AA-filtering CCD and CMOS

[georgl]

The fact is that without real scientific measurements under controlled conditions we cannot say for sure what's actually "better" or "worse" - we not only compare filtered vs. unfiltered images, we compare different processing, different sensors, different lenses and slightly different shooting conditions (like the precise alignment of the sensor grid and photographic detail). I was surprised to see so unsharp unfiltered samples from the 5D which really show barely any more detail than the filtered version - it seems that the lens/focusing limited the IQ already.

The Mamiya ZD was available with and without OLPF - it would be the best "measurement" system we could get hands on.

What we see in these non-scientific "comparisons" are superior (besides absolute image size) MFDB-results. It's not just the AA-filter it's various aspects. But these MFDBs are used for the most demanding photographic situations for a good reason. In the real world, the artifacts which are cannot be handled by post-processing in comparison to the artifacts still left in AA-filtered images are usually negligible.  But every AA-filter reduces contrast below Nyqist and therefore information - that's a scientific fact. All of our bayer-based imaging systems are far from being perfect, most of the image is interpolated! We're talking about poor compromises on both sides anyway.

@Daniel Browning
I'm sorry that I chose my words imprecisely - of course moire or alaising cannot be removed, just as sharpening cannot retain any information. But it can be supressed and it seems that those processed images usually still contain more "real information" than AA-filtered images.
As I mentioned, the Mamiya ZD was available with an OLPF and all MFDBs contain highly specialized technologies (microlenses, cover glasses, sensor technology) - AA-filters aren't different. They chose not to use AA-filters and Pros (and most of them know AA-filtered systems) seem to be happy about that.

[NikoJorj]

Same problem occurs with digital audio !

Sir yes Sir, but with a caveat (at least to my taste) : I'd think human perception is (much) more sensitive to high frequencies in sound, than to those in images.
Eg, the MP3 compression (that among others discards those harmonics we're not supposed to hear at, or am I wrong?) is more evident than the JPEG one.

[Plekto]

http://www.dpreview.com/reviews/sigmadp2/page22.asp

Count the number of lines at one end of the DP2 resolution pattern and compare it to the other end, for any orientation.

Here is what they do with each image:

    *  Load RAW file into Adobe Camera RAW (Auto mode disabled)
    * Set Sharpness to 0 (all other settings default)
    * Open file to Photoshop
    * Apply a Unsharp mask: 80%, Radius 1.0, Threshold 0
    * Save as a TIFF (for cropping) and as a JPEG quality 11 for download

And their comment about that test:
***
Because Adobe Camera Raw doesn't currently render the same level of detail as resolved by the camera's JPEG engine, we've taken the unusual step of using Phase One's Capture One. Again we've minimized sharpening and applied the same unsharp mask in Photoshop.
***

So basically they mangle it by shooting in JPEG format, tweaking and massaging it, and then converting it *again* to another JPEG.  The test is worthless to compare what we're looking for here.

[BernardLanguillier]

The fact is that without real scientific measurements under controlled conditions we cannot say for sure what's actually "better" or "worse" - we not only compare filtered vs. unfiltered images, we compare different processing, different sensors, different lenses and slightly different shooting conditions (like the precise alignment of the sensor grid and photographic detail). I was surprised to see so unsharp unfiltered samples from the 5D which really show barely any more detail than the filtered version - it seems that the lens/focusing limited the IQ already.

Another possibility being that, since DSLR sensors are designed and optimized with the presence of the AA filter in mind, removing it simply turns a highly tuned device into a non optimized one?

Cheers,
Bernard

[bjanes]

The DPReview report on the DP2 provides a good example of the increased "resolution" provided by the lack of an AA filter:

http://www.dpreview.com/reviews/sigmadp2/page22.asp

Count the number of lines at one end of the DP2 resolution pattern and compare it to the other end, for any orientation.  The pattern looks nice and sharp, and if you were just to look at the pattern from say 22 to 24 on the chart you would say that the camera is resolving all the way out to the limit of the chart.  But actually it is just aliasing a more finely spaced set of lines into a more coarsely spaced set of lines; it is not actually resolving at that level.  Aliasing definitely helps the accutance though  

An excellent observation by Emil. The resolution of the DP2 is 1760 pixels/picture height. The Nyquist frequency is 880 lp/ph, whereas DPReview is reporting an "absolute" resolution of 1400 to 1500 lp/ph, which is preposterous!

