Leica M sensor low light performance state of the art?

[tgm]

Advertisement (gone after registration)

Noise, Dynamic Range and Bit Depth in Digital SLRs -- page 4

 

 

Thanks a lot, but this is from 2008.

What about most recent full frame cameras, such as Nikon D800, D600 or Canon 6D or D1X? Do they also use apply on chip noise reduction?

 

Thomas

[Neimad]

One shot I took last weekend in New York (Brooklyn Bridge Park) with the Leica M and the 35Cron ASPH at F2 or 2,8 (not sure about the aperture)

I took it at 800 ISO and underexposed by one stop and pushed at processing with ACR.

 

[MirekE]

Noise, Dynamic Range and Bit Depth in Digital SLRs -- page 4

 

I am not sure if my procedure is exactly the same as the one used in the article, but if I generate FFT spectrum image from M240 exposed at 3sec and ISO 3200, the noise pattern appears even.

 

http://elsners.org/misc/raw-spectrum.png

 

I don't have a Nikon to compare, but I compared with Canon 5D2, same exposure and ISO setting. It also seems even except the center, even though I forced the camera to subtract black exposure.

 

http://elsners.org/misc/rawcomp-c.png

 

And here is the Leica image after processing in LR with noise reduction on. This one is from RGB, the two above are only the green channel.

 

http://elsners.org/misc/raw-filt.png

[mjh]

Thanks a lot, but this is from 2008.

What about most recent full frame cameras, such as Nikon D800, D600 or Canon 6D or D1X? Do they also use apply on chip noise reduction?

The specific noise suppression methods described by Emil Martinec are standard operating procedure for CMOS sensors; the output of a CMOS sensor would be pretty noisy (even at base ISO) if that weren’t the case. This hasn’t changed in recent years. The kind of noise reduction applied in a raw converter or in camera-internal image processing is quite different.

[tgm]

The specific noise suppression methods described by Emil Martinec are standard operating procedure for CMOS sensors; the output of a CMOS sensor would be pretty noisy (even at base ISO) if that weren’t the case. This hasn’t changed in recent years. The kind of noise reduction applied in a raw converter or in camera-internal image processing is quite different.

 

In the article several different aspects of noise reduciton are discussed, so what we are taking here about is noise reduction involving different pixels, "in-camera filtering of raw data". The article by Emil Martinec shows and the post #23 by MikeE show that some cameras apply this type of noise reduction (Sony A700) other do not (Leica M 240 and Canon 5D II), you can clarly see it in the 2D- Fourier transform. It think it is quite a strong statement saying that all, even the most recent cameras with CMOS apply this sort of noise reduction. The two examples above show that this is not the case.

 

Thomas

[JimKasson]

The specific noise suppression methods described by Emil Martinec are standard operating procedure for CMOS sensors; the output of a CMOS sensor would be pretty noisy (even at base ISO) if that weren’t the case. This hasn’t changed in recent years.

 

Let's set aside the long-exposure corrections, which are aimed at dealing with a particular kind of pixel response non-uniformity caused by diode leakage (dark current). If the camera is set at 1/30 or faster, this kind of processing will not be invoked.

 

I would quarrel with the assertion that "the output of a CMOS sensor would be pretty noisy (even at base ISO)" if there weren't in-camera noise reduction, unless the quote was meant to apply to long-exposure noise reduction.

 

I have very sucessfully modeled the Nikon D4 and D800E, and the Sony NEX-7 and RX-1, all CMOS cameras, with no noise reduction at all in the model. The only time the models showed material departure from the real cameras was, as I have posted earlier in this thread, the NEX-7 at ISO 6400.

 

Here are the model's parameters for the D800E. I believe the variable names are self-explanatory.

 

self.D65RedCorrection = 2751.275/4772.923;

self.D65BlueCorrection = 3471.922/4772.923;

self.D65RedWBCoef = 1 / self.D65RedCorrection;

self.D65BlueWBCoef = 1 / self.D65BlueCorrection;

self.quantumEfficiency = 0.56;

self.preAmpReadNoiseMean = 0.6;

self.preAmpReadNoiseStDev = 1.5;

self.postAmpReadNoiseMean = 0.9;

self.postAmpReadNoiseStDev = 1.5;

self.fullWellCount = 47510;

self.nPlanes = 4;

self.baseISO = 100;

self.nBits = 14;

pixelNUStd = 0.003;

 

The units for the read noise measurements are in least-significant bits. The units for the full well count is electrons.

 

One of the things that was surprising to me when I did this modeling work was how close the cameras, used at shutter speeds above 1/30, come to ideal, Poisson-noise-limited performance in the middle parts of their tone curves, above the point where read noise is important, and below the point where the shot noise has fallen to the point where PRNU is material.

 

Jim

[mjh]

Advertisement (gone after registration)

I would quarrel with the assertion that "the output of a CMOS sensor would be pretty noisy (even at base ISO)" if there weren't in-camera noise reduction, unless the quote was meant to apply to long-exposure noise reduction.

No, not long exposure noise reduction (that would be required with CCDs as well), but dealing with non-uniformity. Before Canon’s D30, CMOS sensors were widely believed to be incapable of delivering low noise images, but Canon had demonstrated that by dealing with CMOS-specific issues early on, CMOS could actually be a contender to CCD technology. I have seen the output of CMOS sensors not employing technologies such as double correlated sampling, and the result isn’t pretty.

