M8 at f8 on a California Freeway and Moire at 60mph!

[StuartRae]

I downloaded the file and converted it with the only option available to me, ACR 3.6.

Interestingly there were no artifacts in the parallel lines ahead of the car at 100%. (not to say there weren't some elsewhere)

Then I down-sampled to 800px width using different algorithms. This clearly demonstrated that down-sampling does produce artifacts, and that their severity depends on the method used.

100% Crop

Comparison of down-sampling methods

Regards,

Stuart

 

[StuartRae]

Asher,

As Ian mentioned, is that sensor dust or 60mph flies stuck to the lens?

I also alluded to the sky in a previous post. What did you use to convert the RAW file? The ACR conversion shows nice blue skies, as opposed to your purple ones.

Regards,

Stuart

 

[James Roberts]

Hi James,

Thanks for confirming what I see on the original image. The rest of what is talked about here, is nothing to do with how the camera handles the real world, but how us folk handle the resulting digital image. Its very similar as with video/tv, avoid striped shirts! Possibly, printing at some odd resolution may reduce the visibilty of the artifacts, say at 313 dots per inch may alter things slightly, possibly 433dpi may alter it more.

Since the individual sensor sites are arranged on a regular pattern, green probably being different to the others, then I could expect the colours to be seperated, as we saw.

Best wishes,

Ray

Ray, you're absolutely right. The RIP I use to print is also not using PS output, so that may have something to do with me not seeing the patterns (except of course when zooming the screen).

As I said Asher, I believe what you're seeing is real--the fringing--but the interference patterns are definitely the RAW converter.

I don't believe ACR is ready for the M8 yet, actually. I know it's official with ACR 3.6, but the quality out of C1 is much better IMO.

At 100%--the only magnification at which to judge at all in PS--I just don't see typical moire patterns. Fringes, yeah--which is odd enough--but not moire patterns.

So there's something odd going on here--probably some slight aliasing. Hard to tell for sure though, because it's that slight in prints.

 

[Ray West]

Hi James, et al,

There is nothing odd going on here at all. It is what is expected if you have lines or patterns that do not quite line up with other lines or patterns. Just spend a few minutes with google search on 'moire', etc. Plenty of examples, some of which will 'hurt your eyes'.

Take a picture, with any camera, film if you wish, of similar converging patterns, or just get a cad program or even photoshop, and draw something similar. Then adjust it, so that it 'crosses the pitch' of the display device, either vdu or printer. you will get exactly the same sort of effects as Stuart has shown. As I keep saying, it is not camera related, other than its a sharp image. There are lines in the road, which may or may not be enhanced by the image processing, but they are there, else Asher has a very, very, peculiar camera.

I fail to see what the problem is with the camera - there is no moire, other than the slight effects in the regions I highlighted, as far as I can see. What is being discussed now is a different problem - one of 'how to print a digital image on a digital printer.'

wrt downsampling, if you do it to a slightly different scale, you will see the pattern move, as the pixels hit/miss the region under consideration. Qimage gives a good selection of algorythms, some will be better than others, but all, since they use a regular pattern, will have some sort of fringing occurring. Until software is designed with some sort of AI, you have to correct for it manually. It is a benefit of digital ;-)

Best wishes,

Ray

 

[Bart_van_der_Wolf]

wrt downsampling, if you do it to a slightly different scale, you will see the pattern move, as the pixels hit/miss the region under consideration. Qimage gives a good selection of algorythms, some will be better than others, but all, since they use a regular pattern, will have some sort of fringing occurring. Until software is designed with some sort of AI, you have to correct for it manually. It is a benefit of digital ;-)

In fact, by my request, Mike Chaney did include a type of AA-filtering for down-sampled images in Qimage. The user has a selection of several levels of filtering, but IMHO the default strikes the best balance between aliasing artifacts and blurring (although, again IMHO, not optimal but better than most).

On up-sampled images this will not have any effect. In general, magnification of an image will also magnify (aliasing) artifacts. The aliasing, if visibly disturbing, should be prevented before capture. Otherwise there is little else (other than local color channel editing/blurring) left to mitigate the issue. Prevention is still better than cure ...

Bart

 

[Bart_van_der_Wolf]

I don't know if anyone's looked at the embedded JPEG (320x240) from the DNG file, but having just had a look at Bart's very interesting article about down-sampling algorithms it does seem to confirm his theory.

