D800E and moire - calling all studio shooters

[kers]

Hello Simon, yes the Nikon 24-85mm VR seems pretty good. A lot of clarity in the images..
I used - like Bill - the 28-105 till the d800e came. Then i thought it was not good enough anymore...
There was no aperture that could make a complete sharp image.
But what i liked about the 28-105 was the macro function that is missing on current zooms.
About raw converters - it is interesting to see that moiré is not only a pixel problem but just as well an interpretation problem.

[LKaven]

I wouldn't call that fixed. The false colors are gone, but the luminance moiré remains.

I'd like to suggest that the moiré in this scene is actually in the scene, and produced by the interaction of the venetian blinds with a mesh screen.  The moiré in this case is not produced by an interaction with the sensor so much, but is there to be seen with the naked eye, just as you'd see when looking through two mesh screens.  For that reason, I don't expect the moiré tool to be able to fix that, and I wouldn't fault the camera.

[Bart_van_der_Wolf]

I'd like to suggest that the moiré in this scene is actually in the scene, and produced by the interaction of the venetian blinds with a mesh screen.  The moiré in this case is not produced by an interaction with the sensor so much, but is there to be seen with the naked eye, just as you'd see when looking through two mesh screens.  For that reason, I don't expect the moiré tool to be able to fix that, and I wouldn't fault the camera.

Hi Luke,

It's the camera, really, and it does exactly what it was designed for. The proof is in the detail, as shown here (a crop from Simon's image):



The optics of a camera cannot render detail without a tiny bit of blur, and the demosaicing also loses a few percent of resolution. Yet the zoomed in detail shows single pixel wide horizontal lines, until the phase of the pattern in the image is exactly between two sensels which results in abrubtly no contrast at all. There is no such thing as high contrast single pixel wide detail, unless the detail is exactly the size of the sensel aperture, it is positioned exactly aligned with the sensel, and the lens is (not just theoretically) perfect (= no diffraction and no residual lens aberrations whatsoever).

If it were scene detail that is already aliased because of two overlayed screens, then that wouldn't produce single pixel wide lines, and the angle of the single pixel lines would probably not be as horizontal.

To take away any doubt for the disbelievers, it could help if Simon were able to show a closer view of the detail so we would know what it is that we are looking at.

Mind you, this will only occur in the plane of focus, and with relatively little diffraction, and a decent lens to begin with. This is why it wil not always show, but sometimes it can be a pain for certain subject matter. When prepared for the occasional occurrence of moiré, one could take 2 shots, one regular and one at f/18, and then in postprocessing replace the moiré affected parts with the diffraction smoothed and deconvolution sharpened parts. An even more effective AA-blur would come from a tiny bit of defocus (although we'd have to guess how much is enough), but that could also change the size of the image a bit if the subject is close, making it a bit harder to make a composite.

A shot at f/18 will effectively kill all relevant red and green wavelength resolution (even for the highest subject contrast) beyond the Nyquist frequency of the D800E, so aliasing cannot occur anymore. The Nyquist frequency of the D800(E) sensor is given by its sensel pitch, assumed to be approx. 4.88 micron, which equals 102.459 cycles/mm. Diffraction from a circular aperture for green light (the most important contributor to luminance) will already reduce the modulation transfer at 98 cy/mm to virtually zero. So there is not modulation left at the limit of what the D800E sensor could resolve. Cameras with different sensel pitches will require a different F-stop to totally prevent aliasing, e.g. a 6.4 micron sensor without AA-filter would require f/22.

Cheers,
Bart

[hjulenissen]

One of the interesting aspects of aliasing is that it is non-linear: a frequency of fs/2 will "wrap around" to a frequency of 0. In other words, heavily modulated periodic structures that happen to oscillate at 0.5 linepair/pixel (is this the right terminology?) will result in a sampled signal that has a frequency of zero (depending on phase) - i.e. a constant. In the real world, stuff is never precisely 0.500000..., so you may get something like a rooftop where the tiles are alternating between "black" and "white", where the image file renders the roof as either pure black, pure white, or some non-realistic slow modulation between those two extremes. I can not imageine how a raw developer/plugin could ever detect and suppress such errors without introducing significant errors for desired scene elements. In a properly anti-aliased camera, one would expect such a roof to be rendered as "gray", i.e. the camera/lense works more or less like a linear, space-invariant lowpass filter.

The degree to which this happens (and becomes visible) depends on the scene, and on the degree of prefiltering. Even a D800E, no matter how well focused, expensive optics and heavy stand is used will have some filtering. If nothing else, the spatial extent of the sensel site and micro-filter will "smear" detail and reduce aliasing compared to a hypothetical non-filtered point-sampler. It would be interesting to see examples of such behaviour.

