Home · Register · Join Upload & Sell

Moderated by: Fred Miranda
Username  

  New fredmiranda.com Mobile Site
  New Feature: SMS Notification alert
  New Feature: Buy & Sell Watchlist
  

FM Forums | Leica & Alternative Gear | Join Upload & Sell

  

Archive 2013 · Leica M9 vs MM resolution?

  
 
Mirek Elsner
Offline
• • •
Upload & Sell: Off
p.1 #1 · p.1 #1 · Leica M9 vs MM resolution?


Hello all,

I am looking for some data showing difference in resolution between the two cameras. The only test I am aware of is from Erwin Put's web site (http://www.imx.nl/photo/leica/comparison-m-bodies.html) and I am a bit confused with the resolution going behind Nyquist limit there.

I am trying to understand impact of the Bayer filter in general, so I would be also interested in comparisons of identical MP Bayer vs. Foveon sensors and any other reading on the topic.

Thanks!



Aug 30, 2013 at 07:04 PM
edwardkaraa
Offline
• • • • • •
Upload & Sell: Off
p.1 #2 · p.1 #2 · Leica M9 vs MM resolution?


I think you better drop a pm to Joachim aka the suede. He's the expert on these matters.

If I remember correctly, in theory, you should get double resolution from the MM, but realistically it should not be more than 1.25x to 1.3x.



Aug 30, 2013 at 10:24 PM
Policar
Offline

Upload & Sell: Off
p.1 #3 · p.1 #3 · Leica M9 vs MM resolution?


Very confused by the last comment, but so far as I understand:

You can get plenty of signal past Nyquist, but you risk aliasing. These cameras lack AA filters so no surprise there's detail past Nyquist!

As regards sharpness... look at the area under the mtf curve (should be easy to approximately measure this area) and that's a pretty good indicator of overall apparent sharpness. Extinction of meaningful resolution (let's say arbitrarily 20% mtf) is obviously first on M9 and tellingly last on the monochrome. I'm pretty sure a foveon chip would be similar to the monochrome.

Just based on guesstimate third grade math:

As regards linear sharpness, the monochrome is 32% sharper than the M9.

And the M is 16% sharper than the M9.

(Measured by guessing area under the respective mtf curves, dividing one by the other, and taking the square root of the divisor.)

As regards useful resolution (measured by arbitrarily looking up 20% mtf):

The monochrome resolves 40% more detail than the M9.

The M resolves 26% more detail than the M9.

Looks like the Bayer filter reduces sharpness by about 30% and resolution by about 40%... Both figures will go way down as pixel density increases as the majority of the mtf is already decimated by the lens.

I think?

*Sigh*

This is why I was an English major.

Cute that Velvia appears sharper (with less resolution)!



Aug 30, 2013 at 11:40 PM
Taylor Sherman
Offline
• • •
Upload & Sell: On
p.1 #4 · p.1 #4 · Leica M9 vs MM resolution?


You can't get "signal" past the Nyquist. How could you see detail smaller than a pixel?

The contrast lines on EP's graphs continue past the Nyquist because there is apparent contrast remaining. That is, some of the bright lines correspond to pixels and some of the dark lines correspond to other pixels, and the measurement tool sees there are bright parts and dim parts and computes "contrast". What it doesn't show is that the relationship of the lines is all screwed up, showing moire and other artifacts (many seemingly gray pixels in between bright/dark ones, etc).

If there were an ideal anti-aliasing filter in front of the sensor, the lines on the graph would drop to zero at the Nyquist: the high-frequency image information would be filtered out completely, leading to a uniform gray.

They drop lower on the M9 etc because, though they lack AA filters, they do have Bayer color filters and the averaging done across pixels to estimate color at each location has a smoothing effect (low-pass spatial filter).




Aug 31, 2013 at 12:14 AM
Policar
Offline

Upload & Sell: Off
p.1 #5 · p.1 #5 · Leica M9 vs MM resolution?


