Filtering Gaussian noise in MNIST handwritten digits images using a convolutional autoencoder
A convolutional autoencoder was used to filter Gaussian noise that was added to images in the MNIST handwritten digits dataset. The autoencoder is composed of four convolutional layers: two layers each for encoding the input image and decoding the compressed image. The effectiveness of the autoencoder filter at denoising was assessed using Linfoot's image quality criteria of fidelity, structural content, and correlation quality. The results show that the autoencoder was able to successfully filter noise from images up to a threshold noise factor.