D800E and moire - calling all studio shooters

[NicolasRobidoux]

... However, with upsampling it has an IMO nasty side-effect of exaggerating edge sharpness in an ugly way. I say nasty because we have no control over it, unlike post resampling sharpening...

This agrees with http://www.imagemagick.org/Usage/filter/nicolas.
Note that the next update of these recommendations will specify that filters without a significant negative lobe (like Triangle = bilinear, Quadratic = quadratic B-spline smoothing and Spline = cubic B-spline smoothing, in their tensor or EWA versions) actually give fine results upsampling in linear light.
Basically, the nice "physical" properties of linear light partially go out the window when you use filters with significant negative lobes (which can lead to negative light and "whiter than white" light, which is more acceptable from a physical viewpoint but still iffy).

[NicolasRobidoux]

...On the other hand, when we want to reduce the risk of running into issues with limited bit depth images being converted into linear gamma and back, we might as well (QED !) skip the gamma and sigmoid conversions and just use (deconvolution) sharpening to control the edge sharpness...

Provided conversion between colorspaces is sufficiently accurate (and, between linear RGB and sRGB, it is sufficiently in ImageMagick unless you use an 8-bit version), it is not the input or output bit depth that matters, it's what happens within the processing.
-----
Right now, ImageMagick has a split personality w.r.t. sigmoidal-contrast: it uses a clamping LUT in IM6, and a full double precision, but still clamping, version in IM7. When I have time (and money in the bank), I'll reconcile the two and remove the clamping, which should help a little when using an HDRI version of ImageMagick.
I had to rewrite the whole thing so that sigmoidization works properly, and unfortunately did not quite finish the job.

[hjulenissen]

Provided conversion between colorspaces is sufficiently accurate (and, between linear RGB and sRGB, it is sufficiently in ImageMagick unless you use an 8-bit version), it is not the input or output bit depth that matters, it's what happens within the processing.
-----
Right now, ImageMagick has a split personality w.r.t. sigmoidal-contrast: it uses a clamping LUT in IM6, and a full double precision, but still clamping, version in IM7. When I have time (and money in the bank), I'll reconcile the two and remove the clamping, which should help a little when using an HDRI version of ImageMagick.
I had to rewrite the whole thing so that sigmoidization work properly, and unfortunately did not quite finish the job.

For testing new ideas, and for passing back and forth algorithms/results for comments and tweaking, I much prefer something like MATLAB: it is double-precision float (pretty much by design), it is often "fast enough" (using libraries like FFTW, BLAS etc), and operations like scaling can be implemented with a fraction of the effort it is to do anything at all in C.

If one is annoyed by the proprietary (and expensive) licensing, there are good open-source alternatives (like NumPy). I would never attempt anything like implementing deconvolution in C without having messed around with algorithm in something like MATLAB. YMMV, and other people may be more skilled in prototyping in C than myself.


If your mission is to provide the world with "the best scaler that is humanly possible" with little regard to cpu cycles, user interface etc, people like myself might be willing to install python and live with quirky command-lines if that means faster development cycles.

Just my 2c.

-h

[KevinA]

That can be fixed in Photoshop quickly. Healthy dosage of GB on both A and B channels in LAB should fix it in less than 20 seconds.

I've yet to see any software fix for moire that removes moire, it might disguise it with a certain degree of success, chances are it leaves luminance banding behind.
Kevin.

[Bart_van_der_Wolf]

I've yet to see any software fix for moire that removes moire, it might disguise it with a certain degree of success, chances are it leaves luminance banding behind.
Kevin.

Yes, aliasing artifacts are by definition permanently mixed with real detail, and as such cannot be separated afterwards. As Cliff has pointed out, what may look as aliasing, may sometimes still be shown as real detail, but we will need to upsample and reconstruct some of the detail by amplifying the correct spatial frequencies to do that. The ringing artifacts must mimick the real detail.

The false color artifacts are related to aliasing artifacts, by the difference in sampling density of the Red and Blue filtered sensels versus the Green filtered ones, and the particular way the Raw converter tries to reconstruct color from an insufficient (and partially more aliased) amount of data. Removing the color will at least mask the more eye catching colorful aberrations. Suppressing real luminance artifacts in the plane of focus is much harder.

Cheers,
Bart

[hjulenissen]

Yes, aliasing artifacts are by definition permanently mixed with real detail, and as such cannot be separated afterwards.

There is no general method to separate aliasing and signal once they are overlapping if we make no assumptions about the true signal and the aliasing.

I believe there to also be no general method to separate signal and noise if we make no assumptions about the true signal and the true noise. The thing is, we can often make assumptions about signal and noise that allows us to do noise-reduction that really reduce the noise while not affecting the signal all that much. Many viewers feel that noise reduction can improve noisy images.

-h

[Bart_van_der_Wolf]

There is no general method to separate aliasing and signal once they are overlapping if we make no assumptions about the true signal and the aliasing.

I believe there to also be no general method to separate signal and noise if we make no assumptions about the true signal and the true noise. The thing is, we can often make assumptions about signal and noise that allows us to do noise-reduction that really reduce the noise while not affecting the signal all that much. Many viewers feel that noise reduction can improve noisy images.

That's correct, for single images. Of course when we shoot multiple images of a stationary subject with sub-pixel offsets, then we can resort to things like super-resolution which will allow to increase the Nyquist frequency and thus shift the aliasing threshold. Likewise for S/N ratios we can use HDR exposure stacking, thus cherry picking from the lower shadow SNR levels of the longer exposures, or use exposure averaging which will reduce the noise threshold by the square root of the number of exposures.

Cheers,
Bart

[hjulenissen]

That's correct, for single images. Of course when we shoot multiple images of a stationary subject with sub-pixel offsets, then we can resort to things like super-resolution which will allow to increase the Nyquist frequency and thus shift the aliasing threshold. Likewise for S/N ratios we can use HDR exposure stacking, thus cherry picking from the lower shadow SNR levels of the longer exposures, or use exposure averaging which will reduce the noise threshold by the square root of the number of exposures.

Cheers,
Bart

Sure.

My point was that while there is no general method to remove aliasing in a single picture (and I am convinced that it might be proven that no such general method can ever be found), one can imagine non-general methods that exploit assumptions about typical images (or classes of images) and/or the camera sampling process to reduce aliasing without significantly affecting the desired signal, all in a perceptually pleasing manner.

I say this despite my education in signal processing telling me to always properly pre-filter and post-filter a discrete sampling process in the way suggested by Nyquist and Shannon some 90 and 60 years ago. Following the sampling theoreme takes all of the guess-work, statistics and perceptual complexity out, and you are left with a (theoretically) manageable linear filtering problem.

Interestingly, the sampling theoreme does not say that this filter has to be a lowpass filter, only that the bandwidth must be limited. As such, you could sample a scene by taking multiple exposures using an ideal sampler (approximated by a sensor where each sensel has small coverage, approaching a point-sampler), prefiltering using a set of bandpass filters at successively increasing passband (first [0...fs/2], then [fs/2...fs],...). The successive images would be aliased down to the baseband, but as the prefilter removes any non-aliased signal in that baseband, the aliasing components are uniquely resolvable. And a high-resolution image could be synthesized from the set. All according to a theoretical view, not considering the practical difficulty of actually doing it.

-h