Deconvolution routines and example (in progress) #47
Conversation
|
|
||
| ; Define the elements of the transformation matrix. | ||
|
|
||
| matrix = fltarr( w_dim^2, h_dim^2 ) |
There was a problem hiding this comment.
These squares overflow when one of the dimensions provided is a non-long integer greater than 181, resulting in bad behavior. While you can typecast here, it probably makes more sense to typecast in lines 48 and 49.
|
Here's my playing around with it (original/through-PSF/10-iteration-deconvolution): |
|
Regarding the docs example, to my eye, ~10 iterations is way too many for the coronal source. ~5 iterations seems more appropriate. |
|
@ayshih Thanks for trying this out! I'm pleased to see this recovered the original morphology quite well. This was probably a high-statistics case. In the future I think we can probably come up with some ad-hoc rule of max#iterations/#photons, though the tricky cases will be when we have somethings faint and bright together. For that, eventually, we'll need better stopping rules, but it's probably not necessary for the proposal. I assume you generated a custom instrument response matrix; did this take a long time? I have designs on speeding this up for a future version. |
The 256x64 matrix took about half an hour to generate (and the save file is a beefy ~1 GB). |
|
Wow! Impressive. I've never tried it over such a large FOV (3 arcmin x 3 arcmin) before, on the thinking that it would never be necessary. There are certainly ways the transform calculation can be optimized, if people think FOVs larger than ~1 arcmin are something they'd like to do. |
I'd like to make the method I use for image deconvolution available to everyone. This will probably take some iteration to get it usable by someone who didn't write the code, so I would appreciate if you can start trying it out. Please try the example in the IDL "docs" folder. If it works, please also try it for some other FOXSI simulated image that you might produce. Let me know what works and what breaks, and what is opaque and confusing. Thank you! @ehsteve @ayshih @aringlis @DanRyanIrish @PascalSH