-h

[crames]

Simon, that's a sharp, vivid shot.

About the color and luminance moire defects, it appears that it is purely a software issue, and despite what the AA filter advocates are saying, an AA filter probably would not have made any difference at all in this shot. The frequency of the blind pattern is about 15% below the Nyquist frequency, which means that the blind pattern is actually larger than what an AA filter is designed to remove. So the the blind pattern would not have been removed by the AA filter in the D800, either.

The color moire is due to the way the raw converter is interpreting the raw data. A different converter (or demosaicking method) will give different results and possibly eliminate the color moire, it not there are the usual, effective tools for removing it.

The luminance moire is another matter entirely. Some valid, un-aliased high-frequency information just cannot look right without some interpolation. This is because, in theory, digital images need a reconstruction filter to turn the individual pixels/samples back to the continuous, analog domain. Digital images often get away without the reconstruction filter step because they don't have a lot of very high frequency detail. The D800E with a good lens can certainly capture a lot of the highest image detail, as you have shown, and can easily produce images that need a little reconstruction filtering.

Here is the proof.

The first image is a 100% crop of one of the windows in your image.



The second image is the Fourier transform of the crop, with the bright spots showing the location of the blind frequencies, about 15% from the edge (where the Nyquist frequencies are).



The third image is one channel of the crop, which has been interpolated (2X zoom). Note the clean blind pattern.



The final image has all three channels interpolated. The luminance moire is gone. The color moire remains because interpolation doesn't help that - it's been baked into the file by the raw converter and can be removed by conventional means..



Cliff

[MichaelEzra]

Cliff, I am curious, which interpolation method was used for 2x zoom?

[crames]

I used a pure sinc filter (done by FFT like described here).

A good Lanczos filter should give similar results. The key is that the filter has good response up to the highest frequencies.

[Bart_van_der_Wolf]

The luminance moire is another matter entirely. Some valid, un-aliased high-frequency information just cannot look right without some interpolation. This is because, in theory, digital images need a reconstruction filter to turn the individual pixels/samples back to the continuous, analog domain. Digital images often get away without the reconstruction filter step because they don't have a lot of very high frequency detail. The D800E with a good lens can certainly capture a lot of the highest image detail, as you have shown, and can easily produce images that need a little reconstruction filtering.

Hi Cliff,

That's a good catch by you, again. I was wondering why there were not more obvious luminance aliasing artifacts caused by signal 'folding back' below the Nyquist frequency. As it turns out there was not enough luminance signal to do that because the most significant detail was just below Nyquist, and resolved with enough amplitude. The color demosaicing on the other hand is based on a lower Nyquist limit, and that does mirror back below Nyquist, causing aliases with lower spatial frequencies.

Thanks for spotting that.

Cheers,
Bart

[MichaelEzra]

I used a pure sinc filter (done by FFT like described here).

A good Lanczos filter should give similar results. The key is that the filter has good response up to the highest frequencies.

Thanks. What do you recommend as canvas color in the step "Increase the canvas size on all sides."?
By any chance, would you know any Lanczos filter plugin for photoshop?

[Rand47]

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.

Lightroom moire brush:

[hjulenissen]

Thanks for some interesting posts crames.

-h

[crames]

Michael - use black (0) as the canvas extension color. Feel free to experiment, though.

I don't know of a Lanczos plugin for Photoshop  - somebody needs to create one!

[hjulenissen]

I don't know of a Lanczos plugin for Photoshop   - somebody needs to create one!

(open source) Image magick contains one of the most extensive sets of linear scaling operators that I have seen:
http://www.imagemagick.org/Usage/filter/nicolas/

Board index ‹ ImageMagick ‹ Users < best downsampling method for DSLR photographs


-h

[Bart_van_der_Wolf]

Thanks. What do you recommend as canvas color in the step "Increase the canvas size on all sides."?
By any chance, would you know any Lanczos filter plugin for photoshop?

Hi Michael,

Alternatively you can use ImageJ to do the FFT conversions, and just fill the canvas space added to the FTT conversion with zeroes (center the Real and Imaginary image data, then add canvas to e.g. 2x the original pixel dimensions).
I prefer using an FFT plugin for ImageJ to its built-in FFT functionality. The plug-in allows arbitrary image sizes and multiple threads for the processing, to name a few useful features.

Straight resizing by padding the larger canvas with zeroes does create some visible ringing artifacts to edges.

Cheers,
Bart

[bjanes]

Lightroom moire brush:

I don't see any moire in your reprocessed image, but one should really use a 1:1 view to check for moire. With down sampling, moire can be downsampled out of existence or, alternately, created where it did not previously exist. On one of my computers, the OP's preview image appeared solid gray and the moire was only apparent when viewing the full size image.

