The Convolutional Autoencoder are not only good regenerators of image but also work on noisy input. I have added a sample image for better understanding
I have added weights, incase if you wish to just reuse them and generate the results. You could even train the model. Since it is MNIST dataset, it won't take much time to train it. Also try tweaking with the model structure, it will give you a better understanding of the things.
Happy Coding !!