[falconeye]

I was under the impression AA filters were made of two superimposed layers of birefringent material, typically Lithium Niobate or Magnesium Flouride - the birefringence was used to introduce a slight blurring of the image.

That's correct but you seem to ignore what "birefringent" means: beam-splitting.

(calcite, from the wikipedia article)
The so-called "low-pass filter" or "AA filter" in cameras with a Bayer filter array is no low pass filter. It creates a non-blurry double image and has an infinite frequency response.

The low pass filter is the micro lens array and finite size of a pixel's photo-sensitive area. Characterized by the "fill factor". With fill factors close to 100%, birefringent crystals shift the low pass frequency down.

I mean, it seems to me (without understanding what nyquist frequency is) that we can predict moire.?
...
would that mean to go closer or to go further away?

You're right, it is deterministic.
(btw, nyquist frequency is just pixels per mm as far as this discussion is concerned; a bit more accurate would be 0.5 cycles per pixel = 1 line pair per pixel pair (*)).

But your approach is too complicated. Nobody computes such things in practice. You need a proper understanding to guide your intuition, that's it.

E.g., when I said, change the distance, I meant it. Nobody cares if it is closer or farer away. Just change something to change the pattern magnification (read: frequency). You could also change the zoom a bit.

One problem is you don't see it in LV (LV uses a different sampling frequency), you have to zoom into the final image to check for it.


___
(*) if you want to compute it ... 1 line pair per pixel pair with the D800 is 1 line pair 10µm / Magnification (M). If you shoot a half portrait of 1 m height, M is 36mm/1000mm=0.036. So, a fabric with thread distance of 0.28mm could become a problem. That's an extremely fine texture though, almost invisible to the naked eye. E.g., silk satin has such a fine structure and most people would think it is flat. This is why most moiré patterns emerge when nobody expects them ... Go closer and the threads are resolved with no artefact. Go farer and the silk becomes flat. Note that in the silk example, moiré would be mild, because the silk threads all have the same color. A bigger problem are alternating color patterns, like stockings or patterned fabrics. But you have to be a lot farer away to hit the Nyquist frequency then.

---

Last edited by falconeye; 04-02-2012 at 04:04 AM.

[Digitalis]

That's correct but you seem to ignore what "birefringent" means: beam-splitting.

As a former jeweller i'm am aware of what birefringence is, Sapphires, Rubies, Quartz and Cubic zirconia all exhibit this optical property. Though it is perhaps my understanding of the low pass filter and its application in digital cameras is in need of correction, I assumed the purpose of the "AA filter/low pass filter" is used as a beam splitter to introduce diffusion to the airy disk and spread the light over a greater area of the bayer filter - reducing the potential for aliasing with resolved patterns that are at the nyquist limit. Diffraction works in a similar way albeit the lens itself is the filter (and fundamentally that is exactly what lenses are). However the low pass filters are always in use regardless of image content therefore the entire image will still be slightly un-sharp because if the filtering - and that is why it is so important to sharpen the output on cameras with a LPF to re-construct the contrast lost on the fine details that the low pass filter blurred.

[D4rknezz]

Thanks, falcon.

So again, trying to catch up with what you said, In studio, tethering should help alleviate the problem - whats more, in general studio pictures will usually be close (as you said, magnification of a portrait of 1M height (which is ..waist height). I'd say typical close composition would probably be of a picture of 2-3 meter (do i just multiple with M? for the D800 sensor? so M of 0.072), even though its seldom of material as fine as silk. )

In weddings, the problem could probably rear its head in group pictures. However, given the colours of wedding (white dresses and black tux), it may not be as bad as ...aunt irma in purple in the background. In the case of wedding, you can only spray and pray - since you can't see it in viewfinder.

Once you output the final images - most people will never see a 40MP file at 100%, so they'd see it at the 10% size. Which means the problem will probably seldom show up in their monitor?

And , if I am worried, shoot wide open or shoot stopped down to introduce diffraction. basically, avoid f/8 to be there

That's correct but you seem to ignore what "birefringent" means: beam-splitting.

(calcite, from the wikipedia article)
The so-called "low-pass filter" or "AA filter" in cameras with a Bayer filter array is no low pass filter. It creates a non-blurry double image and has an infinite frequency response.

The low pass filter is the micro lens array and finite size of a pixel's photo-sensitive area. Characterized by the "fill factor". With fill factors close to 100%, birefringent crystals shift the low pass frequency down.


You're right, it is deterministic.
(btw, nyquist frequency is just pixels per mm as far as this discussion is concerned; a bit more accurate would be 0.5 cycles per pixel = 1 line pair per pixel pair (*)).

But your approach is too complicated. Nobody computes such things in practice. You need a proper understanding to guide your intuition, that's it.

E.g., when I said, change the distance, I meant it. Nobody cares if it is closer or farer away. Just change something to change the pattern magnification (read: frequency). You could also change the zoom a bit.

One problem is you don't see it in LV (LV uses a different sampling frequency), you have to zoom into the final image to check for it.


