Leica M9 Serious contender?

[250swb]

What evidence do you have of this conscious design decision?  Seems a huge leap of faith on your part.  

I'm willing to bet it had a lot more to do with engineering than bravery.

Its a 'brave' decision when they face the potential bollocks written about AA filters in reviews and forums like this. Of course its a freaking engineering decision, crikey what do I need to do to stop the odd ironic word being obssesed over?

Steve

[joofa]

What might those conditions be, and how likely are those conditions to be met in real-world digital image capture?

If the sampling rate is greater than max. frequency of the original signal but less than Nyquist rate then aliasing spectrum can be regenerated using an adequate DSB modulator and subtracted from the aliased signal to get rid of aliasing. It is difficult, but perhaps not impossible, to build such systems in the practise.

[telyt]

Nope ... but you might want to read the thread again ...

You might want to see the prints made from those files b4 you put too much stock in the comments on that thread.

[Jonathan Wienke]

If the sampling rate is greater than max. frequency of the original signal but less than Nyquist rate then aliasing spectrum can be regenerated using an adequate DSB modulator and subtracted from the aliased signal to get rid of aliasing. It is difficult, but perhaps not impossible, to build such systems in the practise.

So if the conditions required to remove the aliasing are not met in any typical digital imaging scenario, why even bring it into the discussion? The aliasing that causes visually unpleasant artifacts is due to input frequencies above the Nyquist rate, so what you're talking about is not only of limited practical value under the best of circumstances, it doesn't make any attempt to solve the real-world aliasing problem--incoming data above the Nyquist rate.

[BernardLanguillier]

I read in Digital PhotoPro about the Leica M9. It seems that this camera uses software rather than a "detail-blurring" filter over the sensor to resolve moire issues. The photos from the M9 I have seen look very good. If this is truly an advancement in technology, why hasn't the camera received more publicity regarding this new (?) technology?

There have been numerous threads on this topic recently, no conclusion.

Some people prefer the apparently sharp look of unsharpened AA filterless sensors, others (like myself) prefer the look of a correctly sharpened image coming from a sensor with an AA filter.

In the end, there is very little difference in print at a given pixel count.

Cheers,
Bernard

[Jeremy Payne]

You might want to see the prints made from those files b4 you put too much stock in the comments on that thread.

This thread ... re-read this thread ...

[joofa]

So if the conditions required to remove the aliasing are not met in any typical digital imaging scenario, why even bring it into the discussion? The aliasing that causes visually unpleasant artifacts is due to input frequencies above the Nyquist rate, so what you're talking about is not only of limited practical value under the best of circumstances, it doesn't make any attempt to solve the real-world aliasing problem--incoming data above the Nyquist rate.

Firstly, you are mixing up Nyquist frequency with Nyquist rate. You actually mean Nyquist frequency (f_max) when you mention Nyquist rate (2*f_max) in your comment quote above. Sampling theorem says aliasing will happen for any sampling rate less than Nyquist rate (2*f_max) for a low pass signal. And, I am saying that you can have sampling at f_max < f_sample < 2*f_max = Nyquist, i.e., sampling at less than Nyquist rate and still be able to remove that aliasing resulting from that inadequate sampling. I.e., you can have a lower sampling rate than required by the desired Nyquist rate and still be able to eliminate aliasing.

[Jonathan Wienke]

I think you did not get it.

No, you don't get it. The aliasing that causes objectionable visual artifacts is generally caused by input frequencies higher than the Nyquist rate of the sensor. IOW, if the sensor samples the image at 50 lp/mm and the lens has a decent MTF at 60 lp/mm, you get aliasing artifacts at 10 lp/mm. Your theoretical approach is only useful for frequencies between 25 and 50 lp/mm, and by your own admission, has no practical real-world implementation approach yet. And it offers no help, even theoretically, for situations where the sensor receives an input of frequencies higher than the 50 lp/mm sampling rate.

The only thing I don't get is why you're cluttering up this thread with a discussion of theoretical BS of little practical relevance under any circumstances, that offers no benefit whatsoever to the greatest real-world aliasing problem encountered in digital imaging--input frequencies higher than the sampling rate.

[joofa]

No, you don't get it. The aliasing that causes objectionable visual artifacts is generally caused by input frequencies higher than the Nyquist rate of the sensor. IOW, if the sensor samples the image at 50 lp/mm and the lens has a decent MTF at 60 lp/mm, you get aliasing artifacts at 10 lp/mm. Your theoretical approach is only useful for frequencies between 25 and 50 lp/mm, and by your own admission, has no practical real-world implementation approach yet. And it offers no help, even theoretically, for situations where the sensor receives an input of frequencies higher than the 50 lp/mm sampling rate.

Firstly for the second time: you are mixing up Nyquist frequency with Nyquist rate. You actually mean Nyquist frequency (f_max) when you mention Nyquist rate (2*f_max) in your comment quote above.

