RAW Conversion Snake Oil

[32BT]

This is a zoom of the noise form the other poster above. There is a very clear up/down, left/right component to it.

Good, you are starting to see the light.

That component is exactly the trade-off that each RAW converter needs to make somewhere and at some threshold. That threshold is different for each camera and each iso. It needs to be calibrated. It defines the final look of the file. You can have more blocky noise, or you can have more edgy noise. It would be useful if you could have an interactive slider for it, but alas, it is too deep within the engine to be viable at this stage of computing, and most people wouldn't know what to look for to set it. Not to mention that setting it for one part of the image can completely screw up another part.

Attached you can see what happens if you move the threshold too high. You can have your preferred noise pattern, but the result is checkerboards, or you have to accept some interaction within the noise, but at least get acceptable detail. 

[jeremypayne]

Good, you are starting to see the light.

Can we agree that this noise is not artificially inserted by the developers of the conversion software in an attempt to mislead users?

[32BT]

Can we agree that this noise is not artificially inserted by the developers of the conversion software in an attempt to mislead users?

I should hope I never implied that either directly or indirectly…

The noise is in the capture. Due to the disconnect between adjacent pixels as well as the individual channels, there is no way of knowing what constitutes noise, and what should contribute to interpolated signal. One can set a threshold, and then use different interpolation (or even generating a random value) depending on the signal to noise ratio, but a value has to be selected for the missing channels.

The latter can create more edgy noise (gradient based interpolation), or more blocky noise (linear interpolation), or even more noisy noise (inserting random values), but something has to be chosen below threshold. Incidentally that also coincides with the discussions we had previously about Bart's star image and the AA filter. Beyond nyquist, one can choose several methods of rendition, more blurring or more false detail. In the end it will all contribute to the final look of the image.

Probably even on a subconscious level. If the interpolation routines repeatedly select the opposite direction that the eye-brain combination chooses, then I am convinced that leads to very uncomfortable viewing experience…

[Fine_Art]

People still have to compare their version to the clean natural look of the full size jpg I provided. I doubt anyone can get close. NSA could beat my version, not many others.
Any if anyone doubts that file is from the sameISO3200 version. I am doing a conversion of an ISO6400 Sony A77 provided by someone else. With 24MP on APS-C behind a 1/2 stop pellicle the A77 is a good challenge.

[32BT]

I doubt anyone can get close.

Indeed. Especially by your standards. In fact, why don't you give Fuji a call. I have it from reliable sources that people are seeking a solution to the rather dreadful conversions that the mainstream tools currently offer. I'm sure your methodology will instantly solve 20+ years of failed debayering experience that camera and digital imaging companies apparently exhibit.

[jeremypayne]

I should hope I never implied that either directly or indirectly…

No, you did not ... but Mr Fine Art said so quite explicitly:

"The very strong red and green color noise you showed is only there to make the NR algorithm look good. That noise is not in the picture ...  If a software puts noise in at baseline just to take it out with other sliders its a game."

[32BT]

No, you did not ... but Mr Fine Art said so quite explicitly:

"The very strong red and green color noise you showed is only there to make the NR algorithm look good. That noise is not in the picture ...  If a software puts noise in at baseline just to take it out with other sliders its a game."

Ah, yes, I chose to ignore that for the sake of my own sanity...

[Fine_Art]

Sony A77 ISO 6400 provided by a guy in England. Only the one file was provided.

Rendered with everything off in IDC Note the color noise stays a fine random pattern.


Converted in Images Plus no adjustments. Both versions are weighted to dark colors.


Images Plus with curves, NR, Sharpening


IDC all off 100% crop Note the noise stays a fine spot pattern with no clumps.


Images Plus with curves, NR, Sharpening 100% crop. Notice the clumping of noise. It is created by me with the smoothing routines. The same larger spots of red and green that you see in raw converters. It is not noise from the sensor, it is smoothing routines.


Edit: All these images are 1920x1080 if you open the image.

[Fine_Art]

A different kind of smoothing that makes color clumps! I can make the clumps all kinds of shapes via different smoothing.


So the key to really good noise reduction is here. I'm not going to type it out again.
http://www.dyxum.com/DFORUM/looking-for-high-iso-a77-raw_topic95123_post1129602.html#top

You have to do it yourself in individual channels.

[Bart_van_der_Wolf]

So the key to really good noise reduction is here.

Hi,

Do you realise what it actually is that is happening? If I understand your procedure, you seem to be using the low resolution color separations from the R, G, and B, filtered sensels from the Bayer CFA data. These separations are sampled at every 2nd sensel, therefore at 50% of the Nyquist frequency. You then reduce noise at those spatial frequencies, and then do the Demosaicing on that frequency attenuated/filtered data.

Doing so with Images Plus, will unfortunately mean that Color management is out of the window, but one does have the benefit of doing the noiseband filtering at a high numerical precision, and subsequently converting the gamma to something like 1/2.2 will be able to use that precision for a slightly more accurate result before rounding to 16-bit channels.

It also means that you sacrifice accuracy in luminance demosaicing (we've seen the zipper artifacts), which would not happen with a proper dedicated noise removal tool which masks the detail, and only removes a tweakable amount of noise from low spatial frequency areas. I suppose a profiled (I don't have a profile for the cameras you showed samples of, so I can't show the optimal result) NeatImage run would achieve something similar. Also TopazLabs Denoise plugin delivers very useful results.

Cheers,
Bart

[Fine_Art]

Bart,

Yes, its very valuable to use the independent NR programs. I did my first test using the older noise ninja 2 with pretty amazing results. I expect to pick a newer NR package. Images plus also has extensive NR capability. It has several routines that preserve detail at the pixel level. You will not see your details smeared out. Look at the screenshot above I called "oversharpened". I was running the adaptive richardson-lucy on the luminance channel. The artifacts from too many cycles are easy to spot clean.

Its at a point where any theorizing about lost color channel detail is to ignore experimental results. It becomes mental exercise for it's own sake. This was my first pass at ISO3200.
http://fs05n5.sendspace.com/dl/b6ee75e0ce6682a4ecb9a9d059e5e1ce/50facb6d1781b996/w1tq12/CVT_D800hSLI03200NR0%20curves.jpg
My technique now is much better. The method is posted in that thread.

[Fine_Art]

Actually, I take that back. Theorizing about it is valuable. What just occurred to me is your (Bart) explanation of bayer filling in the "Foveon and full sensor capture" thread is very relevant here. In that thread you convinced me that de-bayer can do as good a job as full sensor capture on some scenes. Ok, look at the noise per channel using this method. A lot of what you are using NR to fix in this method is the black holes and bright jumps that are similar to the missing pixels in bayer.

Closing this down to discuss NR under a less controversial headline.
Arthur