880 lp/ph would be at 8 on their scale.

The conclusion is that aliasing works wonders on a resolution chart.

[BernardLanguillier]

Bumble bees can fly, though aeronautical engineers used to say that it was impossible. They just didn't bother asking the bees. And yes, removing the AA filter on a properly designed system does increase both real and apparent resolution. Talk to the engineers that design these systems, and also just trust your own eyes.

Michael,

When was the last time you spoke with a sensor R&D chief engineer at Canon, Nikon or Sony? Wouldn't be interesting to hear about their motivation for equiping their sensors with AA filters?

If you are interested, I might be able to arrange an interview with one of these people. Cannot commit anything but I would try.

Please let me know.

Cheers,
Bernard

[Daniel Browning]

As I mentioned above, two people can look at the exact same image and get a totally different perception because they have very different tastes. It applies to everything in photography, like saturation. One will look at a certain red tone in a Velvia image and see it as a fake, unnatural-looking color. The other will see it as the ideal pleasing tone. One will look at a certain detail of the image and see it as a fake, unnatural-looking aliasing artifact. The other will see it as the ideal microcontrast. So the very things that are pleasing to one person may be the very thing that another dislikes.

Image A:



Image B:



To see the differences, it may help to open each image in a different browser tab for direct A/B comparison.

Many would prefer the first image, describing it as sharp, crunchy, high microcontrast, with lots of fine detail, such as stubble. Others would see it as fake-looking, with harsh transitions, jagged edges, and lots of false detail, such as stubble that should be too small to see and jagged edges on the ear.

Some would prefer the second image, describing it as smooth, natural, with the appropriate amount of detail for its size. Others would see it as mushy, hazy, low contrast, and lacking in fine detail.

Which image do you prefer?

[Daniel Browning]

Doug,

Show me the artifacts:

Both photos: Leica R8 with DMR digital back (10 MP, no AA filter), 280mm f/4 APO-Telyt-R

I don't see any artifacts in 16" x 24" prints.

I think there is a difference in how you and I perceive fine details. If that's the case, then even if you did provide 100% crops, it would not do any good for me to point out what I consider to be "artifacts", because to you there are not artifacts at all, but the ideal fine detail.

To confirm whether or not we do have a difference in taste, would you consider comparing Image A and B above? Do you see lots of aliasing artifacts? Or just a high-detail vs. low-detail image?

I would also be curious to hear how Michael, George, Nicolas, and Jing Q perceive the difference between the two.

Has the fact that every medium format back since Noah has not had an AA filter escaped anyone's notice? Do those that think that all or even most images taken without an AA have aliasing and artifacting think that the countless pros and highly critical photographers who use MF backs would tolerate this if it were true? Come on!

The reality is that aliasing is only visible in "some" images, and unless you shoot fabrics for a living, it's a very small number.

Michael,

You are correct that moiré does not affect most images taken without an AA filter. It only occurs with regularly-repeating fine patterns, like fabric, which is rare for most types of photography. However, moiré is not the only aliasing artifact. My theory is that the details you like in unfiltered images (high microcontrast) are the very things that I consider aliasing artifacts. If you see good microcontrast in image A, above, then the theory is proven. On the other hand, if you see the image as plagued by terrible aliasing, then the theory is not proven.

The advantage of not having an AA is clearly visible to anyone that has a good eye. In some cases it simply jumps out at you, image after image. In most cases its visible even in small prints and on screen as a type of increased clarity, as if a veil were lifted.

I agree completely. AA filters reduce contrast severely. It's a very steep price to pay, but I'll gladly pay it to avoid aliasing, because I consider that to be an even worse problem.

