2.5.1.4 autofocus fine with Edge11 f/10 0.4" per pixel

Just a quick update that I switched my Edge11 from f/7 to f/10 and 0.4" per pixel and tried some autofocus work with 2.5.1.4 and it all went very well. I kept the range of focus narrow to avoid donuts and got very nice parabolas that were quite repeatable. This is with robofocus on the primary focuser. I also did another unattended imaging session with periodic autofocus - and it went well.

I have some questions about how it’s all implemented - but so far it looks good for me.

Frank

A question for you Frank. Am I correct in assuming that you did not try the 2.5.1.4 with the f/7 configuration and the change to f/10 is just coincidental? And presumably you would expect it to work just as well if you go back to f/7?

Hi-

Yes - I was switching to f/10 anyway. The original autofocus worked ok but not very reliably at f/7. I expect the new version is more robust at both f/7 and f/10. And maybe with hyperstar at f/2.

Frank

I did some testing last night with 2.5.1.4. All I can say is that I am amazed at the improvement. I was imaging with an EdgeHD 14 with the reducer at 2737mm with a scale of 0.42"/px. I tried a number of steps and step sizes, and in almost every case, AF worked very well. I finally settled on a 9-step approach that took the stars well into donut-land, and the result was a very nice, sharp V-curve with straight, steep sides. Before the improvements, about the best I could hope for was a shallow U-curve that left a lot of vagueness in determining optimum focus.

I was imaging M101 with the QHY9, and the galaxy completely fills the frame. M101 is particularly challenging for AF because it has a whole lot of star-like knots that could be misinterpreted as stars. I’m pleased to report that the new algorithm very effectively rejected the galaxy’s knots and only selected real stars.

As the moon gets fuller, I think I’ll do some experimenting at F/11 at 3910mm. In the past, I have not been able to get AF to work at all at this FL. Last night’s results make me encouraged that the AF routine will work for that arrangement as well.

A big thanks to Ken and Jared for the hard work that they put into this. For me, at least, it has very clearly paid off.

Now if you guys get some free time, please start working on a software routine that eliminated differential flexure please.

Tim

I aim for a parabola that is fairly shallow and shows the curve near the minimum is well sampled. So I don’t get donuts but I do need to start near focus. With a steep V I can’t be sure I am near the true focus - and the same thing applied to routines like FocusMax. I’ll be experimenting more with this but so far it looks good.

Frank

Parabola?
Sampling across the bottom is not the idea behind the V curve.
( It might help assure the system is predicting the proper result.)
You should not set it like this for normal use.

HFRs starting at 5-6 are going to give adequate a V.

A V curve makes sense when using a single star - and when relying on a model to deduce where perfect focus is based on very out of focus views of the star. You only need to sample at one or two points and then you know focus. It is very fast and it’s what FocusMax does - but I don’t like that for sct work and I prefer to use many stars and sample through the curve - and not rely on a model.

But when using many stars in longer exposures and sampling many points through the curve there is no need at all to think in terms of a V. If the curve is repeatable and clearly samples the bottom of the curve accurately then you know you are getting optimal focus - and you know how much the HFR will change if you are perhaps 5 steps away. It works and works very well - and you have feedback on how close you are to focus. None of that would be possible if you only have a steep V.

Frank

I agree. Sampling at the bottom of the curve does not add much. That’s where all the noise is. I aim for the longest, steepest, straightest sides that I can get. I don’t really care what the apex looks like if I get good, straight sides that allow for a good linear regression. After every AF routine last night, I would slew to a nearby bright star and check focus with the Bahtinov mask. It was dead on every single time. On a couple of occasions, I ran the routine two or three times in a row, and got results that were within a few steps of each other - a small fraction of my CFZ. So for me at least, going well into donut land produces accurate and repeatable results.

The problem with trying to get a lot of sampling at the bottom of the curve is that the HFR is so noisy there. I can take two identical images at the same focus and get HFR values that might vary by 20%. I haven’t done the math (shouldn’t be that hard), but my sense is that getting a good fit on good asymptote data and then fixing focus at the intersection is probably more accurate at finding “true” focus that trying to map the low point of the curve with a lot of points.

On the other hand, if what Frank does works for him, there’s nothing wrong with that.

Tim

If the hfr is robust and relatively insensitive to intensity - which it should be - and if it doesn’t include false stars - then the parabola will not be noisy and it will give a much more accurate view of ideal focus. In addition - if you repeat focus you can confirm your final focus is right in the center of that parabola.

This approach worked before but not very reliably. Now it appears to work very well.

With sct’s and other scopes the overally V may not be a perfect hyperbola - which it should be according to theory. Therefore the final focus as deduced far from focus may miss the true minimum. There is no way to know for sure unless you see in detail where the final focus lands with respect to the true minimum. So a V far from focus may work ok - and it may not. It’s a blind process you just need to trust if you aren’t actually seeing how the HFR varies close to focus.

Attached is an example. This is a stronger V than I would like because those samples far from focus aren’t contributing anything. But it does show the shape of the bowl in detail and lets me see how the HFR varies near focus.

Frank

Frank:

I think we’re saying much the same thing. My curves look a lot like the one you showed with the notable exception that my asymptotes are quite a bit steeper and more balanced. I have also notices that going further out of focus makes the slope of the two fitted lines much more similar. With 9 points, I find that I typically get 3 points in a nearly straight line, three more points near focus (just like your pic) where the line starts to curve and then three more that form the other nearly straight line.

In the case of your graph, the intersection of the lines is slightly to the left of where I would guess the true focus would be. That’s because the right side fitted line has a shallower slope than the left side. If you go further out of focus, the two sides become more symmetrical, leading to a calculated focus value that is closer to the minimum of the curve. It looks like your step size is 5 steps (what focuser are you using? - that seems like a very small number). The plot minimum appears to be around 39091, but the SGP calculated minimum looks like it is somewhere around 39089. That is a difference that is somewhere around half your step size.

That is, if I am understanding how SGP calculates the focus point correctly. My understanding is that it picks the intersection of the two lines as the focus point - unless it doesn’t get a good fit, when it calculates a weighted minimum based on the lowest three values.

Like I said before, if it works well for you, that’s all that matters. For me, though, I have found that the new routine produces a much more symmetrical plot when it is taken substantially out of focus, and for me, I believe that produces more accurate results.

Tim

This is the honest truth… these rigs are all unique. People will often find success empirically… always go with that route.