I can't see any "larger scale" contrast differences with/without AA filters. I do however see aliasing and digital artefacts screaming out at me in any DMF / Leica digital picture I see. And it's certainly not a "natural" look.
The point spread of a dual birefringent with the layer thickness usually used in normal cameras is below 5µm. And that's the 99% energy limit, not some "halfway" point. Two pixels away, the signal is more than 10Ev down - wich is a lot more than you can say about the diffraction at F4.0....
The main reason why AA-filtered images don't "sharpen up" like images fram an AA-less camera is that you're sharpening stuff that's already been destroyed by curves, NR and gamma. If you know how to counter this, then using an AA-filter preserves a lot more of the reality in front of the lens than a wildly aliasing Leica (or DMF) picture does.
That some people may prefer the artefacts, since they create a false impression of "detail" is something that has to be seen as a different question. All aliasing is distortion of the signal, inducing stuff that wasn't (or shouldn't be) there.
3) That an unsharpened image looks less visually appealing than a sharpened one
True, but again nothing new.
and that the sharpened one closer resembles qualitatively the higher resolution original.
You need to define/quantify "closer resembles qualitatively".
4) That to truly remove all aliasing you need to low-pass filter the image beyond what most people would consider acceptable
That is true (although the sinc filter would do just that for an infinite image). Fortunately a small degree of remaining aliasing is hardly a problem. It is complete omission of an AA filter that turns a picture downsize into a disaster.
I'll show three examples here. Open them in separate tabs in your browser and flip between them.
I really miss your point here. Are you saying that, since aliasing is objectionable and since sharpening purportedly undoes the effect of an AA-filter, it is forbidden to sharpen the lanczos3 resize? Come on. It is not true anyway.
In order to compare apples with apples, you should compare a multistep resize with a single-step resize, or a multistep resize plus sharpening to a single-step resize plus sharpening. I just did, and if there is any more detail in the former than in the latter I am not able to find it.
theSuede wrote:
... I do however see aliasing and digital artefacts screaming out at me in any DMF / Leica digital picture I see. And it's certainly not a "natural" look.
Can you point out the aliasing and artifacts in my DMR photos?
theSuede wrote:
I can't see any "larger scale" contrast differences with/without AA filters. I do however see aliasing and digital artefacts screaming out at me in any DMF / Leica digital picture I see. And it's certainly not a "natural" look.
What's natural? By what standard? You seem to imply that the designers of the highest performance cameras in the world - DMFs are clueless, and their users as well.
Toothwalker wrote:
If that were true you would be able to make the left image look like the right one by means of sharpening.
No, but I can get the same level of aliasing. I'll destroy the rest of the picture with sharpening but the aliasing will be there.
You need to define/quantify "closer resembles qualitatively".
I'm vague on purpose, but I'll show what I mean in my next post with the experimental results.
I really miss your point here. Are you saying that, since aliasing is objectionable and since sharpening purportedly undoes the effect of an AA-filter, it is forbidden to sharpen the lanczos3 resize? Come on. It is not true anyway.
I'm saying that if aliasing is so objectable then you wouldn't want to sharpen the image as it makes aliasing more visible. You can't have it both ways, either aliasing for good or for bad is an acceptable price to pay for increased crispness or it's not. If it is a question of degree (which it obviously is) then we're back to aesthetic preferences.
In order to compare apples with apples, you should compare a multistep resize with a single-step resize, or a multistep resize plus sharpening to a single-step resize plus sharpening. I just did, and if there is any more detail in the former than in the latter I am not able to find it.
I can post examples later but based on a couple of real world images, I could not tell bicubic photoshop and bicubic lanczos3 apart at all. I could tell them and the multi-step method apart when applying sharpening to the former. The results were different, although it's hard to say that one is better than the other without getting into the whole discussion of aesthetic preferences.
By the way, there is no 'canonical' multi-step sharpening. For instance the method I use on my 5DII files produce crap when applied to my M9 files (this lack of generality is why I'm interested in a more general approach).
Anyway, stay tuned I've got some nice experimental results coming up
denoir wrote:
Now in a real world image things are more complex where you have a blend of multiple spatial frequencies. And in effect you'd be blending the poorly reconstructed signal with a properly reconstructed one. So which is worse, sabotaging the good one or throwing away information from the poor one? I don't know and I have so far not heard any convincing arguments either way.
If you have not heard any convincing arguments in favor of throwing away ambiguous information that obscures legitimate detail you need catch up with half a century of literature in signal processing.
