Demosaic - Quick comparison of three raw converters and four algorithms

[Bart_van_der_Wolf]

Your animation is very instructive and complements Erik's excellent work. When viewing the extinction radii of the star chart, I see that the radii are approximately the same and I interpret the radii as the resolution near the Rayleigh limit. But how do the MTFs compare at lower frequencies? To my eye, contrast at higher frequencies appears slightly better in the renderings showing more Moire.

Hi Bill,

The blur radii are approaching Nyquist (at 4x 92 = 368 pixels diameter), but Capture One 7 (and later versions) does extract more detail (smaller blur diameter). Also, because of the lower amplitude of artifacts near Nyquist, more deconvolution iterations can be used, to restore resolution instead of amplifying artifacts.

Cheers,
Bart

[bjanes]

Hi Bill,

The blur radii are approaching Nyquist (at 4x 92 = 368 pixels diameter), but Capture One 7 (and later versions) does extract more detail (smaller blur diameter). Also, because of the lower amplitude of artifacts near Nyquist, more deconvolution iterations can be used, to restore resolution instead of amplifying artifacts.

Cheers,
Bart

Bart,

Thanks for the information. Artifacts are never welcome. They might not show up in one's prints with routine processing, but they could be amplified by deconvolution and limit the number of iterations. Adobe hired Bruce Fraser and Jeff Schewe to help implement capture and print sharpening in Lightroom and perhaps they should hire Emil Martinec to implement amaze demosaicing for LR/ACR. The LR/ACR capture sharpening is also getting dated.

LR/ACR is good for volume work, but when one needs to go into Photothop and break the parametric workflow and render to a TIFF, it might be advisable to demosaic with Capture One or RawTherapee and capture sharpen with one of the better deconvolution algorithms.

Regards,

Bill

[ErikKaffehr]

Hi Bill,

Pretty much my opinion. I would like to thank all you gentlemen posting on this thread for turning a boring post into an interesting discussion.

Hopefully you don't mind I get back to this thread in a few days as I happen to have an interesting couple of images to look into. To me it seems that RawTherapee gets more and more interesting now that OLP filtering has been dropped more and more in the camera industry.

Best regards
Erik

Bart,

Thanks for the information. Artifacts are never welcome. They might not show up in one's prints with routine processing, but they could be amplified by deconvolution and limit the number of iterations. Adobe hired Bruce Fraser and Jeff Schewe to help implement capture and print sharpening in Lightroom and perhaps they should hire Emil Martinec to implement amaze demosaicing for LR/ACR. The LR/ACR capture sharpening is also getting dated.

LR/ACR is good for volume work, but when one needs to go into Photothop and break the parametric workflow and render to a TIFF, it might be advisable to demosaic with Capture One or RawTherapee and capture sharpen with one of the better deconvolution algorithms.

Regards,

Bill

[Pictus]

AMaZE demosaicing is amazing, my preferred algorithm.
But for images with moiré LMMSE is better.

Raw file http://movies.dpreview.com.s3.amazonaws.com/pentax_k5iis/IMGP0101.DNG.zip


RAW file http://www.imaging-resource.com/PRODS/sony-a7r-ii/A7R2hSLI000100NR0.ARW.HTM

Only LMMSE or IGV render the fabric without moiré...
Unfortunately the best commercial RAW converters do not support AMaZE or LMMSE.

[ErikKaffehr]

Thanks for very good demo, carry on Sir!

Erik

AMaZE demosaicing is amazing, my preferred algorithm.
But for images with moiré LMMSE is better.

Raw file http://movies.dpreview.com.s3.amazonaws.com/pentax_k5iis/IMGP0101.DNG.zip


RAW file http://www.imaging-resource.com/PRODS/sony-a7r-ii/A7R2hSLI000100NR0.ARW.HTM

Only LMMSE or IGV render the fabric without moiré...
Unfortunately the best commercial RAW converters do not support AMaZE or LMMSE.

[Bart_van_der_Wolf]

AMaZE demosaicing is amazing, my preferred algorithm.
But for images with moiré LMMSE is better.

