CMOS (Software) Binning, When and How?

Topic starter

Jun 20, 2021 · Edited: Jun 20, 2021

Hi,

I'm still searching for a good solution for my workflow. With my 8" SCT (reduced to f/7) and my CMOS, I would end up with approximately 0.35"p.pxl. I've integrated the unbinned data and worked with different integer downsampling and it seems that with a 3x3 binning, I would achieve an ideal sampling for my typical seeing conditions. At 4x downsampling one can clearly see a loss of image detail.

My question now is, when and how do you downsample/bin?

Many people capture in native resolution as for CMOS it's pure math, contrary to CCD. From a (simplified) mathematical point of view it wouldn't make a difference to integrate the subs and then bin but certainly it should be done on the linear image. I haven't elaborated how the game changes if clipping algorithms are involved.

From the tools that I've been working with, I haven't found a step/function to bin images. E.g., in the AstroPixelProcessor forums, the recommendation is actually to integrate in native resolution and then downsample through a bilinear interpolation. A common argument found is also that the high-res image would give a better star registration.

The reason for me not to bin on the camera/driver is that it would be much easier for me to create a dark library.

CS!
Björn

Engaging

dkamen

Jun 20, 2021

Hi Bjorn,

If you are oversampled then you must bin. It makes very little difference when you shall bin from the POV of the final result. I would say the sooner the better, because smaller raw files have obvious advantages: less storage, faster download time, faster and less demanding to integrate. Still, unless you are doing thousands of short exposures, it doesn't really matter and you can stay at 1x1 for simplicity.

If you are not oversampled it makes sense to work at native resolution all the way and then rescale the final result to improve its SNR and contrast. This has the advantage that you can do fractional scaling e.g, 1.7x1.7 binning which might suit your data best.

Now, unless your gear is obviously oversampling (e.g, 2 meters focal length with 2 micron pixels) you generally cannot know if a particular dataset is oversampled or not, as it depends on the seeing and usually can ne revealed only with close examination of the data after the fact.

So I would lean towards capturing 1x1 most of the time, and have the flexibility to bin/scale or not afterwards.

Btw, regardless of the above, I doubt the unbinned image gives better star registration. It has lower SNR and is far more likely to have stars with deformations that make the detector think they are not stars or place their center wrong. This is why we usually bin when platesolving. APP in particular in a recent project of mine would detect ~20-30 stars in the unbinned image and ~50-60 in the binned (2x2) one.

Cheers,
D.

Helpful Engaging

Björn Arnold

Topic starter

Jun 20, 2021 · Edited: Jun 20, 2021

Hi Dimitris,

Yes, I am talking about the cases where I'm likely oversampling. I determined the FWHM statistics in SiriL for several subs from several nights from different objects and the FHWM reported was always close to 2" (using the 8"SCT). Therefore, I would say that the 0.35"p.pxl. and 0.7"p.pxl are highly oversampled.
Storage is likely not a problem as I'm trying to use maximum dynamic range and also expose rather long.

I also agree with you on the registration. To me it would make more sense to improve the SNR as it would also allow to detect more stars (as you've observed) which should improve the registration fit.

How did you handle it then in APP?

I actually would consider to create binned calibrated subs which are then used for registration and stacking. However, I haven't found a tool yet which would do this as a batch process with explicit binning.

CS!

Björn

EDIT: PS: what algorithm do you use to downsample the integrated image? Do you use actual binning or an interpolation algorithm like bilinear?

Helpful Insightful Respectful Engaging

Björn Arnold

Topic starter

Jun 21, 2021

I've found out that ASTAP from Han Kleijn supports batch processing with binning 2x2 and 3x3. It also adjusts the fits headers accordingly and a repetitive 2x2 binning can be used to achieve 4x4. For the moment this seems to be the way to go.

Björn

Well Written Concise

SemiPro

IOTD Staff

Jun 21, 2021 · Edited: Jun 21, 2021

I'll usually mess with the resolution once it's all processed and ready to go. Once I'm done messing around with it in PixInsight, I'll mess with the zoom and see how much I can get away with and resize accordingly. Say I think it looks good at 80% percent its original resolution, then that is what it gets resized to.

That being said, I have a 5800x processor, lots of RAM and a good GPU so processing takes no time at all.

At 0.35 arc seconds a pixel I would dare suggest 2x2 from the get go might be a good idea. I don't know what kind of skies you are under but I think you would have to be extremely lucky to make use of such a resolution and 0.70 is still a really great resolution. Even at a 1.3 KM elevation and a relatively low humid environment I am lucky to see the lads at MetBlue forecast any lower than 0.8 . I doubt you would notice a loss of detail, but the business end of things (integrating, processing, etc) would go much faster.

Helpful

Most of the times, when I am binning, I ...