Ken brought up this point as well, but I think you underestimate the lenses in question. For instance I used a Zeiss 25/2.8 ZM in a previous example which is diffraction limited at f/4 and resolves up to 400 lp/mm center frame. It is way beyond the resolving power of any 35mm sensor. The crop above was shot with a lens that is around 90% MTF at 40 lp/mm, so you can rest assured that it outresolves the sensor by a large margin.
That 400 lp/mm figure is for 10% MTF. You would need to know the value at 72 lp/mm, or in fact the entire curve. There is some AA effect of these lenses as 72 lp/mm is in the transition region, how much it is precisely we do not know.
So I very much doubt there is any AA effect from the lens.
Well, we can safely assume that frequencies of 1000 lp/mm and beyond are thoroughly AA-ed by the lens.
This however gave me an idea of a rather simple experiment to perform. One could point a high resolving lens on an AA-less camera at a subject where you know moire patterns will be evident. Then by very slight steps you move focus and for each move take a shot. At some point the moire pattern will disappear as the misfocused lens will act as an AA filter. Unlike the camera based AA filters this will be a very high quality AA filter. At that point where the moire disappears one should check how much detail has been lost (if any) in other regions where there was no apparent aliasing. ...Show more →
The idea of such an experiment is interesting. As Ken suggested variation of the aperture gives a more controlled approach, and I think the diffraction blur pattern may yield better filter action than the uniform (in theory) blur disk of defocusing.
If tuned precisely the lens should at one point cut off at the nyquist limit and in theory there should be no loss of detail according to the "aliasing is bad" theory. Correct?
That is correct if you assume a perfect AA filter, which is unlikely. Hence there is not a sudden cut-off point but a transition. As you stop down the legitimate detail will slowly disappear, and so does the aliasing.
By the way, the "aliasing is bad" statement does not say anything about loss of detail. It just says that aliasing is bad.
kwalsh wrote:
As a side note, for your experiment proposal I'd think maybe an easier an more controlled approach would be to just keep stopping down the aperture to increase the size of the Airy disc. Much easier to model and back out what exactly happened I'd expect.
Ken that's an excellent suggestion and precisely what I ended up doing. I hooked up a follow focus to the aperture control of the lens which allowed me very fine control.
I varied the aperture between f/8 and f/11 and took 38 photos at evenly spaced rotations of the aperture ring. Then I downloaded the images, picked a spot where aliasing was evident and went through all the pictures in sequence. Around frame 24 the moire pattern started disappearing and by frame 26 it was gone. Hence the Nyquist limit should be around there.
You can see Frame 0 (plain f/8) where the image hasn't hit the diffraction limit.
Now, you may want to compare detail, but I've got a better crop for that (taken from frame 0 and frame 26).
First however I want you to look at a detail from the same scene, but taken with a 200mm lens + 1.4 extender on a Canon 7D (an AOV equivalent to 448mm on 35mm FF):
This very much relates to an image qualitatively resembling reality. On the anti-aliased images one doesn't see the gaps in the vent. And from a signal processing point of view - you shouldn't if the AA filtering has done its job. However on the plain image without an AA filter it is visible - granted, at the wrong frequency - but one can clearly see that it is actually a vent and not some non-descript brown square. And no amount of sharpening in post can bring out that detail that has been nuked.
Now look at the blinds. In the AA image we can barely see them (in fact had they been perfectly anti aliased we shouldn't see them at all) while in the plain image they are clearly visible - again, with the wrong frequency. But what is more important qualitatively - to show that there are blinds in the window or to show them at the right frequency? What possible improvement would a gray blur have been?
One can also see that the sharpened version cannot bring out the detail that has been lost in the AA process. You just end up with sharpening artifacts.
carstenw wrote:
Toothwalker and Luka, aren't you focusing too much on point sampling when sensors really do an area sample?
Area sampling affects aliasing, so it is relevant to digital image capture. I suppose theSuede can tell us all about that. The issue of area vs. point sampling is however not relevant to the fundamental questions that have been raised in this thread.
Toothwalker wrote:
If you have not heard any convincing arguments in favor of throwing away ambiguous information that obscures legitimate detail you need catch up with half a century of literature in signal processing.
In the context of producing aesthetically pleasing images? Please point me to a reference.
What you seem to be consistently confusing is some information theoretical ideal of producing an image with no incomplete signals. That's just completely meaningless in this context. And in many other contexts as well where your primary purpose is to maximize information retrieval at the expense of the integrity of the signal.