And as for removing moire in software – don't be too sure it can't be done. I'm working with one company now that is likely to have something exciting in this area in the months ahead.

Thanks for the heads up, I'll look forward to it.

AA blurring is an artifact too.  Pick the one(s) you're willing to live with

Doug, that's a good way to put it. Aliasing and AA blurring are both artifacts in a sense. To me, blurring is far more acceptable than aliasing.

But every AA-filter reduces contrast below Nyqist and therefore information - that's a scientific fact.

George, I agree with you on that. Fortunately, there is a very straightforward method to get that contrast back: reduce pixel size by 30%. If you want 12 MP with no loss of contrast from the AA filter, you have two choices:

• 12 MP with AA-filter removed (results in aliasing artifacts)
• 20 WP with AA-filter kept in place (no aliasing artifacts)

That is one of the things I noticed about the 5D2, even in just an 8x10 print. Although the AA filter was still the same relative to Nyquist, it was much weaker in the absolute sense. That weaker filtration resulted in much higher contrast even in print sizes where the pixel resolution itself did not contribute (8x10).

That's why I prefer filtered photographs.

[ejmartin]

Here is what they do with each image:

    *  Load RAW file into Adobe Camera RAW (Auto mode disabled)
    * Set Sharpness to 0 (all other settings default)
    * Open file to Photoshop
    * Apply a Unsharp mask: 80%, Radius 1.0, Threshold 0
    * Save as a TIFF (for cropping) and as a JPEG quality 11 for download

And their comment about that test:
***
Because Adobe Camera Raw doesn't currently render the same level of detail as resolved by the camera's JPEG engine, we've taken the unusual step of using Phase One's Capture One. Again we've minimized sharpening and applied the same unsharp mask in Photoshop.
***

So basically they mangle it by shooting in JPEG format, tweaking and massaging it, and then converting it *again* to another JPEG.  The test is worthless to compare what we're looking for here.

That's not how I understand what is written.  It looks to me as though they used C1 rather than ACR to convert the RAW files, and then proceeding according to their usual protocol.  The jpeg resolution test is on the previous page.

[ErikKaffehr]

Hi Michael,

I have great respect for your views and writings. That said I have a couple of points to make.

The first one is that many posters on this forum are involved with imaging theory and it seems mostly that AA filtering is necessary from a theoretical viewpoint.  As an example. it's quite obvious that we get som artifacts like the spurious resolution on the Sigma DP2 in the DPReview test on systems without OLP.

As you have pointed out MF Digital backs didn't employ OLPs since Noah. The problem in my view is that it is very difficult to compare filtered and nonfiltered systems, as we normally don't have access to identical systems in two versions, one having OLP and the other not. A pair of cameras of which one is modifed by Mad Max may be a good comparison. (AFAIK MadMax only removes one of the OLP-filters as the other is bonded to the IR-filter or sensor).

That MF cameras perform better on an per pixel basis may depend on other factors, like better optics (relative to pixel size), plentyful resolution or different processing.

An intresting comparison was made by Erwin Puts, here: http://www.imx.nl/photo/leica/camera/page157/page157.html

His writing indicates that the Nikon D3X is actually resolving better than the M9 but he feels that the Leica images are more pleasant to the eye.

Best regards
Erik

I can't believe the amount of misinformation and disinformation that this topic has engendered.

Has the fact that every medium format back since Noah has not had an AA filter escaped anyone's notice? Do those that think that all or even most images taken without an AA have aliasing and artifacting think that the countless pros and highly critical photographers who use MF backs would tolerate this if it were true? Come on!

The reality is that aliasing is only visible in "some" images, and unless you shoot fabrics for a living, it's a very small number.

The advantage of not having an AA is clearly visible to anyone that has a good eye. In some cases it simply jumps out at you, image after image. In most cases its visible even in small prints and on screen as a type of increased clarity, as if a veil were lifted.