Just in case one wonders, Stuart is (probably) referring to my theoretical and practical down-sampling example pages. As sensor sizes increase, the chance of having to down-sample increases (e.g. for web publishing of samples), so I created those pages in an attempt to avoid having to repeat myself ...

Those pages demonstrate artifacts due to down-sampling, making it 100% certain that detail will hit the capability limits of faithful reproduction in a regularly sampled discrete image, although similar effects can be expected from under-sampled sensor images, like the one discussed in this thread.

Bart

 

[Ray West]

Hi Bart,

Whatever you do before the light hits the sensor, will be a trade off, as you mentioned earlier. I believe that all you do is shift the problem further away, without solving it. I think, since it is digital artifacts, that a digital solution, i.e. software, will provide the answer. It's just, at the moment, nobody, afaik is really looking at getting a solution. I'm thinking it would need some in built intelligence, to detect the interference patterns, and blur them out, or whatever, in that area only. A sort of auto morphing tool, to give a gentle change between the blurred distance, and the sharper mid ground, in this instance. If this particular image was worth the effort, maybe an hour with a brush tool, or five minutes if I was sixteen years old ;-)

When you think about it, adobe and the like are so primitive, its quite unbelievable, a bit like trying to travel to the moon by horse and cart....

Best wishes,

Ray
.

 

[Bart_van_der_Wolf]

Hi Bart,

Whatever you do before the light hits the sensor, will be a trade off, as you mentioned earlier. I believe that all you do is shift the problem further away, without solving it.

Hi Ray,

I agree that AA-filtering is not a solution, but it is an attempt to prevent it from happening. Mathematically aliasing is an easy problem to prevent, just remove the highest spatial frequencies in the projected image before it hits the regularly spaced sensels. The catch is that it will reduce the resulting resolution by a factor of 2-3. Given that aliasing artifacts are not always as obvious as in the example shown in this thread, or they can sometimes be remedied rather easily after the fact, most of us are unwilling to pay the price of lost resolution just to be on the safe side.

The most elegant prevention is the use of an AA-filter, since it strikes a balance between reducing the chance of producing artifacts with only a limited impact on resolution. Because it is not optimal (or probably even possible in the M8), there will still be a need for some postprocessing. However, if we start without any prevention the task will be much harder.

I think, since it is digital artifacts, that a digital solution, i.e. software, will provide the answer. It's just, at the moment, nobody, afaik is really looking at getting a solution. I'm thinking it would need some in built intelligence, to detect the interference patterns, and blur them out, or whatever, in that area only.

The real problem here is that the larger sized aliases of impossible to reliably image small features (smaller than 2 samples, or 4 pixels in a Bayer CFA) are intermixed with real detail, and it has become impossible to remove one without damaging the other. That's why it should be prevented before it happens.
I agree that since human vision is good at spotting unexpected patterns, there may be some room for AI solutions in the future, but again it can't be perfect.

A sort of auto morphing tool, to give a gentle change between the blurred distance, and the sharper mid ground, in this instance. If this particular image was worth the effort, maybe an hour with a brush tool, or five minutes if I was sixteen years old ;-)

When you think about it, adobe and the like are so primitive, its quite unbelievable, a bit like trying to travel to the moon by horse and cart....

Yes, there is still a long way to go, but the laws of physics will pose a solid restriction on what can ultimately be done. I remain a strong advocate for 'prevention is better than cure', because GIGO (garbage-in garbage-out) still rules, even if we get better tools to cover things up a bit.

Sofar, as a postprocessing kludge, I occasionally use a severely (color)noise reduced layer or even blurred layer with mask to locally reduce the visibility of the most glaring artifacts. Local desaturation may also help somewhat in the case of false color artifacts.

Bart

 

[Ray West]

Hi Bart,

Thanks for your considered reply. I agree with what you say, but I personally think, whatever that may be worth, that the balance leica have is not too bad - just avoid tapered stripes... In more 'natural' scenes, its not a visible problem, afaik, but to be honest I've not checked in any detail. For example, I do not think there have been any complaints here, http://www.openphotographyforums.com/forums/showthread.php?t=1684 and I guess enough folk have been looking.

Best wishes,

Ray

 

[Asher Kelman]

Asher,

As Ian mentioned, is that sensor dust or 60mph flies stuck to the lens?

