Quote:
Originally Posted by Placidus
The Nyquist limit for a 2-dimensional image (as opposed to a 1-dimensional signal) is 3 pixels to the smallest recordable detail.
|
It's fantastic to hear you are across the conditions for when you'd want to bin (or not bin) and/or apply deconvolution. I'm hopeful then that you'd agree that the smallest resolved detail in your dataset is very, very far from 3 pixels.
E.g. a point-light at infinite distance (e.g a star) in your "Big One" image looks like this;
http://download.startools.org/Tutori...ts/MnT_PSF.jpg
This close-up of a star is very close to the point spread function (PSF - though a non-linearly stretched version of course) in your entire image. Every bit of real recordable detail is "smeared out" in this way. As you can see, the pixel in the middle (the center of the star) looks virtually identical in brightness to its surroundings. E.g. much of its energy bleeds into neighboring pixels. It's text-book oversampling and makes the scene blurry/soft for no good reason.
Quote:
Binning is usually done....
|
Firstly, I think you're confusing one particular flavour/application of binning ("hardware" binning) with the general procedure itself, which can be done in software. If the read noise consideration you expertly lay out does not apply, then it is
preferably done in software these days, as it gives you more flexibility (e.g. fractional binning/resampling or rejection algorithms) and avoids possible CCD well limitations too.
Atmospheric seeing is only half the story; your optical train and/or its configuration can equally have these effects (or worse). Perfect optics don't exist, and a circular opening comes with an innate PSF (the Airy disc) to begin with.
Quote:
In conclusion, in our case, the hard cold mathematics is that binning would achieve little, and inevitably throw away some information.
|
All I'm saying is that when reality doesn't match up with the mathematics and assumptions, it's usually not a flaw in the fabric of reality
Quote:
What would help us a lot is moving to 2.5 Km up in the Atacama desert, but it would be hard to raise our chooks steers and veggies there.
|
The animals can rest easy, as there is an easy (though not perfect) solution to this.
Bin (if it benefits your signal) and apply deconvolution to restore the lost detail. PI should have modules/functionality for both. It is important to note that binning lets deconvolution restore more using the improved signal, hence the recommendation to combine measures of both.
Right now, you are leaving detail on the table, which is a shame!
http://download.startools.org/Tutori.../MnT_Decon.jpg(your JPEG binned to a quarter resolution, then deconned; applying decon restoration to an already processed lossily compressed 8-bit image is a big no-no of course, so you should be able to do a lot better with much fewer ringing artefacts when using the real linear data)
Hope this helps!