And as for sharpening compensating for having an AA, what are you thinking? Sharpening is about edge sharpness; AKA accutance. Removing an AA filter is about increasing resolution. Yes, adding USM increases apparent sharpness, but not real resolution. Two different things, though related in the real world.

Bumble bees can fly, though aeronautical engineers used to say that it was impossible. They just didn't bother asking the bees. And yes, removing the AA filter on a properly designed system does increase both real and apparent resolution. Talk to the engineers that design these systems, and also just trust your own eyes.

And as for removing moire in software – don't be too sure it can't be done. I'm working with one company now that is likely to have something exciting in this area in the months ahead.  

Michael

[telyt]

I think there is a difference in how you and I perceive fine details. If that's the case, then even if you did provide 100% crops, it would not do any good for me to point out what I consider to be "artifacts", because to you there are not artifacts at all, but the ideal fine detail.

Here are a couple of 100% crops, no sharpening applied:





Keep in mind, these are real-world photos not test photos and as such were made under less-than-ideal conditions: 1/125 sec shutter speed, moving subjects, imperfect support.

[Daniel Browning]

Here are a couple of 100% crops, no sharpening applied:

Thanks for the response, Doug. The subject, composition, blur, and light are really great; you are very talented.

Keep in mind, these are real-world photos not test photos and as such were made under less-than-ideal conditions: 1/125 sec shutter speed, moving subjects, imperfect support.

That's an excellent point. The more imperfect the conditions, the less likely there is to be aliasing. All it takes to reduce aliasing is a very slight blur, and there are so many things that can cause a very slight blur, including DOF, camera shake, slightly missed focus, subject motion, diffraction, slight lens imperfections, etc. Plus, the smaller and smaller the pixels are, the harder and harder it is to get aliasing.

I don't see much aliasing in the first photo, but in the second one I do. Keep in mind that this is my perception of the detail, I'm sure that others see it very differently. There are many hairs with the width of 1 pixel that occur on a perfectly horizontal line. A bit of stair-stepping from one horizontal line to the next. Some of the very short, bright hairs appear to me to sparkle slightly. The 1-pixel hairs that bend significantly do it on a jagged curve. There are a few hairs that are completely unattached from the head, such as the one in the top left. Some hairs are missing a few pixels half-way through the hair, so that the ends appear unattached to the base. The two-pixel hairs become irregular in shape and thickness when they pass over contrasting backgrounds, like a dark hair on a light background, so they appear to have little bumps. There appear to be tiny spots of green discoloration, though perhaps that is a normal part of the fur.

If any image has sharp transitions from dark to light in just two pixels (black, white), I think we would agree that it has such content. The only difference of opinion is in how we interpret that content. I interpret two-pixel transitions as unnatural. To me, a transition requires 3 pixels (black, gray, white) in order to have a natural appearance. To other viewers, a two-pixel transition is the sought-after ideal of microcontrast.

[jing q]

wow now that is serious pixel peeping.

what's your opinion on MFDB images then?

I don't see much aliasing in the first photo, but in the second one I do. Keep in mind that this is my perception of the detail, I'm sure that others see it very differently. There are many hairs with the width of 1 pixel that occur on a perfectly horizontal line. A bit of stair-stepping from one horizontal line to the next. Some of the very short, bright hairs appear to me to sparkle slightly. The 1-pixel hairs that bend significantly do it on a jagged curve. There are a few hairs that are completely unattached from the head, such as the one in the top left. Some hairs are missing a few pixels half-way through the hair, so that the ends appear unattached to the base. The two-pixel hairs become irregular in shape and thickness when they pass over contrasting backgrounds, like a dark hair on a light background, so they appear to have little bumps. There appear to be tiny spots of green discoloration, though perhaps that is a normal part of the fur.

If any image has sharp transitions from dark to light in just two pixels (black, white), I think we would agree that it has such content. The only difference of opinion is in how we interpret that content. I interpret two-pixel transitions as unnatural. To me, a transition requires 3 pixels (black, gray, white) in order to have a natural appearance. To other viewers, a two-pixel transition is the sought-after ideal of microcontrast.