Taylor Sherman wrote:
You can't get "signal" past the Nyquist. How could you see detail smaller than a pixel?


The Nyquist limit is 2 pixels. In theory.



Aug 31, 2013 at 10:16 AM
theSuede
Offline
• • • •
Upload & Sell: Off
p.1 #6 · p.1 #6 · Leica M9 vs MM resolution?


Firstly, I think some of you are thinking about this from the wrong starting point. The output signal (the sensor, the image) has a fixed resolution, and I would guess that's where the confusion originates. Of course the output signal can't have higher resolution than what it physically has (the basis of discrete sampling...), and this might cause a slight logic hiccup when you think about MTF at "frequencies above Nyquist".

But you have to get stuff in order to realize what is what. The MTF frequency is the TARGET's frequency as interpreted by the recording device's resolution - and that gives "an output result" - regardless of the sensor resolution!

So if you measure (shoot in this case...) a sine target with a resolution higher than what the recording media (sensor in this case) can register, but the sensor STILL shows some kind of remaining fluctuations (contrast patterns) in the output signal, then you have output above Nyquist. This is above what the sensor ideally SHOULD be able to record.

The reason for this is that even when a pattern is smaller than a pixel, the value recorded in the pixel is still an integral. To return to the sine wave pattern, imagine that you have a SLIGHTLY higher target resolution than than what the sensor can handle. Not much, maybe 10% higher. Ten half-cycles of signal only covers nine pixel widths...

Then the first pixel will still record a signal that is mostly the first postive part of the sine - and then include 10% of the negative part. The pixel will still be white, only a tiny bit less white than if the signal was a perfect fit.
The second pixel will start measuring about 10% into the negative swing, get the rest of the negative swing and 20% of the next positive part. It will still mostly be black...!

Around pixel number five, you've split the signal polarities beween two pixels so that one pixel gets half the negative and half the positive part (resulting in mid gray), then the next pixel will get half the positive and half the negative - ALSO resulting in mid gray. The contrast between those two is close to zero.

Around pixel 9-10 you've gone a full phase-difference cycle, so that you yet again have maximum output swing. Pixel 9 is close to white, 10 is close to black.

This will result in a pattern with a nine-pixel cycle. This pattern goes from almost maximum contrast between one pixel and the next (pixels 0, 9, 18, 27 and so on) to close to zero contrast (pixels 5, 14, 23, 32 and so on).

This is called aliasing, and it can be a very strongly visible effect in a monochrome sensor.
..................

I'm mobile at a new MBP, so I'm sorry I can't simplify the explanations with some graphics. But a clarification that might be necessary for everything to fall into place is that MTF is a model based on how a sine signal is modulated by a system, regardless of the system resolution. From a purely mathematical PoV, the MTF is the integral of the original contrast over the integral of the image contrast, without any regard what so ever to what resolution/frequency they are.

Understanding that, you might also instinctively grasp that the resulting MTF contrast is the same as the average "interaction result" over a large sample area. So the resulting contrast in the mathematical model is the perfect average of all possible interactions/positions of the pixels and the sine signals.



Aug 31, 2013 at 05:39 PM
snapsy
Offline
• • • • • •
Upload & Sell: On
p.1 #7 · p.1 #7 · Leica M9 vs MM resolution?


Thanks theSuede for your nyquist sampling theory explanation. How does a bayer vs monochrome sensor interact with this sampling though?


Aug 31, 2013 at 05:56 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.1 #8 · p.1 #8 · Leica M9 vs MM resolution?


To get to the problem with Bayer without spending to much time on details, just consider what records "color" in the image.

Red dominance is determined by one pixel in every 2x2-group of pixels. If you have a very thin red line that passes through that group of pixels without at any time even partly covering the red sensitive pixel, the sensor won't even know the red line was there...

This is what causes moire, and it's not reall the same as aliasing if you want to be strict. Aliasing happens even when all of the target is sampled, like in the monochrome M, moire happens when the sensor actually totally misses parts of important information in the target.

