Right now these functions are hard-coded to work using a (51,51) image inputs, padding it to (64,64) and cropping it back to (51,51). See here.

They sould adapt to different input size. This might need to add as new input parameter the input size of the images which we would use to define the right padding and cropping to the allowed values of the Learnlets (powers of 2).

The input dataset normalisation is done here.

This is not great, as the pixel value range can be quite high depending on the input energy and distribution of the dataset. It is better to scale the input value. Make sure the input eigenPSFs belong to the [-0.5, 0.5 ] interval.

• Fix the scaling on the network training.
• Fix the scaling on the proximal algorithm to replicate the required input of the networks.

This class can be found here.

• The prox class should have a different name than the Learnlet model.
• The input size is hard coded, it should be an input when instantiating the class.
• The parameters of the Learnlet class is hard coded too. This should not be the case, it should be an input of the proximal operator class as different Learnlet models could be used.

This should be the standard way to do things.

• One denoiser trained with global eigenPSFs
• One denoiser trained with local eigenPSFs

We need two proximal operator functions.

There is some bug with the classical wavelet denoising.

It can come from the threshold or the new wavelet prox denoising class.