The First Part of the Trilogy on the Use of Autoencoders – Noise Reduction

This time focusing on noise reduction. First, I will contaminate the entire dataset with a large amount of colour noise. I will compress this damaged image (to 512 digits) and then decompress it again. The excess noise was complex for the network to remember, so it is gone after the adjustment. 🙂

Generator of colour noise. Not too much, not too little – therein lies the charm of moderation – I take the second one from the right (10%).

Noise removed, albeit at the cost of some blurriness, but can you do better in Photoshop?

Sometimes I find the reconstructed photos more beautiful than the originals. 🙂

And once again, we lost glasses along with the noise. Unnecessary details must go.