search

CosmoStat / ShapeDeconv

Repo for Machine Learning-based deconvolution
MIT License
7 stars 1 forks source link

Setting Tikhonov parameter #11

Closed fadinammour closed 3 years ago

fadinammour commented 3 years ago
fadinammour commented 3 years ago

In the Unet evaluation notebook, in the preprocessing function pre_proc_unet, after applying the Tikhonov filter, a convolution is performed using gf.kconvolve.

I assume this convolution is used to give the image the resolution of the target which is HST. However the PSF that is used corresponds to CFHT which is the input.

Is there something wrong or did I miss something?

fadinammour commented 3 years ago

Error confirmed. The convolution after the Tikhonov filtering is performed with the HST PSF (target PSF).

fadinammour commented 3 years ago

The Tikhonov implementation has been checked and seems correct. The option to switch between the identity and the Laplacian regularizations has been added. The line on the value of the regularization weight has been performed on the identity regularization. In fact, do not need an accurate estimation of this parameter. We only need a value that stabilizes the output. This approach is inspired by the ForWaRD method. In the notebook for the estimation of the Tikhonov parameter, the value that was calculated is 9E-3.