As Bart has pointed out, improper downsizing can introduce moire where it was not previously present and low pass prefiltering may be necessary to prevent introduction of moire.'

Regards,

Bill

[crames]

(open source) Image magick contains one of the most extensive sets of linear scaling operators that I have seen:
http://www.imagemagick.org/Usage/filter/nicolas/

Board index ‹ ImageMagick ‹ Users < best downsampling method for DSLR photographs


-h

Hi h,

I was aware of imagemagick and its reputation for quality results, but it's not a PS plugin and I have never managed to sit down and plow through all the options. It is certainly a smorgasbord of lanczos and other filters.

Do you have a script that might work well with Simon's image?

[Bart_van_der_Wolf]

Do you have a script that might work well with Simon's image?

Hi Cliff,

When you use a Windows system, you can add a batch file to your "Send to" files so the command can be invoked by a right-mouse-click. The batch file could contain something like this:

convert %1 -filter LanczosRadius -define filter:lobes=8 -distort Resize 200%% "%~dpn1_200pct.png"

Otherwise you can use a Command line interface with a command like:

convert {input} -filter LanczosRadius -define filter:lobes=8 -distort Resize 200%% {output}

where {input} and {output} are replaced by the appropriate path/filename.ext .

In the examples above I've used the most recent (LanczosRadius) filter settings that are not available in older versions of ImageMagick, and I've added an option to vary the number of lobes which controls the amount of sharpening/ringing. It is generally not advised to use Lanczos for upsampling (because of the ringing artifacts), but if we want to reconstruct the original signal as good as possible then that's the side-effect (just like with simple FFT upsampling). With a 300% upsampling the result looks even better. Reducing the number of lobes to 3 will reduce the ringing but also reduce the sharpness.

For general upsampling the "-filter LanczosRadius -define filter:lobes=8" part is left out, which will then be replaced by the default filtering that produces a trade-off between various artifacts, but is not as effective in suppressing the artifacts in this example image.
For less specific images that require 'artifact free' upsampling, I get very good upsampling results with the following batch file entry:

convert %1 -set colorspace sRGB -colorspace RGB +sigmoidal-contrast 12.09375 -filter Robidoux -distort Resize 200%% -sigmoidal-contrast 12.09375 -colorspace sRGB -compress None "%~dpn1_200pct-CSpc+sigmoid.tif"

It can also do good downsampling, in this example to an uncompressed TIF format, with modified resize percentages below 100%.

I've used the slower but more accurate (-distort Resize) EWA (cylindrical) version of resizing, IMHO far superior to the traditional (-resize) 2 orthogonal tensor method.

Cheers,
Bart

[crames]

Thanks for the info, Bart.

The 8-lobe Lanczos command does give results very similar to my 2X zoom (FFT) example above. It is actually a little less ripply.

The 3-lobe and your general purpose one don't do the job on this image, however.

We have to come up with some ways to reduce the rippling, so D800e users can get the most from their cameras and in an easy way!

attached: the 8-lobe Lanczos version.

[hjulenissen]

The 8-lobe Lanczos command does give results very similar to my 2X zoom (FFT) example above. It is actually a little less ripply.
...
We have to come up with some ways to reduce the rippling, so D800e users can get the most from their cameras and in an easy way!

I am not surprised that 8-lobe lanczos is visually similar to FFT-based scaling.

For linear filtering, there is a relationship between frequency response (wide & flat passband, narrow transition band and lots of attenuation in the stop-band) and ringing. In order to approach a perfect lowpass filter (sin(x)/x), you need to have lots of filter coefficients, some negative and some positive. This means that an image edge will spread out in space and cause undesirable artifacts.

The common (linear) solution to this problem is to either sacrifice passband (blurriness) or stop-band (aliasing). A more advanced solution might be adaptive/non-linear processing.

Did you consider DCT-based scaling rather that FFT-based? I believe that the implicit periodic extension built into DCT might be more suited than the periodicity built into DFT (FFT).

-h

[Bart_van_der_Wolf]

Thanks for the info, Bart.

The 8-lobe Lanczos command does give results very similar to my 2X zoom (FFT) example above. It is actually a little less ripply.

Hi Cliff,

You're welcome. That's probably due to the Lanczos windowing. More lobes will steepen the first ripple amplitude even more, and reduce the other ripples.

The 3-lobe and your general purpose one don't do the job on this image, however.

We have to come up with some ways to reduce the rippling, so D800e users can get the most from their cameras and in an easy way!

I'm not sure whether both constraints, high amplitude single pixel edges and few artifacts, can be satisfied at the same time in a universal approach. There are also things like diagonal lines that should not go jagged. There can also be a slightly different response between JPEG images and other non-wave based sources. We'll have to see what can be done with some of the other filters.

Cheers,
Bart