Reconstructed digits.

📂 Project Structure

notebooks/
autoencoder_mnist.ipynb    # main notebook (training & evaluation)
data/                      # MNIST is downloaded automatically
requirements.txt           # dependencies
README.md

🗃️ Dataset

The notebook automatically downloads MNIST using torchvision.datasets.MNIST.
Each image is 28x28 grayscale, with labels (0–9). Labels are not used in training (unsupervised).

▶️ How to Run

1. Create an environment and install dependencies:

   pip install -r requirements.txt

2. Open the notebook:

   jupyter notebook notebooks/autoencoder_mnist.ipynb

3. Run all cells to train and visualize results.

🔍 Results

• Input digits are corrupted with random Gaussian noise.
• The Autoencoder reconstructs clean versions of the digits.
• Evaluation metrics:
  • MSE Loss
  • PSNR (Peak Signal-to-Noise Ratio)

🛠️ Tech Stack

• Python
• PyTorch
• Torchvision
• Matplotlib
• NumPy

⚠️ Notes

• Training runs on CPU by default. If CUDA is available, the notebook uses it automatically.
• Results may vary slightly due to randomness in noise injection.