[Daniel Browning]

wow now that is serious pixel peeping.

Thanks!  I appreciate your complement on my attention to detail; however, aliasing doesn't actually require pixel peeping to see, it affects the perception of the entire image as a whole. Detailed descriptions can help illustrate how individual artifacts contribute to that perception, and since Doug asked me to, I tried to describe some.

what's your opinion on MFDB images then?

Compared to 35mm, as far as image quality is concerned, I like the improved contrast resolution of the lenses operating at lower spatial frequencies. I also appreciate the improved SNR over the most important stops of the dynamic range, leading to less noise and deeper color (tonal resolution/gradations). I appreciate the correct implementation of metadata ISO. What I dislike the most is the aliasing.

[ErikKaffehr]

Hi!

So you can observe some aliasing in the image. I presume that this is one of the things that you can train eye/brain to see. What happens when you print the image?

I found this image on the Zeiss website demonstrating aspect of resolution/MTF/sharpening: http://www.zeiss.de/C12567A8003B8B6F/Graph...ile/Bild_10.jpg

It's an illustration to this article: http://www.smt.zeiss.com/C12567A8003B8B6F/...Kurven_2_en.pdf

I have printed that image in A2 and studied it along the article,  a recommended exercise.

See also below
http://83.177.178.241/ekr/index.php/photoa...-and-perception

Best regards
Erik

My experience is that the perceptive system can be trained to detect small differences, once you have seen them you detect quite easily.

Thanks!  I appreciate your complement on my attention to detail; however, aliasing doesn't actually require pixel peeping to see, it affects the perception of the entire image as a whole. Detailed descriptions can help illustrate how individual artifacts contribute to that perception, and since Doug asked me to, I tried to describe some.



Compared to 35mm, as far as image quality is concerned, I like the improved contrast resolution of the lenses operating at lower spatial frequencies. I also appreciate the improved SNR over the most important stops of the dynamic range, leading to less noise and deeper color (tonal resolution/gradations). I appreciate the correct implementation of metadata ISO. What I dislike the most is the aliasing.

[ErikKaffehr]

Hi!

If you draw a line from just below the nose to just below the eye you can see four regions:

1) At the nose no hair is actually resolved, this may depend on very little or thin hair or beeing slightly out of focus
2) A bit from the nose the hair is actually resolved
3) About half way there is a lot of artifacing. Many strains are discontinous
4) Just below the hair is not resolved but just goes mush. This is not likely a depth of field effect as the eye and is well resolved.

That said I'm quite impressed with your technique! The sharpness is excellent even without sharpening.

Best regards
Erik

Here are a couple of 100% crops, no sharpening applied:





Keep in mind, these are real-world photos not test photos and as such were made under less-than-ideal conditions: 1/125 sec shutter speed, moving subjects, imperfect support.

[jing q]

Thanks!  I appreciate your complement on my attention to detail; however, aliasing doesn't actually require pixel peeping to see, it affects the perception of the entire image as a whole. Detailed descriptions can help illustrate how individual artifacts contribute to that perception, and since Doug asked me to, I tried to describe some.



Compared to 35mm, as far as image quality is concerned, I like the improved contrast resolution of the lenses operating at lower spatial frequencies. I also appreciate the improved SNR over the most important stops of the dynamic range, leading to less noise and deeper color (tonal resolution/gradations). I appreciate the correct implementation of metadata ISO. What I dislike the most is the aliasing.

After all the technical talk, I still don't see your point or why aliasing is such a big issue.
Perhaps that's where the problem lies, a lot of people who use AA filterless images actually are not bothered by the aliasing?
Granted, there are times when it is visible (esp in certain highlights)
but from my personal experience no matter how much I sharpen images I still get a slightly mushy and blurred effect with DSLRs with their AA filters.
And the mushiness/blurring is MUCH more disturbing to me than any aliasing.