[JimKasson]

No, not long exposure noise reduction (that would be required with CCDs as well), but dealing with non-uniformity. Before Canon’s D30, CMOS sensors were widely believed to be incapable of delivering low noise images, but Canon had demonstrated that by dealing with CMOS-specific issues early on, CMOS could actually be a contender to CCD technology. I have seen the output of CMOS sensors not employing technologies such as double correlated sampling, and the result isn’t pretty.

 

Upon examination of my model, I find that correlated double sampling, pattern read noise subtraction, and PRNU compensation could all exist and have the model work well, albeit with different parameters than a camera without these mechanisms. Since I derived my model parameters from testing actual cameras, my parameters include the effects of these three types of corrections. So, I, like the cameras, stand corrected. However, if the camera used the ISO setting in any way other than as gain to apply any of these corrections, then the model would deviate from the actual camera performance.

 

 

Jim

[JimKasson]

Upon examination of my model, I find that correlated double sampling, pattern read noise subtraction, and PRNU compensation could all exist and have the model work well, albeit with different parameters than a camera without these mechanisms.

Jim

 

As Martinec himself points out (using different language), these three are point processes, and do not trade resolution for SNR.

 

Jim

[JimKasson]

dwbell, don't understand why it's ironic when it's relevant: somewhere around page 5 or 6 of the thread you linked Jim Kasson talks about how the technique discussed in that thread (maximizing the M9's IQ by shooting at ISO 640 and pushing in LR4/5) could be applied to the M-Monochrom — he surmises that for the M-Monochrm the technique would probably be to shoot at 1250 and push in post. It would be interesting of he could test the M240 in this respect. I'd be interested in how Jim responds: it would be good if he could post here.

 

Initial testing indicates that the M240 is ISO-less.

 

http://blog.kasson.com/?p=3138

 

More to come.

 

Jim

[dwbell]

Great that you're on it already Jim, much appreciated.

 

Couple of thoughts;

 

I really do hope they haven't added noise to remove banding. If we go to a push in post workflow on the M240 that banding could have been easier to remove than the noise hiding it. All speculation on my part of course.

 

Maybe the specific in-camera noise reduction of Sony, Fuji etc is where their advantage lies, as they know specifically what they need to remove and have the capacity to do so? Perhaps an area where Leica would need to invest / improve / develop to remain competitive?

 

Looking forward to further instalments and thanks again for making your findings public.

[tobey bilek]

Noise reduction is a skill and many photogs do not have it or like having it done for them. Plus Nikon can boast their sensor as being low noise when in fact it is in camera software. So if you did 1000 photos at night football, it saves a step.

 

The downside of noise reduction is loss of sharpness, they are polar opposites regardless of how it is done. If you did it yourself, you can apply an edge mask in photoshop to preserve edges or brush it in certain areas as required. I would not want to do this with too many photos regardless of superior results and there is no way to batch process it. Can you even make an edge mask in Lightroom? I have not seen that tool.

 

I think Leica has taken the better road. Nikon/Canon should at least provide a "off switch".

 

But then if you are building cameras to pass tests, no switch is better.

[dwbell]

Good points indeed, but don't they kind of assume that the photographer, replete with the appropriate skills, will do a better job than that of the huge multi-million dollar tech giant who can R&D the hell out of their own sensor's output and fix what's wrong before the user even sees it? Maybe there's more to it than the simple kind of noise reduction vs sharpness most of us deal with?

[JimKasson]

The downside of noise reduction is loss of sharpness, they are polar opposites regardless of how it is done.

 

As I said earlier, correlated double sampling, pattern read noise subtraction, and PRNU compensation are point processes, and do not trade resolution for SNR. By the way, none of these techniques are used in post. Well, uh, PRNU compensation can, but only if the photographer is willing to average a bunch of images to get the compensating image. I've done it using a D800E and 128 images, and it works, but requires writing code or using astronomical software, takes a lot of time and is not for the faint of heart. Pattern read noise subtraction needs a positive offset before the ADC to work properly. Canon does that, but not Nikon. Again, to do it in post requires a compensating image. There is no way to do correlated double sampling in post.

 

I think Leica has taken the better road. Nikon/Canon should at least provide a "off switch".

 

I don't know abut Canon, but neither the D800E nor the D4 show evidence of noise reduction that compromises resolution in their raw files, at least as far as I can see. Do you have any evidence that these cameras use that kind of noise reduction?

[JimKasson]

In the moderately bright parts of the image, there is a small benefit to incresing the ISO setting on the camera if you can't get a proper exposure at base ISO because you can't use that slow an f-stop and/or that wide an aperture.

 

I think the shadow data is more important, since that's usually where the noise is visible; maximizing the mid/high lever SNR is solving a problem that doesn't usually exist.

 

Details here.

[viramati]

Ok I'm officially lost. the main problem I find with using the higher iso settings is banding in the shadow areas so is this reduced by keeping the iso down and pushing processing the image?

[JimKasson]

In the deep shadows, turning up the ISO from 100 to 3200 creates about a stop more noise -- twice as much -- as making the correction in your raw processor.

 

Characterizing the Leica M240, part 5 | The Last Word

[JimKasson]

Ok I'm officially lost. the main problem I find with using the higher iso settings is banding in the shadow areas so is this reduced by keeping the iso down and pushing processing the image?

 

I will come up with a test and post the results. Warning: you'll have to look at the images and make your own judgement; I don't feel competent to devise a mathematical function to evaluate banding. The images may end up being quite large.

 

Jim

[dwbell]

 

Are pictures of your bookcase coming Jim?

[Advertisement]

Hallo,

Look here: Leica M

[JimKasson]

 

Are pictures of your bookcase coming Jim?

 

They were going to be next, before I got to thinking about a test for banding. Maybe I can do both today.

 

Jim