🖼️ AIMS Coursework

Image Denoising Autoencoder

Late 2024 — Fully Convolutional Network for Noise Removal

Autoencoder Denoising FCN LFWcrop

About

A fully convolutional autoencoder that removes Gaussian noise from face images. Trained on the LFWcrop dataset (64×64 color images), the network learns to reconstruct clean images from noisy inputs.

View on GitHub →

Network Architecture

Encoder

  • • 4 convolutional layers with ReLU
  • • 3 average pooling layers
  • • Input: (3, 64, 64) → Output: (32, 8, 8)

Decoder

  • • 4 transposed convolutional layers
  • • Restores spatial dimensions
  • • Input: (32, 8, 8) → Output: (3, 64, 64)

Training Details:

Loss: MSE
Optimizer: Adam
LR: 0.001
Epochs: 50

Results

0.0014

MSE (batch)

0.0015

MSE (full test set)

Training & Validation Loss

Training and Validation Loss

Smooth convergence over 50 epochs

Qualitative Results: Noisy → Denoised

Qualitative Evaluation

Top row: noisy inputs | Bottom row: denoised reconstructions

Example: Noisy vs Clean Training Data

Example Processed Images

Discussion

  • • Model effectively removes noise but sometimes produces slightly blurry reconstructions
  • • Given the low quality of original LFWcrop images, results are impressive
  • • Blurring occurs due to MSE loss smoothing pixel values

Key Takeaways

  • Autoencoders work well for image denoising tasks
  • Fully convolutional architecture preserves spatial information
  • MSE loss achieves good quantitative results but can cause blurring
  • Future work: UNet architecture, perceptual loss, other noise types