Hi,

The great thing about RawTherapee is that it offers options, lots of them. LMMSE is a demosaicing algorithm aimed at conversion of Higher ISO images, and because some aliasing artifacts may look like noise to the demosaicing algorithm, it may indeed help to try it. 'AMaZE' (Aliasing Minimization and Zipper Elimination) on the other hand has an additional final step in its algorithm, a Zipper artifact Elimination step, which softens the image of edges a bit to avoid artifacts on sharp edges, like those from sensors without an Optical Low-Pass Filter (OLPF). Another algorithm, DCB, produces similar results to AMaZE. From the documentation, "AMaZE can often be slightly superior in recovering details, while DCB can be better at avoiding false colors especially in images from cameras without anti-aliasing filters."

The RawPedia documentation page for RT mentions:"After reading though this article you should know that LMMSE and IGV are to be used on high ISO photos and that for the majority of other cases you should stick with the default AMaZE method, but of course you are free to explore for yourself and to test each method out on your own raw files."

Cheers,
Bart

[r010159]

Hi Bill,

The blur radii are approaching Nyquist (at 4x 92 = 368 pixels diameter), but Capture One 7 (and later versions) does extract more detail (smaller blur diameter). Also, because of the lower amplitude of artifacts near Nyquist, more deconvolution iterations can be used, to restore resolution instead of amplifying artifacts.

Cheers,
Bart

What is "deconvolution' with respect to image processing? This is a term I came across in an engineering class many many years ago.

Bob

[Bart_van_der_Wolf]

What is "deconvolution' with respect to image processing? This is a term I came across in an engineering class many many years ago.

Hi Bob,

In Digital Signal Processing (DSP), e.g. of images, one can describe the blur operation (caused by e.g. optics) as a convolution function. The convolution is usually referred to as the Point Spread Function (PSF), the diffused pattern of a point light source.

To invert the blur function, one uses a mathematical operation called Deconvolution, which restores the original source signal from the blurred one. That's why it's usually called 'restoration'  rather than 'sharpening', because the impression of sharpening can also be simulated by boosting edge contrast (but that doesn't really restore resolution, it just mimics it by fooling our eyes).

Cheers,
Bart

[Hoggy]

Wow..  Some of you guys are far more advanced in all this.   

But I have lately been thinking that the LR pv2012 is becoming long in the tooth.  It's mostly good, but I'm hoping the future of LR brings with it alternate process versions (including demosaicing algorithms) - like pv2016-DXO and pv2016-C1. And thanks to Bart, I have a new respect for Capture One now.  If only the usability/workflow issues can become improved to the level of LR, with [better] support for DNG/XMP-stored-snapshots(variants)/stacking(as in derivative files)/duplicate-checking/history/etc.

I have to say though - I'm feeling better about having a Pentax K-3.  AA-Off, AA-level 1, AA-level 2, due to in-body shake reduction.  I just wish my Canon G7 X didn't have an AA filter as well - since it's a tad easier to carry around, just a tad.    I almost want to send it in somewhere to remove the AA filter, but I think it would be prohibitively expensive - like with LifePixel, for instance.

Also, like Denis, I'm thinking that processing technology for raw files is still in its infancy. ...  And even sensor technology, for that mater.

[Pictus]

Thanks for very good demo, carry on Sir!

Erik

You are welcome, oops!
I forgot C1!!
C1 V8 also did a good job with the red fabric, considering that
LMMSE/IGV are not the best for the whole image as they are made
for higher ISOs, C1 is the best overall with this image...

[Pictus]

Hi,
(...)
The RawPedia documentation page for RT mentions:"After reading though this article you should know that LMMSE and IGV are to be used on high ISO photos and that for the majority of other cases you should stick with the default AMaZE method, but of course you are free to explore for yourself and to test each method out on your own raw files."

Cheers,
Bart

Hi,
If got moiré I do 2 RAW conversions, one AMaZE  and another with LMMSE and only use the LMMSE version for the moiré parts.
If the moiré still there I will try the default tools and this way

[Bart_van_der_Wolf]

Hi,
If got moiré I do 2 RAW conversions, one AMaZE  and another with LMMSE and only use the LMMSE version for the moiré parts.

Yes, that's one of the benefits of having a choice of algorithms in the same application.

Cheers,
Bart