That 400 lp/mm figure is for 10% MTF.
So? MTF10 is easy to spot with ocular inspection and after you add sharpening you increase it anyway.
Well, we can safely assume that frequencies of 1000 lp/mm and beyond are thoroughly AA-ed by the lens.
Yes we can, but as the sensor can't resolve more than 72 lp/mm what the lens can or cannot do at much higher spatial frequencies are besides the point.
By the way, the "aliasing is bad" statement does not say anything about loss of detail. It just says that aliasing is bad.
Yeah, but it does for two reasons. First the practical that any type of real world (optical or digital) low pass filtering will result in collateral damage. Second you have the loss of the "false" detail. Like the blinds and the vent in the experimental results. Call it false detail if you wish, but it contains information about the real world that an AA filter promptly removes (or tries to anyway) just because a perfect reconstruction of the signal isn't possible.
denoir wrote:
I'm saying that if aliasing is so objectable then you wouldn't want to sharpen the image as it makes aliasing more visible. You can't have it both ways, either aliasing for good or for bad is an acceptable price to pay for increased crispness or it's not.
What counts is the ratio of nonsense to legitimate signal. Sharpening does not change that ratio, because like any other algorithm it cannot tell them apart.
I can post examples later but based on a couple of real world images, I could not tell bicubic photoshop and bicubic lanczos3 apart at all. I could tell them and the multi-step method apart when applying sharpening to the former. The results were different, although it's hard to say that one is better than the other without getting into the whole discussion of aesthetic preferences.
I feel that we are converging
By the way, there is no 'canonical' multi-step sharpening. For instance the method I use on my 5DII files produce crap when applied to my M9 files (this lack of generality is why I'm interested in a more general approach).
Lanczos will do a proper job on both the 5DII and M9 files. Unfortunately it is not more general in the sense that photo editing programs offer the algorithm. (Irfanview does, but the implementation is bad.)
denoir wrote:
In the context of producing aesthetically pleasing images? Please point me to a reference.
Do I really need to show you what happens if you downsize an image by just taking every Nth sample? I think you already know how aesthetically pleasing that will be.
Telyt:
That would be hard when the samples you linked are at best 600x600, or a print barely covering my palm. I have no doubt your pictures are very sharp, but then again I used to be used to the files from the P45+ (that I used), and the HB's our art department used. And they were crap when you compared "accuracy per pixel". Try shooting a redhead in studio with a camera like that - or some product. It usually takes the retouch artist at least TWICE as long as if the photographer used the 1Ds3 or the D3x. For us, that meant an additional cost of a few millions per year (a few 100k€).
In print, the artefacting is largely a non-issue at some certain magnifications, then it starts to be an issue again at higher magnifications. Colour aliasing issues can never be "fixed" though, and the destructive power that effect wields is present at any magnification. The original information in the reality in front of the lens has been destroyed - forever. You would then have to resort to manual chroma in-painting and severe signal decomposition filtering techniques (which is what most "de-moire"-filters use). That's like using a smart sledgehammer on details. It doesn't really matter how smart the sledgehammer is, it's still a sledgehammer.
denoir:
That's why you use an AA filter, and not diffraction (when you can). Thank you for the comparison pictures re apertures btw.
AA-filters are really beamsplitters, with a very controlled spread ratio - and very close to zero spread outside the limit the designer choses. You can specify that you want a 6µm spread, and that will result in an octagon-shaped PSF with a very strictly defined outer rim.
The reasons why Leica don't incorporate AA filters into their designs are quite simple:
1) they cannot afford it
2) since they're already using the cheapest, cheesiest materials available for the filter package, an AA filter would add to much thickness, and make the edge colour vignette even worse.
The reasons why MF manufacturers don't inlude AA filters:
1) they cannot afford it
2) they expect their users to be aware of the effect and do what they can when shooting to avoid it.
.....................................
Point vs Area sampling:
When you use normal focal lengths at apertures >F2.0 you could say that the sensors we use today are almost perfectly area-sampling devices. For apertures <F2.0 and for lenses with very short distances out to the exit pupil - a bit less so.
-until you add a Bayer filter...
What makes a bayer sensor a "point-sampling device" is that you're not sampling the entire area for every colour. Red, which also is a very "strong" colour from a vision acuity PoV is only sampled in 1 of 4, which means that you have a 75% information loss even though your sensor is "area-sampling". A very small red point only has a 25% chance of actually being picked up by the sensor...