I also alluded to the sky in a previous post. What did you use to convert the RAW file? The ACR conversion shows nice blue skies, as opposed to your purple ones.

Regards,

Stuart

Sorry Stuart for missing your post!

(I was taking a workshop on Leaf Aptus Backs! I'm now a certified Leaf Technician!)

The spots must be on the sensor. I have not cleaned the sensor and have not botheredhoot as I normally shoot much wider apertures. Out of a car, I wanted to get more depth and so those bugs are dirt on the sensor!

The sky color? I cant remember. I liked it. I didn't see purple. Is your monitor calibrated?

I'll recalibrate my Eizo and see how it is!

Asher

 

[StuartRae]

Asher,

I liked your sky as well, but the red channel is significantly higher. Maybe "purple" is the wrong word, but have a look at these two crops from the top left corner.

ACR conversion RGB = 43:110:161

Your version RGB = 105:116:168

My monitor is as well calibrated as I can get it, but alas it's dying, and must soon be replaced.

Regards,

Stuart

 

[Nathaniel Alpert]

I have been following this discussion with interest and some dismay. I wish to clarify some of the points that Bart van der Wolf and other have made.

van der Wolf wote "... I agree that AA-filtering is not a solution, but it is an attempt to prevent it from happening. Mathematically aliasing is an easy problem to prevent, just remove the highest spatial frequencies in the projected image before it hits the regularly spaced sensels. The catch is that it will reduce the resulting resolution by a factor of 2-3. Given that aliasing artifacts are not always as obvious as in the example shown in this thread, or they can sometimes be remedied rather easily after the fact, most of us are unwilling to pay the price of lost resolution just to be on the safe side."

There seems to be a rampant misconception, namely that one is "entitled to all" the resolution available from a given lens, independent of the properties of the sensor. But you really cannot go beyond the resolution limits imposed by the sampling of a digital sensor without corrupting the image. And do we really need to push the envelope? The 10-16 Mpixels sensors available today can provide excellent enlargements, better in some ways than can be achieved with 35 mm film. Adding a properly designed AA filter does not reduce the resolution by a factor of 2 to 3. I have never seen any evidence of that. Theories (and practice) in digital signal processing show that it is possible under somewhat idealized conditions to **exactly** reconstruct an analog signal from a finite number of digital samples, provided there is no information in the incoming signal that exceeds the Nyquist frequency of the detector. In terms of digital imaging it is only in those frequencies below Nyquist that one may gain an advantage with a good lens. (Some photographers use Leica lenses on the 1DsII for just that reason). One needs to consider the lens and sensor together when designing the system. I have not seen any evidence that the beloved Leica look now seen by digital adherents comes from the inclusion of information that exceeds the Nyquist frequency. My expectation is that whatever we are seeing comes from better performance at lower frequencies.

Imaging without an AA filter violates the no free lunch theorem. Mathematical analysis shows that it is not possible to remove the effects of undersampling (aliasing). The only valid way to obviate the AA filter is to make the sampling in the sensor so high that its Nyquist frequency exceeds the cut off of the lenses MTF. Probably, the current sensors are not that far from matching lens performance. Notice that no one dropped the AA filter with a 6 Mpixel camera. The aliasing would have been intolerable.

There will not be any practical antialiasing software. If we are lucky some mitigation may be found to disguise the harmful effects. Lieca surely understood the trade off imposed by a lack of AA filter. In certain scenes whose frequency content somehow just matches the properties of their sensor, you will get a little improvement. In all other cases their camera will perform less optimally than it would with a good AA filter. Leica photographers will just have to accept and adapt to this trade off.


I hope this post adds to rather than subtracts from the debate.

 

[Asher Kelman]

How close to the sensor does an anti-aliasing filter have to be?

IOW, could one design such a filter in front of the lens? Oe else could one have a filter that popped down close to the sensor when you activated that "function"?

Asher

 

[Nathaniel Alpert]

How close to the sensor does an anti-aliasing filter have to be?

IOW, could one design such a filter in front of the lens? Oe else could one have a filter that popped down close to the sensor when you activated that "function"?

Asher

The object of the filter is to remove (image smoothing) higher frequencies. It could be part of the lens design, as you are in effect degrading the lens resolution.

