Another approach for OSC/DSLR cameras

[Astrocava]

Hi all!

I've been thinking if there is another way to preprocess images taken with OSC/DSLR cameras. I don't like the interpolations every debayer method makes. So... why not to use the real information converting each image in 4 channels (RGdGiB) using only the corresponding colour pixel? Each channel size will be 1/4 of the original image (that will improve the performance of old CPUs). This way I think we can do a better job with noise at pixel scale, without propagating to near pixels through the debayering process. 

However I'm not sure about its implications with spatial resolution or other problems that could arise. And what to do with two G channels?

I have no results yet, and even, I don't know how to evaluate them. It's only thinking out loud.

I'll be experimenting with the tools we already have: I can get the channels extracted (cannot separate each green channel) using this process cycle:

1.Read raw files with Create RAW Bayer image (RGB)
2.Channel extraction
3. Superpixel debayering each channel

The problem is that you can only use this with raw formats (cr2). The best solution is a small script that read the original file and extract the pixel values to make the 4 channels without limitations related with file format (cr2 or fits). This way we can use Master Darks, Bias and flats integrated the usual way to feed the script and obtain the masters for each channel to calibrate them. But my programming skills are...

What do you think?

Sergio


EDIT: Perhaps is better loading raw as CFA,  superpixel debayering and extract channels. We have the same problem with green channel averaged but we can use fits as input.

[Carlos Milovic]

There is a problem with the green data, since it is averaged in each superpixel. So, calibration frames may be messed up, and not work as they should. Preprocessing should be done always before debayerization.

[mschuster]

Tony Hallas presented a totally different procedure at AIC this weekend.

Debayer, stretch and process one single subframe to produce an RGB image. He used a different software app to do this. Then he applied the same processing to all of the dithered subframes. Then he integrated the RGB subframes. Then he tweaked the final integration.

I don't think he calibrated anything at all, arguing that modern cameras with good lenses and dithering there seems to be no need.

He sees "color noise" as the primary issue with modern equipment. His procedure mitigates this.

The result seemed excellent to me.

Mike

[Astrocava]

There is a problem with the green data, since it is averaged in each superpixel. So, calibration frames may be messed up, and not work as they should. Preprocessing should be done always before debayerization.

Yes, you are right. The best (and only) point to apply is to calibrated subs, just before registering them.

[Carlos Milovic]

Mike, what you describe is equivalent to process JPEG data (without a lossy compression). Staying in the linear stage have a lot of meaning, specially for calibration, deconvolution, gradient removal, etc. I understand that Tony uses general use photo-manipulation software, so it is understandable that he is trying to do what he can with the tools he have. We have not such limitations here.

[pfile]

i thought and thought about what tony presented, and i guess i have the following comments. obviously he knows what he is doing, but he's been doing years of deep sky imaging with a CCD and photoshop.

1) he's using a state of the art canon full-frame sensor camera (6D), which has been modified for Ha transmission, from a very dark site. i think that this is the main reason why his images turned out so well.
2) he described the dominant problem with DSLR images as "color mottling". but i'd argue that what he's seeing is dark current and sky noise in RGB, all at once. i can't think of a reason why a DSLR image would be different from an OSC CCD image, except that the OSC is usually cooled, hence you don't see as much effect from dark current, especially when the frames have been calibrated.
3) canon does play tricks with dark current suppression and the 6D is supposed to have really amazing high-ISO performance, perhaps better than the 5dmk3.

one very positive aspect of his approach is the use of adobe camera raw to correct lens aberrations. adobe has an extensive database of lens profiles and can quickly and accurately correct barrel/pincushion distortions, as well as CA, for a variety of lens/extender/camera body combinations.

it *may* be possible to calibrate a DSLR image and then bring it into camera raw for lens corrections, if the file is a 32-bit tiff. that would be the best of all worlds. luckily a guy who is a deep-sky imager is now on adobe's camera raw team so it's possible CR may support the kinds of stuff we need to do in AP vs. general DSLR photography.

i'll bet tony's images could be improved upon if they first went thru the normal calibration flow. but, if the end result is "good enough", then perhaps many people would be happy with a flow like this. but i agree that it is contrary to proper data reduction.

rob

[Ignacio]

I have played with workflows such as Tony's a couple of years ago (those were my DSS days), not by doing full processing of each sub, but by ad-hoc "normalizing" the histogram of the session's set by sample white-balancing, adjusting the black point and a little gamma, saving as debayered tiffs, and stacking. No calibration, and often results were better than doing the standard drill in DSS with full calibration.

My conclusion at the time, was that the debayering algorithm in Canon DPP was making the difference. This is something I notice even today, when I open a raw sub in Canon DPP, am always impressed by the outcome. Not sure what method they use.

Ignacio

[mschuster]

My conclusion at the time, was that the debayering algorithm in Canon DPP was making the difference. ... Not sure what method they use.

Proprietary algorithms. The camera manufacturers, Apple, Adobe, and other vendors are all competing for quality.

Mike