Secondly, here is what I am saying: Suppose in a system the max analog frequency is 10 Hz (Nyquist frequency). You then need to sample at 20 samples/sec (Nyquist rate) at least. However, if you sample between 10 and 20 you get aliasing. However, with certain additional mechanism you may still be able to get rid of that aliasing even if you sampled below 20 but over 10. That is all I am saying.

The only thing I don't get is why you're cluttering up this thread with a discussion of theoretical BS of little practical relevance under any circumstances, that offers no benefit whatsoever to the greatest real-world aliasing problem encountered in digital imaging--input frequencies higher than the sampling rate.

Jonathan, I offered you to take it offline.

[BJL]

If the sampling rate is greater than max. frequency of the original signal but less than Nyquist rate then aliasing spectrum can be regenerated using an adequate DSB modulator and subtracted from the aliased signal to get rid of aliasing. It is difficult, but perhaps not impossible, to build such systems in the practise.

Joofa,
    the problem is that some lenses have enough resolution to deliver signals to the sensor with maximum frequencies higher than the sampling rate (whatever you call it) of the sensor, and these high frequency signal will thus be completely indistinguishable in the sensor's output from a certain lower frequency signals: this is what is meant by "aliasing", and no post-processing can completely eliminate it. For some painfully obvious examples, look at images from Foveon X3 sensors of resolution test charts, where the number of line pairs changes as the line pairs get closer together, while still showing sharp black and white alternation instead of more gracefully and honestly blurring to gray. Gray is an honest "I don't know" whereas aliasing can answer a confident "black" at a place where the correct answer is "white": sharpness with no connection to reality.

It is another question, more esthetic than technical, how often and how badly this aliasing actually harms the final "print image quality".

[Jonathan Wienke]

Firstly for the second time: you are mixing up Nyquist frequency with Nyquist rate. You actually mean Nyquist frequency (f_max) when you mention Nyquist rate (2*f_max) in your comment quote above.

Secondly, here is what I am saying: Suppose in a system the max analog frequency is 10 Hz (Nyquist frequency). You then need to sample at 20 samples/sec (Nyquist rate) at least. However, if you sample between 10 and 20 you get aliasing. However, with certain additional mechanism you may still be able to get rid of that aliasing even if you sampled below 20 but over 10. That is all I am saying.

OK, great. Explain how, if you are sampling the image at 50 lp/mm (Nyquist frequency) or 100 pixels/mm (Nyquist rate), you can sample a 75 lp/mm input signal (which is going to alias down to 25 lp/mm) and reliably distinguish the 25 lp/mm aliased signal from a non-aliased signal that was 25 lp/mm before sampling? How you can transform an aliased 25 lp/mm signal back to 75 lp/mm without taking any un-aliased 25 lp/mm signal along for the ride? How does a DSB (dual-sideband modulator?) help solve this problem when the only data you have to work with is the sampled signal?

If what you are saying is valid, I should be able to take an un-aliased image 1000 pixels square, downsample it to 750 pixels using nearest-neighbor resampling (so that no anti-aliasing filtration is done), and then run some kind of software transform on the image to expand it back to 1000 pixels with little or no distortion or loss of detail.

[Jonathan Wienke]

Here's a test image, downsampled from 512 to 362 pixels square using nearest-neighbor interpolation, so that it contains aliasing only between the Nyquist frequency and Nyquist rate.

[attachment=19160:Noise_A.bmp]

If your assertions are valid, you should be able to upsample this image back to 512x512 pixels, and your upsampled image and the original image should be essentially identical.

[telyt]

It is another question, more esthetic than technical, how often and how badly this aliasing actually harms the final "print image quality".

Exactly.  For those who care to go beyond theory to prints, go to Appel Gallery in Sacramento and ask to see my prints.  Take a loupe for all I care.  Examine the prints at your leisure and see how "bad" it gets.  All of my prints in the gallery were made with the Leica DMR (no AA filter), and most were made with a lens that resolves far more than the sensor's sampling rate (Leica 280mm f/4 APO).  The aliasing in these prints is so "bad" that skeptical gallery owners can't resist them.

[Jonathan Wienke]

Exactly.  For those who care to go beyond theory to prints, go to Appel Gallery in Sacramento and ask to see my prints.  Take a loupe for all I care.  Examine the prints at your leisure and see how "bad" it gets.

Whether aliasing is visually objectionable depends heavily on the subject matter. For most natural subjects (like the stuff you shoot, judging by your web site), a modest amount of aliasing can be visually indistinguishable from enhanced sharpness and detail. But for some subjects, particularly cloth and other man-made objects that have regular, repeating patterns, aliasing an be a serious problem.

[attachment=19161:2004_07_02_0017_a.jpg]

This image is an excellent example. It has been resized to 25% of its original dimensions using nearest-neighbor resampling. Even though the "detail" in the grass is really mostly aliasing, it looks OK because we see roughly what we expect to see there. The aliased false detail is similar to what true detail might look like, so the aliasing is not an issue in that area.