 

[James Roberts]

Nathaniel,

I actually agree with your statement that the obvious (to me, anyway) sharpness and detail improvement of the M8 over, say, a 1ds2, has to do with sensor performance unaffected by aliasing.

And I'm not sure, really, how Leica has designed any of the filtering that is on the camera--IR or AA, though I have little reason to doubt that you are correct theoretically speaking.

All I know is that--practically speaking--the shot as posted by Asher presents little real-world difficulty to print.

I've also just now started to see the combination of new M8 (new hardware and firmware) coupled with new ACR (in PS CS3 beta). So far, the combination is looking very good indeed!

 

[Asher Kelman]

Thanks Jamie for the helpful and positive response.

I am as you know a strong admirer of the new M8. Still, I do find things and getting the moire minimized conveniently is needed.

I'll use CS3.

I also have newer images!

Asher

 

[Asher Kelman]

Looking at your pictures on my apple monitor clearly shows what you are saying!

What can I say?

I have weird tastes!

Asher

Asher,

I liked your sky as well, but the red channel is significantly higher. Maybe "purple" is the wrong word, but have a look at these two crops from the top left corner.

ACR conversion RGB = 43:110:161

Your version RGB = 105:116:168

My monitor is as well calibrated as I can get it, but alas it's dying, and must soon be replaced.

Regards,

Stuart

 

[Asher Kelman]

I have another approach. I gave this issue to an imaging specialist who designs compression algorithms to see if he can get this issue solved mathematically.

Asher

 

[John Sheehy]

There's some component that's related to the screen and magnification. Here, in PS with a 1280x1024 LCD at 33% it's very apparent, at 50% it almost vanishes and it comes back strongly at 66%. I'll be curious to see what happens on the 1600x1200 LCD.

LCDs are sharper than CRTs at the pixel level, so any aliasing is going to be more obvious.

As far as zoom percentages are concerned, many viewing/editing programs use nearest neighbor algorithm for some or all ratios, so the display will have additional aliasing beyond what the original has. 66% is a notoriously bad viewing resolution in Photoshop. I never use it.

The original in this case is full of aliasing and moire. All high-contrast edges have color fringing in the RAW. A shot like this should be shot at a very high f-stop, so diffraction will eliminate the alasing, IMO. For shallow DOF, you are stuck with aliasing unless you use some kind of filter that throws some photons off a little bit.

IMO, this is not a good design philosophy. A proper AA filter used with this camera would result in highly sharpenable RAWs. The sharper your lens is at the pixel level (and Leica lenses are very sharp), the more sharpenable the image is without sharpening noise. The non-noise-related sharpening artifacts would be no worse than the aliasing.

 

[Asher Kelman]

John,

Could one design an external filter which would be exactly matched for f stop etc that one could slip in front of the lens of this particiular camera.

Asher

 

[John Sheehy]

John,

Could one design an external filter which would be exactly matched for f stop etc that one could slip in front of the lens of this particiular camera.

You'd want something that would spread photons about 1/2 to one pixel width off course. I don't know if its possile to do correctly from the front threads. The best place for this is right over the sensor.

Have you tried high levels of caffeine?

 

[Asher Kelman]

John, maybe I'm already over the caffeine limits! I can still move the camera exactly 0.6 pixels. I can't change the tremor to integers or half integers.

To the moire, I wonder whether poor subsampling technic from the data incoming from the chip might be responsible in part. I'll put online images taken under better conditions.

It seems that when any digital image files are encoded and then stored, data is discarded in uniform non-detail rich areas. So that is one place that the M8 CCD encoding could worsen the already present detail mismatch, (that the sensels are insufficent for the incoming detail). I am told that such subsampling could be improved.

One especially difficult area is the removal of moire when the repeating lines occur in more than one axis. In this image, there is an added issue, the lines are also converging.

Asher

 

[James Roberts]

{snipped}
The original in this case is full of aliasing and moire. All high-contrast edges have color fringing in the RAW.
{snipped}
IMO, this is not a good design philosophy. A proper AA filter used with this camera would result in highly sharpenable RAWs. The sharper your lens is at the pixel level (and Leica lenses are very sharp), the more sharpenable the image is without sharpening noise. The non-noise-related sharpening artifacts would be no worse than the aliasing.

John, with all due respect, I am not seeing what you're evidently seeing. In C1, the "original RAW" is not full of aliasing. The high contrast edges DON'T have any colour fringing whatsoever.

