Wiener filter output doesn't look good enough...

Is it possible to add a trainable filter after an autoencoder?...

Where is the Wiener filter implemented by Scipy derived from?...