But the barn roof is a completely different story. It is made of corrugated sheet metal that has rusted in stripes along the corrugations. The corrugations are vertically oriented, going perpendicularly from the eaves to the peak of the roof, as shown in this crop:

[attachment=19162:2004_07_...0017crop.jpg]

Obviously, there is a major difference between the aliased image and the original subject in this area. The bottom line is that even if aliasing is advantageous for what you typically shoot, that does not mean that it is a good thing for everyone, particularly portrait/fashion shooters where aliased cloth textures can be a huge problem.

[ErikKaffehr]

Jonathan,

This is a very good example of the problem.

So you essentially "emulate undersampling" by doing nearest neighbor interpolation to 25% scale. This undersampling results in a visible pattern in the 45 degree direction although the real pattern is vertical. The pattern is visible on the roof. Did I get it right?!

Best regards
Erik

Whether aliasing is visually objectionable depends heavily on the subject matter. For most natural subjects (like the stuff you shoot, judging by your web site), a modest amount of aliasing can be visually indistinguishable from enhanced sharpness and detail. But for some subjects, particularly cloth and other man-made objects that have regular, repeating patterns, aliasing an be a serious problem.

[attachment=19161:2004_07_02_0017_a.jpg]

This image is an excellent example. It has been resized to 25% of its original dimensions using nearest-neighbor resampling. Even though the "detail" in the grass is really mostly aliasing, it looks OK because we see roughly what we expect to see there. The aliased false detail is similar to what true detail might look like, so the aliasing is not an issue in that area.

But the barn roof is a completely different story. It is made of corrugated sheet metal that has rusted in stripes along the corrugations. The corrugations are vertically oriented, going perpendicularly from the eaves to the peak of the roof, as shown in this crop:

[attachment=19162:2004_07_...0017crop.jpg]

Obviously, there is a major difference between the aliased image and the original subject in this area. The bottom line is that even if aliasing is advantageous for what you typically shoot, that does not mean that it is a good thing for everyone, particularly portrait/fashion shooters where aliased cloth textures can be a huge problem.

[Jonathan Wienke]

So you essentially "emulate undersampling" by doing nearest neighbor interpolation to 25% scale. This undersampling results in a visible pattern in the 45 degree direction although the real pattern is vertical. the pattern is visible on the roof,. Did I get it right?!

That is exactly the point of the exercise. Downsizing with nearest-neighbor causes the same kind of aliasing as one sees when shooting without an AA filter. With the right subject matter, a bit of aliasing can increase perceived detail/sharpness. But with the wrong subject matter, it's just ugly.

[ErikKaffehr]

Jonathan,

What's your view on oversampling, like ignoring diffraction and sampling with say 3 micron pitch? The image would obviously have a lot of non-resolved detail if not a truly excellent lens would be used at a very large aperture. We would obviously down sample using something better than nearest neighbor.

Best regards
Erik

That is exactly the point of the exercise. Downsizing with nearest-neighbor causes the same kind of aliasing as one sees when shooting without an AA filter. With the right subject matter, a bit of aliasing can increase perceived detail/sharpness. But with the wrong subject matter, it's just ugly.

[Jonathan Wienke]

Jonathan,

What's your view on oversampling, like ignoring diffraction and sampling with say 3 micron pitch. The image would obviously have a lot of non-resolved detail if not a truly excellent lens would be used at a very large aperture.

IMO, the loss of DR from halving pixel size/pitch is offset by the downsampling, so that's a wash. You could do noise reduction and demosaic prior to downsampling, so that each output pixel would be true RGB (having real sampled data in each color channel), and demosaic and noise reduction artifacts would be minimized by the resampling. I think a credible case could be made for doing most of this in-camera (except for noise reduction) and outputting a downsampled RGB RAW.

[Daniel Browning]

With the right subject matter, a bit of aliasing can increase perceived detail/sharpness. But with the wrong subject matter, it's just ugly.

That's something everyone can agree with. I would add that some people find aliasing to be ugly even the "right" subject matter.

Some prefer aliased images, describing them as sharp, crunchy, high microcontrast, with lots of fine detail. Others would see it as fake-looking, with harsh transitions, jagged edges, and lots of false detail.

Others would prefer anti-aliased images, describing them as smooth, natural, with the appropriate amount of detail for their size. Some would see them as mushy, hazy, low contrast, and lacking in fine detail.

[BernardLanguillier]

That's something everyone can agree with. I would add that some people find aliasing to be ugly even the "right" subject matter.

Some prefer aliased images, describing them as sharp, crunchy, high microcontrast, with lots of fine detail. Others would see it as fake-looking, with harsh transitions, jagged edges, and lots of false detail.

Others would prefer anti-aliased images, describing them as smooth, natural, with the appropriate amount of detail for their size. Some would see them as mushy, hazy, low contrast, and lacking in fine detail.

Yes indeed. I would add one thing to this, images from AA filterless cameras typically have undergone some form of in camera processing to reduce artifcats (the M9 and S2 come to mind), that are also easy to recognize and can be perceived as being pleasing or not. I would describe this look as micro-level painterly.

Cheers,
Bernard