They DO in ACR, so I can only think that they're not uncompressing the RAW file correctly. It's a 16 bit (actually 14 bit) image compressed to 8. As far as I know, Leica only worked with C1 on the M8.

(ACR 4 does work a lot better than ACR 3.x with the M8--it has a better set of controls--but it's still not as good as C1).

The only thing a C1 16 bit TIFF produces--at 100% in PS or in a print--is the colour "bleed" already reported. That may be a product of aliasing--absolutely--but it's not a bad problem at all.

Even the 1ds2 displays moire under certain circumstances, and IMO the M8 is sharper, and (I'm going to get flamed for this) produces as good or better enlargements in print (which is saying something given the smaller size of the sensor).

 

[James Roberts]

More moire...(say that 5 times quickly!)

Ok, my point about Asher's photo was that the color fringing I was seeing in a print was minimal, and that all cameras have moire to some degree...not that the post Asher posted didn't have artifacts.

So in the interest of full disclosure, have a look here at this thread on the LUF talking about moire and showing some "good examples" of "bad moire" from things like rooftops.

http://www.leica-camera-user.com/digital-forum/11763-have-look-m8-image.html

Note that I haven't seen these at 100%, but they look more typically--and more severly--flawed to me than Asher's shot.

Note, too, that there's a link at the bottom of the thread to the 5d doing the same sorts of things, and it certainly has an AA filter

 

[Asher Kelman]

Does anyone have pre-release M8 files which I hear were 16 BIT? I'd like some moire examples to analyse.

Also Anyone have M8 files with no sub-sampling from a very early Leica test bed?

Asher

 

[James Roberts]

Does anyone have pre-release M8 files which I hear were 16 BIT? I'd like some moire examples to analyse.

Also Anyone have M8 files with no sub-sampling from a very early Leica test bed?

Asher

Asher, three quick thoughts...

First, the M8 shots today are compressed to 8 bits, with a maximum loss of 2 bits (if I understand the compression algorithm) and so uncompressed are about 14 bits (essentially the same as the DMR). This allows Leica to claim they are, in fact, 16 bit RAWs.

Secondly, I've heard through a pretty reliable source that a nearly-done level of firmware will let you turn off the compression, and get a DMR-like DNG. Personally, the difference between consistently large (20MB) and slow to work with DNGs and the new M8 ones are so great that if I could compress my DMR files in the same way I most certainly would!

Last--what do you mean by sub-sampling?

 

[Asher Kelman]

Jamie,

We may have a new compression utility available. I just want to be able to look at very primitve Lecia image files even before any data is compromised to make the DNG.

Anyway, I'm very happy with the M8, even like the noise at high ISO! Has a painterly feeling!

Subsampling: going from the digital conversion data set from the chip to the RAW file involves a lot of sampling decisions and deciding what to keep. Id like an software designer to try and put his hand to these primitive files before the anything else is done.

Asher

 

[John Sheehy]

Asher, three quick thoughts...

First, the M8 shots today are compressed to 8 bits, with a maximum loss of 2 bits (if I understand the compression algorithm) and so uncompressed are about 14 bits (essentially the same as the DMR). This allows Leica to claim they are, in fact, 16 bit RAWs.

The only "compression" used is gamma. The highlights, therefore, are a bit lossy, but the deepest shadows could hypothetically contain far more depth than a 12- or 16-bit linear RAW.

 

[Asher Kelman]

John, could you expand on that? The shadows.

Asher

 

[John Sheehy]

John, could you expand on that? The shadows.

It's what the values represent.

For 12-bit linear, you have 0, 1/4095, 2/4095, 3/4095, etc. Actually, most cameras don't have 4096 levels, but something around 3500 - 4000.

For 8-bit data with 2.2 gamma, the values, when converted to linear, are 1^2.2/255^2.2, 2^2.2/255^2.2, 3^2.2/255^2.2, etc. Translated to 1/4095ths, they come out to be 0.021/4095, 0.096/4095, and 0.233/4095. You pay for it at the highlight end, where the three highest 8-bit values represent 4025/4095, 4060/4095, and 4095/4095.

I spoke too soon about 16-bit linear in a previous post; 8-bit gamma 2.2 is only very slightly advantageous for the deepest shadows, and only has a couple levels lower than 16-bit's "1".