I guess I like my images out of the box having a nice pop instead of having to go through rounds of processing with it.
And also although people keep repeating that you can sharpen to regain detail, from personal experience no matter how much you sharpen a mushy edge you still get stuck with softer, mushier transitions upon interpolation.

which goes back to your explanation of 3 pixel natural edge vs 2 pixel unnatural edge
I actually don't find 2 pixel edges unnatural. I actually think that's what makes lines so much more defined.

Another thing is that how visible are aliasing effects when files are actually printed?

[telyt]

I don't see much aliasing in the first photo, but in the second one I do.

You're only seeing aliasing with some extreme pixel-peeping, and in details that would have been mushed out by an AA filter.  In a 16" x 24" print you'd have to use a loupe and your imagination to see any of this.  So, pick your artifacts: mush or aliasing.

[telyt]

1) At the nose no hair is actually resolved, this may depend on very little or thin hair or beeing slightly out of focus
2) A bit from the nose the hair is actually resolved

limited DOF.  The photo was made with a 280mm lens @ f/4 and about 3 meters distance.

3) About half way there is a lot of artifacing. Many strains are discontinous

can you point these out?  With extreme pixel-peeping there are some whiskers that show stair-stepping but this is not evident in a large print.  Many hairs are crossing over other hairs, is this what you're seeing?

4) Just below the hair is not resolved but just goes mush. This is not likely a depth of field effect as the eye and is well resolved.

I'm not sure which area you're referring to.  Mush isn't an aliasing artifact.

[michael]

Trying to bring things back to practical reality...

I've spent the past 10 days shooting with four different cameras, a Phase One 645 with P65+, a Sony A900 with Zeiss glass, a Leica M9 and a Panasonic GF1. These were all of fall colour in Northern Ontario, almost all of it tripod or monopod mounted. (One of the most spectacular fall colour shows in recorded history I'm told by park rangers).

I have now made some 40 prints, ranging in size from 13X19" to 24X36".

Here's what I see. The GF1 shots are very nice in print sizes up to 11X17". They are easily distinguished though from any of the others. A wonderful casual use camera.

The Sony A900 with Zeiss glass comes next. Really excellent. No one would kick it out of bed.

Then there's the Leica M9. Visibly superior to the A900, even on small prints. This is a combination of resolution, clarity, accutance and modeling. Much of this may well be caused by the Leica lenses. The point being that while the A900 at low ISOs, and when used with Zeiss lenses, is as good as any DSLR on the market, the M9 using Leica glass just looks superior to my eyes. There's a palpable three dimenionality to the images that's hard to put into words, but is akin to the sweetness of a tube amp over a transistor amp. (Pure Class A over Class A/B).

The P65+ is in a class of it's own, but only when prints over about 20X24" are made. Otherwise I find it hard to differentiate between them and the M9, except that I usually can identify the M9 shots by their other "drawing" characteristics, especially with the 35mm Summilux, which seems to be in a class by itself; almost luminous, if you'll pardon the expression. The P65+ has other superior characteristics, such as dynamic range, but that's another discussion.

The question is, could I make these differentiations in a double blind test? Probably not 100%, but since I took the opportunity to do some sider-by-side shots, and can clearly see the differences (though admittedly they usually are subtile) I'm confident that likely score quite a bit better than random chance much of the time.

As for artifacting, well, try as I might, on prints and at 300% on screen, I just don't see it, as long as sharpening is carefully and properly done. If I'm confusing artifacting with resolution, then long live the revolution.

As for the question asked – if AA filters are so limiting, why do smart companies like Nikon and Canon use them? I would conversely ask, if AA filters are such a problem, why do smart companies like Phase One and Hasselblad and Leica not use them? Clearly designers and engineers make design choices based on what they'd like their equipment to accomplish. We as consumers do similarly. We each have different criteria and needs, not to mention budgets.

The discussions of theory are fine, and I enjoy reading them. But, just as in the world of audio I trust my ears over measurements, in photography I trust my eyes above all else, even if what they are seeing can be described as an illusion.

If you want more from me on this, come to my talk at B&H Photo next Thursday, 10am till noon (though it appears to be sold out). It's titled "The Photographic Illusion".

Michael