Denoising Autoencoder Project

Overview

This project demonstrates the training and deployment of a denoising autoencoder on 100 grayscale images (64×64). It includes:

• A Jupyter notebook for model development, evaluation, and analysis.
• A Gradio web application for interactive denoising.

Links

• Colab Notebook: Open in Google Colab
• Gradio App (Hugging Face Space): Try it here

What It Does

1. Jupyter Notebook

• Data Preparation:
  • Fetches the Olivetti Faces dataset (400 × 64×64 grayscale images).
  • Randomly samples 100 images, splits into 80% train / 20% test.
  • Adds Gaussian noise (σ=0.1) to both sets.
• Model Definition:
  • Builds a Residual U-Net autoencoder with skip-connections and residual blocks.
• Training:
  • Trains with MSE loss, using EarlyStopping and ModelCheckpoint to save best_model.h5.
• Evaluation:
  • Computes Mean Squared Error (MSE) on train and test sets.
  • Plots training vs. validation loss curves.
  • Visualizes side-by-side comparisons of noisy, reconstructed, and original images.
• Analysis:
  • Discusses reconstruction errors, loss curves, and visual results to assess generalization.

2. Gradio Web App

• Interactive Interface:
  • Upload any image (grayscale or color).
  • Displays three outputs:
    1. Original (resized to 128×128 for display)
    2. Noisy (Gaussian-corrupted, σ=0.1)
    3. Reconstructed (denoised by the autoencoder)
• Deployment:
  • Simple app.py script and requirements.txt file enable one-step launch and public sharing.

Model & Tech Stack

• Model:
  • Denoising autoencoder based on a Residual U-Net backbone.
• Libraries & Frameworks:
  • TensorFlow / Keras for model building and training
  • scikit-learn for dataset loading
  • OpenCV for image processing
  • NumPy & Pandas for data handling and metrics
  • Matplotlib for plotting loss curves and examples
  • Gradio for the interactive web interface

Analysis & Observations

• Noise Removal: The model cleans up most of the noise while keeping faces recognizable and clear.
• Consistent Performance: Training and test errors are both low and nearly the same, which tells us the model learned useful patterns and works well on new images.
• Smooth Learning Curve: The loss steadily went down during training without large jumps or gaps, so it didn’t overfit or underfit.
• Visual Quality: Reconstructed images look sharp in general, although very fine details can appear slightly blurred. Trying different loss functions (like SSIM) might help bring back those tiny details.

Conclusion:
The Residual U-Net autoencoder exhibits strong denoising performance and generalization. Future enhancements could include multi-noise conditioning, alternative loss functions, or extension to color images and higher resolutions.