A raw converter's job is to guess the two missing datapoints for each pixel, and it does that by (more or less) intelligently looking at the pixel values closest to it. That the red and green values in a "blue" pixel gets correctly estimated necessitates that all the surrounding pixels next to it contains VALID and ACCURATE results.
In the example where the thin red line misses the red pixels completely, not only are the pixels where the red line "should" have been given totally wrong values, but all the surrounding pixels too - since the red line data just isn't there, it can't give the surrounding pixel the information support that THEY need to get correct estimates.



Aug 31, 2013 at 05:58 PM
snapsy
Offline
• • • • • •
Upload & Sell: On
p.1 #9 · p.1 #9 · Leica M9 vs MM resolution?


Thanks again. So is that to say the resolution increase for a monochrome sensor is mostly for color aliasing situations where either the color detail would be completely missed or undersampled? Stated alternately, is there any resolution increase for green-dominated detail on a monochrome sensor vs bayer?


Aug 31, 2013 at 06:09 PM
Mirek Elsner
Offline
• • •
Upload & Sell: Off
p.1 #10 · p.1 #10 · Leica M9 vs MM resolution?


I would think that most real world detail will have signal in all four channels, so the demosaicing software might be able to reconstruct it pretty well and close to the Monochrom. But both the Monochrom and the Sigma "look" more detailed...


Aug 31, 2013 at 06:21 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.1 #11 · p.1 #11 · Leica M9 vs MM resolution?


Yes, there is (more resolution and better contrast) in a monochrome sensor than in a Bayer based sensor even when you shoot a continuous hue surface, and that's true even when it's the dominant green.

In a Bayer interpolation you have to weigh color against luma against noise - for two unknowns per position. So let's say you have a dark green line on a slightly brighter green solid. IS that line (from a speculative PoV)
*Just a darker version of the same colour?
*Actually a different colour, with the same brightness?
*Just noise? Remember that sqrt(signal) will always be noise due to photonic noise...

A good interpolation algorithm will make a good blend compromise between the three options based on surrounding data, but in detail that is so small that there actually IS no really supporting data to be found in the pixel's neighbours - what do you do?
The normal action is "average compromise". Add in a little bit of all three solutions. This will of course add in colour errors (though small from a human-perception PoV) while lowering contrast.

The monochrome doesn't have to do this, the base parameters is already set by the filtration. If a pixel says it should have "this" value, then the ONLY source of error can be noise of some sort. And in a low-ISO image, you don't have to do any NR at all, since the noise at that exposure should be low enough to be almost imperceptible.



Aug 31, 2013 at 06:38 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.1 #12 · p.1 #12 · Leica M9 vs MM resolution?


For an average image (Kodak and SPIEE figures) you only lose about 4-5% due to the Bayer interpolation, but that's an area average, not a very accurate visual interpretation.

The Bayer works very well for situations where you either have:
*A solid colour background and a low-medium contrast luma detail
*A solid edge between two continuous colours, where each side have at least 3px support

This covers most of the image area in most images... But small pixel-level detail is usually only present in a small fraction of the total image area, so that 4-5% estimate is way off.

Using a much stricter set of parameters my own "average" estimate is close to 30% loss. I might have mentioned that a few times, but according to my own findings a bayer based image gets per-pixel PSNR figures approaching "perfect, noise limited" values when downsampling to about half MP amount, or 71% linear scale (29% loss). At that scale the best raw converters provides no reasonable further PSNR increase (noise effect subtracted of course) with further downsampling. That is: Further downsampling removes much more actual image detail than what it removes interpolation artefacts.

Depending on how you strict you want to be (and what the image contains of course!), that figure will vary between 10% and 35% loss IMO. 35% or even 40% for lots of small chroma detail, 10% for scenes with little or no chroma detail.