(and hence the trouble with redheads in the studio.... single strands of hair get "point-sampled" to oblivion and you get all sorts of nasty pictorial effects).
In areas where the interpolation can safely assume reality to be "unicoloured" this is not a problem. In areas where there actually exists some colour information detail, you get a problem.
To make the Bayer principle work right you would have set the AA-filter spread at exactly one pixel width - this would give no gaps between the commutation "scene" > "measurement". Most manufacturers use less than the pixel c-c now.
theSuede wrote:
That would be hard when the samples you linked are at best 600x600, or a print barely covering my palm. I have no doubt your pictures are very sharp, but then again I used to be used to the files from the P45+ (that I used), and the HB's our art department used. And they were crap when you compared "accuracy per pixel". Try shooting a redhead in studio with a camera like that - or some product.
I have no desire to photograph redheads in a studio or products. Show me screaming artifacts:
This is an unsharped 100% crop from this photo:
I haven't printed this bigger than 12" x 18" because my mat cutting equipment can't handle anything bigger, but I assure you this print doesn't show artifacts no matter how close to press your nose to the print. This is one of several printed this size, some with cropping, that leaves gallery owners at a loss for words except to remark at the detail and color quality. This with very minimal processing work (color balance, resize, USM). That's good enough for me.
Toothwalker wrote:
Do I really need to show you what happens if you downsize an image by just taking every Nth sample? I think you already know how aesthetically pleasing that will be.
but really it's not like anything below nyquist is unaffected by sampling frequency. there is a pretty severe phase amplitude interaction near (but below) nyquist. would you throw out that data too because it is corrupted? anybody who is serious about getting an artifact free signal will low pass filter well below the nyquist, choosing the appropriate sample rate to allow them to do this of course, in applications where this is possible. obviously photography is not such an application. everyone wants as much detail as is possible and in many cases adding false high frequency details can, i suspect, be shown to be preferred by a significant majority of photo viewers for many types of images.
Personally I have problem with a lot of images where hunting for detail often results in too much detail; meaning over-sharpening/artifacts. That include my own images as well, the balancing act is really difficult. And to add to the stew, watching it on a (my) screen which happen to have 135dpi gives another look than a more normal screen of 100dpi where the same image can look awful. Add to that viewing it on the iPhone with 300+ dpi. Very confusing. Colors ignored of course. But perhaps this is a different, but related subject: "when does the hunt for detail deteriorate into over-sharpening".
theSuede wrote:
The reasons why Leica don't incorporate AA filters into their designs are quite simple:
1) they cannot afford it
2) since they're already using the cheapest, cheesiest materials available for the filter package, an AA filter would add to much thickness, and make the edge colour vignette even worse.
The reasons why MF manufacturers don't inlude AA filters:
1) they cannot afford it
Um..No, I'm not buying that. For several reasons. First if you look at papers published on the subject you'll see that it's a very old discussion. Here's for instance a paper from 1974(!). You can look in modern literature as well, like The Manual of Photography 10th edition (2011) and you'll get back to the same conclusions:
There's always a tradeoff between aliasing and MTF and when you put an optical low pass filter in the way, you have to deal with its MTF as well. Its optical characteristics will inevitably negatively impact what you get out of the lens - and not just at the frequencies you want.
Second, the economic argument doesn't cut it. AA filters are not something particularly expensive - even the cheapest P&S cameras have them.Furthermore thinking that digital backs with $50,000 price tags that have been designed to maximize optical performance would cut costs by skipping such a trivial component isn't plausible. Leica's ways when it comes to cost cutting may be mysterious, but the same cannot be said for for instance the Leaf Aptus 80. Of course, that wasn't the reason for Leica either - they declared their design goals when they released the M8. In fact with the M8 they went further and did not include an IR filter claiming that one would affect image quality and that it was better to have one mounted on the lens.
Not a very practical decision and most people hated it, but they were right. The M8's per pixel sharpness is slightly (really really slightly...but still) higher than the one of the M9. You are right that the M9's IR filter is crap as IR filters go, but that's again by design. They chose the thinnest (most flimsy one as you expressed it) one and optically least effective they could get away with. Subsequently the IR issue is solved for the most part, but not entirely - you can still get a purple cast on some black synthetics. The impact on image quality is though as minimal as it can be.
