This was a master course project completed at Utrecht University, Netherlands, on April 2021-June 2021
Main jobs:
• Designed Variational Auto-encoder(VAE) by Pytorch to investigate effectiveness for image denoising problems on MNIST- Fashion dataset. Trained DVAE model by carrying unsupervised training with L2-norm loss. Added additive Gaussian noise to images and compared reconstruction with original for varying latent dimensions.
• Improved VAE model by fine tuning hyper-parameters based on Structural Image Similarity Score (SSIM) and Peak signal- to-noise ratio (PSNR) criteria.
• Achieved 20% image denoising efficiency by DVAE model with truncated principal components analysis, compared with Fourier low-pass filter analysis.