___
(*) if you want to compute it ... 1 line pair per pixel pair with the D800 is 1 line pair 10µm / Magnification (M). If you shoot a half portrait of 1 m height, M is 36mm/1000mm=0.036. So, a fabric with thread distance of 0.28mm could become a problem. That's an extremely fine texture though, almost invisible to the naked eye. E.g., silk satin has such a fine structure and most people would think it is flat. This is why most moiré patterns emerge when nobody expects them ... Go closer and the threads are resolved with no artefact. Go farer and the silk becomes flat. Note that in the silk example, moiré would be mild, because the silk threads all have the same color. A bigger problem are alternating color patterns, like stockings or patterned fabrics. But you have to be a lot farer away to hit the Nyquist frequency then.

[falconeye]

I assumed the purpose of the "AA filter/low pass filter" is used as a beam splitter to introduce diffusion to the airy disk

Airy disks are much smaller than the pixel distance (normally). Beam splitting works because the spatial extent of a pixel itself acts as a low pass filter in the first place.

Because you're so knowledgeable, I just wanted to make sure you understand that what is called "low pass filter" (also by the vendors) technically isn't a low pass filter. But combined with the pixel structure, it spreads the spatial uncertainty from 1 pixel surface to 4 pixels surface (or less for a weaker "AA" filter).

So, we're splitting hairs here.

But is has an implication: What everybody calls Moiré pattern actually isn't one. Moiré is already killed by the large fill factor. It is a low frequency pattern of color-detection errors created during by the demosaicing algorithm. This is why I often call it "false color moiré". A better (entropy-based, non-local) demosaicing algorithm could actually avoid it and false specular highlights would be the only remaining problem. If you look at a raw file prior to demosaicing (which some raw converters output) then you shouldn't see Moiré for a gray input and proper rgb channel calibration (that's one way to deal with it: compare the frequency responses before and after demosaicing).

---

Last edited by falconeye; 04-02-2012 at 05:48 AM.

[Digitalis]

Airy disks are much smaller than the pixel distance (normally). Beam splitting works because the spatial extent of a pixel itself acts as a low pass filter in the first place. Because you're so knowledgeable, I just wanted to make sure you understand that what is called "low pass filter" (also by the vendors) technically isn't a low pass filter. But combined with the pixel structure, it spreads the spatial uncertainty from 1 pixel surface to 4 pixels surface (or less for a weaker "AA" filter).

Now I got it Falk, thanks for the explanation. Now at least I will have a precise explanation on the "low pass filter" to share with my more inquisitive students.

[falconeye]

I think some of the compression options in Nikon's NEF format are closer to the idea.

I found this source:
-> Noise, Dynamic Range and Bit Depth in Digital SLRs -- page 3

I.e., the 12 Bit NEF lossy compressed file format is already very close to my idea. However, Nikon forgot the rationale with the 14 Bit NEF lossy compressed format (they store additional photon shot noise rather than only additional shadow detail).

Worse though, Nikon probably forgets to restore the noise upon loading raw images
-> Jeffrey Friedl's Blog
so that the effect can become visible (lack of dithering smoothing in the highlights). However, Jeffrey Friedl didn't look to well at that particular issue. So, maybe, Nikon adds noise. A histogram inspection might reveal it but I could find no source of such a study.

See here -> Noise, Dynamic Range and Bit Depth in Digital SLRs -- page 3
why it matters to restore noise when loading the raw file. And how even then resizing/blurring can make the compression visible (Fig. 17a). Which is why I want to include a downsampled but uncompressed 16 bit version in my raw file format (it would be used to slightly adjust the expectation value for the random generation at each pixel). This is necessary because even if a value may be random, its expectation value (i.e., when taking enough samples) may not.

[poetictrance]

Is there a 645D2 in the works? Or better yet, a 67D?

[Anvh]

I hope there is a 645D2 on the way, the sensor is fine but a faster processor would do wonders for studio photography.
67D would only be a dream, such a large sensor are very expensive and i doubt actually someone makes them so they need to be specially make...

[jogiba]

Is there a 645D2 in the works? Or better yet, a 67D?

Red has a 261mp 168x56mm sensor on it's roadmap.

[Digitalis]

Red has a 261mp 168x56mm sensor on it's roadmap.

oh good , bill gates could buy one - too bad for everyone else.

you would have a fun time finding lenses to suit a sensor with that size and resolution - you would be looking at large format Schneider APO or Rodenstock APO lenses, many of which are f/5.6 and slower and there isn't a wide variety of focal lengths that you have the option of using with 35mm systems. 90mm lenses are popular on 6X17 format cameras but there aren't many that are suited for digital sensors.

the D800(e) is really showing the weaknesses of 35mm optics - I have had the opportunity to use a D800, the Nikkor 58mm f/1.2 ASPH didn't do all that great on it, even when I stopped it down. But then again Nikon's fast 50's have always been behind the pack, It would be interesting to try a 50mm f/1.4 takumar on the D800. The Nippon kogaku 105mm f/2.5 actually did surprisingly well, but the jewel was the Voigtlander 125mm f/2.5 APO Lanthar which offered absolutely stunning performance on the D800 - If I was in the market for a High MP camera body, I would get a D800E and use the 125mm f/2.5 Lanthar - which would be a combination that would give the Pentax 645D a real run for its money.