- New Auto Focus and Close Stars

Hello all,
Tis daylight and I downloaded and tried out the new star detection system using an existing .Fit sub I had just to see what happens. Not sure what star size slider is set to but it is about half way between the two middle notches. Clicking the ‘Calculate HFR’ button after opening my Fit file does a nice job of detection of the stars but I did notice a couple of things I thought I’d mention.

SGP seems to be seeing two separate stars which are very close as a single star, see pic’s for examples:

My image is available via Dropbox in case:


Is this situation likely to cause problems with the accuracy of SGP’s HFR calculations when Auto Focus is running ??


I have also seen this on a noise pattern about 1-2 pixels wide and about 10 pixels long that is on only one 1 frame. These appear in random places with my cooled 6D.

Ken, I would suggest you add a numeric display of the star size slider value. Several people have mentioned that they are guessing at what it is set at, including me.

I just finished analyzing the autofocus packs that I took with under some clouds using Discovered some things that should be very helpful regardless of the seeing conditions.
I am assuming the new autofocus routine uses the exact same code as the Star button does, since I was using it with to analyze the AFpacks I took with

Target was a very star poor area. Imaging 8300M on RC12. Min star size set to 4 for all packs.


  1. new AF routine has a hard time distinguishing real stars from noise, even with the min star size set to 4, in the special case of 0 star images.
  2. when there are real stars in the image and no donuts, star detection was excellent.

The first 2 images in the 1st AF pack were quite illuminating. Both were images of the inside of my dome so there were 0 stars in either one. In the directory for AF pack #1 are 2 PNG images showing my screen with the Star turned on.

Amazingly, the first image (#0) found 300 stars, all of size between 20x20 to 22x22 with average intensity of about 1.5 standard deviations from the whole image mean.
And equally interesting is the next image, found 111 stars, most with a size of 4x4 and a few 5x5 and 6x6. None larger. All with a mean close to 2 standard deviations.

Seems to me, and please forgive me if I am grossly over simplifying this, but it should be fairly easy to never find stars that are actually noise. I would only look at possible real stars if their mean intensity is at least 3 standard deviations from image mean. When I looked at the final (best focus) image of the 2nd focus pack, it had good focus and all 94 identified stars were really stars, and the dimmest one I found was 2.1 standard devs. Almost all the rest were at least 4x or much higher.

Adopting this approach should completely eliminate any need for a minimum star size slider. The program can just find that optimum size.

For the several images I looked at that were actually sky, star detection was excellent. There were many somewhat fuzzy small stars, clearly bigger than 4 pixels that were not identified as stars. Not sure if that is a problem or not.

Overall, AF routine definitely getting much better. I am anxious to get to testing the donut phase, which awaits another collimation attempt.

I don’t think so… if the stars are touching there is nothing to be done (in any way that is still computationally fast)… there are essentially the same shape. We are taking shortcuts to make all of this work in a way that is highly accurate (not 100% accurate) and reasonably fast.

It’s not that amazing… I can do all sorts of cool stuff.

Ken you are Amazingly witty. And I fully agree with “make all of this work in a way that is highly accurate (not 100% accurate) and reasonably fast”.

Jmacon - since updating my dark frames, false readings off hot pixels has virtually gone away. Are you using dark frame subtraction?

No buzz, I have never used dark frame subtraction in the auto-focus module. I don’t think it matters for me because I have never noticed any hot pixel stars in the AF frames for either the SBIG or the Canon. It has been curious to me that many others have reported these, but I never see any.
In the two initial inside of dome images I referred to above, those are not hot pixel stars. They are simply 1.5 sigma bulges in the mean noise level which the routine is very cleverly finding and crying, “a ha! I found another star”. Definitely cool Ken. I love humor in this hobby.
In that second frame where it found 111 stars, none of the ones I looked at closely were one of the very few actually somewhat hot pixels. SGP said the image mean was 2617, stddev 56, and max 9137, which makes sense for a 2x2 binned image. I also confirmed these numbers in PixInsight.

Thanks, by hot, I probably mean warm. I saw a definite improvement in false star detection by creating a set of integrated dark frames, ( 2-30 seconds in 2 second intervals, 10 frames each). It doesn’t hurt and I guess I’m making a point in a round about way that if the program has the facility to do dark frame subtraction, it is difficult to judge false star detection if we are not using it.

Excellent point buzz, and I will definitely crank some out. I have been intending to do so since I first noticed this option a few weeks ago, but too busy putting out the fires to get to it.


I don’t think the new AF routine will be enhanced by darks… this functionality may just be removed completely. Odds of false detection of CCD defects as stars are pretty much eliminated.


I messed around with std deviation based star filtering and it works pretty well. I was never able to get the image of the inside of your observatory to report 0 stars… well I mean I can get it to do that, but at the cost of eliminating good data from good frames. Somewhere in the middle… 2 std deviations? from background seems like a pretty good balance. Still allows for the detection of faint stars, but doesn’t find a thousand stars in a field of nothing now (it still finds about 5 though). It can also make the routine a great deal faster in some cases…

That sounds great Ken. When I saw those first 2 image results, it really seemed to click for me.
Maybe a useful additional rule is to boost it a little higher even if you lose some good data. After all, apparently you don’t need many stars for a good result. Have you done any studies comparing reliability vs star count? Why not purposely squeeze the star count on both ends to just end up with say 20 to 50 stars, but stars of the most reliable profile?
Plot a profile of overall reliability vs star size or some other star quality. If you can identify a more favored set of star traits, use that to filter the used star count down to an optimum total number.
Easy, right!