CNN based autoencoder for image denoising on the MNIST dataset

Dataset used

MNIST: http://yann.lecun.com/exdb/mnist/

The noisy dataset was created by adding a Gaussian noise of $\mu = 0$ and $\sigma = 0.5$. Test train splits are then made accordingly for the noisy and clean parts.

Results

Test result on epoch 10: Avg loss is 17.3354560546875

In the plot below, the first line of images are the originals, followed by their noisy counterparts, and then finally the denoised output from the CNN