In the real world small detail is in almost all cases (except human made fabrics or weaves) either stochastic or fractal. Very few objects are as ordered and geometrically perfect (and still constant-chroma) as a black/white resolution target... So the line width resolution target is an unnaturally "perfect" condition for a raw converter. It can achieve much higher target>image per pixel PSNR here than in almost any conceivable natural target.

But if you accept lower chroma resolution, and SOME amounts of artefacts per image, Bayer losses can be said to be low in most imaging situations.



Aug 31, 2013 at 06:59 PM
Mirek Elsner
Offline
• • •
Upload & Sell: Off
p.1 #13 · p.1 #13 · Leica M9 vs MM resolution?


Excellent information, thanks! Would you care to comment on Foveon type of sensors? Do they have any resolution loss due to the design? Images from new Sigma cameras look so detailed - is this real detail that we are seeing, or result of image processing?


Aug 31, 2013 at 07:43 PM
theSuede
Offline
• • • •
Upload & Sell: Off
p.1 #14 · p.1 #14 · Leica M9 vs MM resolution?


It's mostly "real" detail, though the red layer is subjected to quite a lot of sharpening (helped by looking at the two other layers). The red is quite deep into the silicon surface, meaning that diffusion and reflections move quite a lot of the "red" hitting one pixel surface is measured by the surrounding pixels in stead.

When you combine this with an extremely low color separation, you can get detail over-attenuation as a natural part of the processing.

But that part is rather low in strength compared to a full Bayer interpolation. You don't lose any detail contrast, quite the opposite (though over-contrast is also a contrast "loss" from a PSNR or "error" point of view".
What Foveon processing (in general) does is to translate small changes in colour into changes in luma contrast. This, together with an artificial red sharpening is part of what makes the Foveon images appear more detailed than life (as long as you're not sensitive to color resolution).

Hue resolution is also quite low, but that seems to be less of a problem to most users. I would never recommend a Foveon for anything that requires colour accuracy though. They probably have the worst colour accuracy of any device I've ever measured north of 20$ phone camera modules.



Sep 01, 2013 at 06:08 AM
Mirek Elsner
Offline
• • •
Upload & Sell: Off
p.1 #15 · p.1 #15 · Leica M9 vs MM resolution?


Thanks, that connects the dots.


Sep 01, 2013 at 11:24 AM
Taylor Sherman
Offline
• • •
Upload & Sell: On
p.1 #16 · p.1 #16 · Leica M9 vs MM resolution?


Thanks for all the explanations! Also thanks Policar for the correction, right - Nyquist is 2 pixels.

Interesting to hear about the sharpening on the red channel for Foveon images. Most all of the Foveon pics I see in the DP2M thread here look over-sharpened to me.



Sep 01, 2013 at 12:51 PM
Mirek Elsner
Offline
• • •
Upload & Sell: Off
p.1 #17 · p.1 #17 · Leica M9 vs MM resolution?


Policar wrote:
Cute that Velvia appears sharper (with less resolution)!


If I remember correctly (and I may not, it has been almost 10 years since I shot film), Velvia has some 160lp/mm at 1:1000 contrast ratio and 80lp/mm at 1:1.6 contrast ratio. The Leica bodies resolve around 80lp/mm, correct?



Sep 01, 2013 at 01:29 PM
edwardkaraa
Offline
• • • • • •
Upload & Sell: Off
p.1 #18 · p.1 #18 · Leica M9 vs MM resolution?


Mirek Elsner wrote:
If I remember correctly (and I may not, it has been almost 10 years since I shot film), Velvia has some 160lp/mm at 1:1000 contrast ratio and 80lp/mm at 1:1.6 contrast ratio. The Leica bodies resolve around 80lp/mm, correct?


Yes, correct, except that 1:1000 contrast ratio is impossible to achieve with an optical lens.



Sep 01, 2013 at 01:33 PM





FM Forums | Leica & Alternative Gear | Join Upload & Sell

    
 

You are not logged in. Login or Register

Username       Or Reset password



This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.