Can we replicate the effect of optical anti-alias filter in digital processing?

[earlybird]

Hi,
 I was reading a debate about the merits for use of, or cancellation of, optical anti alias filters on the Canon 5DS and 5DS R cameras.

 The debate quickly devolved in to claims that persons either did or didn't understand sampling theory.

 One idea I thought I understood but realize has been more likely a presumption is the idea that if you skip the opportunity to use an "analog" anti alias filter prior to sampling that you may not recover any of the anti aliasing effect in a digital process.

 I am having second thoughts about this; why can't we use a digital process to mimic or at least closely replicate the effect of the analog filter? Please note that I am asking specifically about a comparison to the effect of the analog filter rather than considering other ideas about anti-aliasing. For example; the effect of anti aliasing through super sampling as it might apply to downsizing an image.

 These second thoughts are based on my impression that the analog filter doesn't output or introduce "reality" to the signal but merely cuts off (with varied effectiveness) the frequency information above the Nyquist limit. I can't shake the idea that the effect of an analog AA filter is no more "real" or accurate than an effect of aliasing.

 If you miss the opportunity to average out values with an analog filter before sampling how closely can you approximate the analog filtering by analyzing neighboring data points in a digital file and reversing the aliased results?

 Is it possible, impossible, or would a digital process provide some sort of choice of a third impressionistic representation of "reality".

 Thank you.

[TonyW]

I wouldn't attempt to try and answer the questions relating to the technical aspects of sampling theory - I just do not know enough, but...

I wonder if the concerns about removing AA filter has been somewhat overplayed - maybe at least good reason from the manufacturers POV ?

When I first got my D800E being concerned about the potential I set about to try and get moire and false colours without success shooting fine detail in fabrics etc.  Camera on tripod and mirror up using remote release.

I have seen one occasion of false colour on a pano shot showing a bridge across an estuary as a small part of the scene.  Easily fixed in PS by just painting over the offending colour stripes on a new layer set to colour.

Best methods to avoid:
* If seen at capture time change position a little and try again

* Stop down to introduce greater diffraction effect

* Hand hold a high MP camera ( only slightly tongue in cheek  ) the potential shake introduced should be enough to degrade IQ to the point of no moire 

Pentax have even included an AA filter simulator in some of their cameras
https://www.youtube.com/watch?v=82ym_GdFlcE

I am not aware of any software that would actually mimic an AA filter in camera and TBH cannot see any need for such as editing in PS or LR should resolve the issues most times with little if any loss of IQ

[sandymc]

You can replicate the effect of a low pass filter in digital processing. You can't replicate the effect of a low pass filter (or any other class of filter) acting as an anti-alias filter in digital processing because by the time the signal reached the digital filter, the aliasing will already have taken place.

[GrahamBy]

Pixels imply sampling at some spatial frequency. Anything above half that frequency will be "reflected" around the sampling frequency and become part of the signal.

Once it's part of the signal... it's part of the signal. There is no way to tell whether it was put there by the thing being photographed, or via aliasing, so unless the digital processing has god's phone number, it can't possibly fix it.

Maybe it's easier to understand in audio, where we think more naturally in frequency terms. Suppose you are using an AD converter sampling at 48kHz, without an AA filter. In that case, a tone arriving at 25kHz will be encoded in exactly the same way as a tone at 23kHz. So there is nothing to be done... In practice, you either hope there is no >24kHz sound present, or you apply an analogue AA filter to remove it.

Of course what happened in digital audio was that AD technology improved, and it was possible to sample at (for eg) 192kHz and you only needed to worry about sounds above 170kHz: those between 96kHz and 170kHz would be encoded between 96kHz and 22kHz, and could be filtered in the digital domain. The important difference is that ears only work up to 16-20kHz... whereas in spatial sampling, you can enlarge the image change the spatial frequency. The audio equivalent of pixel peeping would be to slow the play-back speed, but so far as I'm aware no one has been nuts enough to propose this...

So... yes you can make use of the higher res of a 50Mpx sensor to do digital AA, provided you're happy to get a de-aliased version of a lower-resolution. Ie you could remove the Moiré by down-sizing your file to be equivalent to that from a 12Mpx sensor... but no one wants to do that.

Btw, this essential information issue is why there can never be a perfect de-noising or sharpening technique: necessarily if you photograph something that looks exactly like noise, and apply de-noise, it will be removed. If you photograph something with slightly fuzzy edges and sharpen, the fuzziness will disappear. Those techniques only work because you know that what you photographed, and that it didn't look like noise or have fuzzy edges. If you apply them automatically, you're asking the software to guess what you photographed. In the case of case of digital AA, you would have to know that you hadn't photographed something that actually looked like a Moiré pattern... or be happy to live with a lower resolution that averages it away. It's a cake vs full stomach issue.

[earlybird]

Thank you suggesting for the ideas to consider.

I now realize that this morning I was disregarding the reflection of the frequencies past the cutoff when no filter is used. I was preoccupied with thinking about the spectrum between the filter's cutoff and the Nyquist limit and not considering how the the range effected by reflection may be much wider.