Image Integration Question

[wa4hnd]

Although everything seems to be working correctly, I am seeing the following during an Image Integration process.

Warning - No Convergence in MRS Noise Estimation Routine using K-sigma noise estimate.

I figure since is says "warning", there may be something I'm doing wrong or something I should be doing and not. <g>

Anyone have any ideas as to what direction I should look?

Like I said on top, the integration seems to be working fine... just thought maybe I should be doing something different.

Thanks,
   Al

[pfile]

long, long thread on this topic, good stuff though:

http://pixinsight.com/forum/index.php?topic=3320.0

[wa4hnd]

Thanks, I'll take a look.

Al

[wa4hnd]

Well I read through the thread but I didn't see a real conclusion on how to fix the problem... all my images are dithered and most of the time, the subs can be separated by mins, hrs or days ( sometime weeks ), so when I star align, there is always a certain amount of back along the edges after integration.

So the answer could be align / crop each sub, re-align / stack ?  If that's the case, would be nice to automate the crop... ie starting from the center... crop a percent.. ie, 10 percent  or save a percent... 90 percent.

Has there been a conclusion or is work being done about the problem?

Thanks,
  Al

[pfile]

well my takeaway from that thread is that there seems to be something funny about the way bilinear interpolation is debayering DSLR images, and that alternate debayering methods may not cause the noise evaluation problems. i'm not sure if juan is working on anything like that or not. i was confused by the mention of the different methods to debayer in the DSLR_RAW module because in the flow we use you have to debayer using sander's script which only offers superpixel and bilinear debayering of the calibrated subs.

[ruediger]

Has there been a conclusion or is work being done about the problem?

In my opinion, some very simple modifications should be done:

1) MRE noise estimation should ignore black pixels
2) noise weight based on a "rectangle of interest"
3) k-sigma noise estimate as a robust alternative weighting scheme
4) more demosaicing algorithms should be implemented
For "educational purposes" I implemented Hamilton Adams, LMMSE [1] and three variants of a weighted median filter based algorithm [2] in PSJR. They all perform visibly better than bilinear demosaicing. But there are bazillions of more algorithms out there and it is hard to make any recommendation. When reading the scientific papers, most algorithms are compared against a well-known set of scanned slides which have properties like smooth hue transitions, low contrast and so on, but are in no means a meaningful test set for "real life" pictures [3]. And noisy astropictures probably are even more challenging.

But despite of replacing bilinear debayering with other algorithms, I already found examples where MRE noise estimation fails on some pictures of a stack. So a combination of 2) and 3) would be really helpful.

Regards,
 Rüdiger

[1] http://www4.comp.polyu.edu.hk/~cslzhang/paper/LMMSEdemosaicing.pdf
[2] http://www.dei.unipd.it/~calvagno/Demosaicking/LiRandhawa-IEEETIP2009.pdf
[3] http://www.csee.wvu.edu/~xinl/papers/demosaicing_survey.pdf

[wa4hnd]

Number 2 would really be nice to have.... I don't use a DSLR but I am using an ST-8300c so I still need to debayer so more options there could maybe help.  I'm a PixInsight newbie here so I'm not in a good position to advise a course of action but it's good to know what the experts are thinking and what if anything is being worked on.

Al