Convolutional Auto-Encoder for Image Reconstruction and Denoising

This project involves building and training a convolutional auto-encoder model for the tasks of image reconstruction and denoising using the CIFAR10 dataset.

Table of Contents

• Image Reconstruction
• Denoising
• Results
• Libraries Used
• References

Image Reconstruction

Objective

Design and train a convolutional auto-encoder for image reconstruction on the CIFAR10 dataset.

Methodology

• Utilized the CIFAR10 data loader from PyTorch.
• The auto-encoder model consists of 2 encoder blocks (2D convolution) and 2 decoder blocks (transpose 2D convolution).
• Channel configurations:
  • Encoder: (3,8), (8,8)
  • Decoder: (8,8), (8,3)
• Activation functions: Relu and Tanh.
• Loss: Mean Squared Error (MSE).
• Optimizer: Adam.

Results

• The model converged at a loss of 0.005 by epoch 9. The plot can be seen below:

Denoising

Objective

Train the convolutional auto-encoder for image denoising on CIFAR10.

Methodology

• The input images were perturbed with Gaussian noise (mean=0, variance=0.1) using the torch.randn function.
• Same model and training configurations as the image reconstruction task.

Results

• The model converged at a loss of 0.017 by epoch 8. The plo can be seen below:

- Performance Metrics: - Average PSNR: `23.860` - Average SSIM: `0.867`

Libraries Used

• PyTorch
• skimage (for PSNR and SSIM calculations)

References

1. PyTorch CIFAR10 DataLoader