Their reason for not having an AA filter is the same - they prefer to use the full sensor resolution rather than solving the problem by increasing pixel density. My 18 megapixel M9 easily outresolves my 21 megapixel 5DII and I can make larger high quality prints with it. You have to go up to the 24 megapixel D3X to match it. Some tests show that it even outresolves the D3X, but I personally don't put too much stock in them - from most tests I've seen they seem to be matched in that department.
2) they expect their users to be aware of the effect and do what they can when shooting to avoid it.
That I do buy. Aliasing is problematic in some cases and you have either avoid it, fix it in post or simply ignore it. If there was somebody demanding AA filters it would potentially be medium format digital back users as MFDBs are very often used in fashion photography and you how "lovely" the combination of textiles + no AA filter can be. And yet they don't. The reason is simple MFDBs are in a megapixel race, but not one equivalent of the P&S cameras a couple of years ago. Here the quality of the pixels matter as the resolution tells you how large you can print. And that's why most MFDB users would rather eat their camera than to let the integrity of their dear expensively purchased pixels to be violated by a brute optical low pass filter.
Colour aliasing issues can never be "fixed" though, and the destructive power that effect wields is present at any magnification.
While I agree that it can look nasty, fixing color aliasing is actually trivial. Switch to lab mode and lowpass filter the a & b channels.
As you can see all the color artifacts are gone. And the drop in color detail is not something you can readily pick up. Our eyes are far more sensitive to b/w detail than to details in colors.
wfrank wrote:
But perhaps this is a different, but related subject: "when does the hunt for detail deteriorate into over-sharpening".
depends to a large degree both on the subject and medium of display i'm afraid. optimally, i suspect we should all have different sharpening algorithms depending on the frequency content of the scene. even then though something sharpened perfectly for the pixel pitch and viewing distance of a standard laptop screen will probably not be optimal even for other lcds...
wfrank wrote:
Personally I have problem with a lot of images where hunting for detail often results in too much detail; meaning over-sharpening/artifacts. That include my own images as well, the balancing act is really difficult. And to add to the stew, watching it on a (my) screen which happen to have 135dpi gives another look than a more normal screen of 100dpi where the same image can look awful. Add to that viewing it on the iPhone with 300+ dpi. Very confusing. Colors ignored of course. But perhaps this is a different, but related subject: "when does the hunt for detail deteriorate into over-sharpening"....Show more →
It's actually a rather complicated topic, but in my experience an image can become too saturated with detail if the MTF is very high at all spatial frequencies and the subject happens have a very broad spatial spectrum. Then you get a very busy image. Ideally resize & sharpen should be lens specific, but well.. you see that it's difficult to even agree on what is desirable in a resized image.
sebboh wrote:
depends to a large degree both on the subject and medium of display i'm afraid. optimally, i suspect we should all have different sharpening algorithms depending on the frequency content of the scene. even then though something sharpened perfectly for the pixel pitch and viewing distance of a standard laptop screen will probably not be optimal even for other lcds...
Indeed. Just thinking about it makes me uneasy. I don't know how many times I've posted some image that looked OK at home and terribly over sharpened or too soft when viewed on some monitor at work or on my laptop (screens with different pixel pitch).
denoir wrote:
Indeed. Just thinking about it makes me uneasy. I don't know how many times I've posted some image that looked OK at home and terribly over sharpened or too soft when viewed on some monitor at work or on my laptop (screens with different pixel pitch).
yeah, for a while i had an previous generation of dell monitor hooked up to an imac at work. the difference in apparent sharpness of images when dragged from one screen to the other was quite shocking.
As you can see all the color artifacts are gone. ...
This crop demonstrates the crux of the problem. The "color artifacts" are a result of moire brought about by high frequencies that should not be present at the time of sampling. Having now been sampled by the sensor, it's too late to correct the error. One can only fix the image a posteriori with additional knowledge of the scene (perhap you took a Polaroid ). An example: image of a man in a gray suit (badly color-aliased) unhappily holding up a photograph of himself wearing a grey suit (badly color-aliased). Do you hack the man's suit, the photograph, both? Maybe his suit isn't gray, and really looks that way.
Moire (and color artifacts) can engulf the entire image in an obvious way, but an over-sampled image is also noticeable: I sense it as noise, or grain. If certain photogs and artists see over-sampling as Good Stuff, well, good luck. I prefer to start with a sampling of reality - hold the signal ambiguity - and add my own flavor of fantasy in post-processing. That's Art!
