Native ISO's on M9?

[fates]

So here's a question that I haven't found answered out on the net, are certain ISO settings native to the M9 chip (i.e. requiring no processing by camera to achieve value)?

For example the 5D has native ISO settings that have less noise than slower values! On a 5D, 640 is actually cleaner than 500. The official native ISO values haven't been discussed by Canon, I don't believe, but the studies basically show this pattern:

ISO 100 (ISO 160 - digital pull of -1/3 EV)
ISO 160
ISO 240 (ISO 160 + digital push of +1/3 EV)
(etc...)

So I'm wondering does the same thing hold true for the M9 senor? Has anyone tested this. Shooting a closed lens and pixel peeping the noise on black would show if this is the case...

[kzphoto]

I haven't heard of this phenomenon before. It would be interesting to have a full list of values for different cameras from different manufacturers.

[jaapv]

Native ISO equivalent of both M8 and M9 is 160.

You can find this value both in Leica publications and the Kodak factsheet for the sensor.

The ISO 80 setting is marked on the M9 as "pull".

The value is of most interest for the max dynamic range. In daily practice it is not very relevant.

Note -sensitivity is defined on the relevant ISO sheet for film only - sensor sensitivity is defined as being comparable to the equivalent film sensitivity - leaving the method of measurement unspecified. This means each manufacturer can define the sensor sensitivity differently. Thus "Canon ISO" does not neccesarily equal "Leica ISO" etc. (In fact this has been remerked on often by reviewers) For the Leica M8 the real ISO value in daily use, for instance for flashes or exposure meters can best be regarded as 200, for the M9 the setting is closer to 160, probably because Leica shifted the black point.

Btw, to be pedantic "needing no processing" is an impossiblity. You cannot transform the stream of electrons coming out of the sensor into a measurable value without processing, and with a CMos the processing and amplification starts on the sensor on a per-pixel basis.

On the M9 (and M8) shooting a closed lens would not work . As soon as the shutterspeed exceeds one second, the camera will do a second exposure black frame subtraction to achieve heat-generated noise elimination. This cannot be switched off.

[sabrewolf]

http://www.dxomark.com/index.php/Cam...nsors/Leica/M9

[sojournerphoto]

BAse iso on the M9 is 160 as Jaap says. If you look at DXomark their methodology places this at iso 144 (the same as iso 200 on a Canon 1Ds3). I suspect the DXO approach measures iso using a well saturation approach implying that the Canon has greater highlight headroom for a given iso than the Leica. Both have similar dtnamic range at this point.

Higher isos on the Leica appear to lose dynamic range on a stop for stop basis, suggesting that higher isos are implemented primarily by multiplying the raw numbers. This is unlike the canon's cmos sensor where the sensor output voltage can be multiplied to maintain higher DR for longer (do a DXo comparison). The latter approach could be considered to be a real increase in sensitivity.

On Canon's lower level camera's get 'real' iso increases for whole stops and digital for intermediate values. On pro series camers real isos are implemented for all values. Not sure which category the 5D fell into.

One corollary of different approaches (and their effectiveness) to implementing high iso is that sometimes it's better or no worse from a noise perspective to underexpose a lower iso and push in processing than to shoot at a higher camera iso. This always helps to protect highlights and may reduce noise (Canon 1Ds3 above iso 800 for example). I suspcet that the M9 is a bit like this and underexposing at 160 or 320 may not look much or any worse than exposing at 1250 or 2500, and will give you highlight protection. However, before I'm flamed, I haven't tested this thoroughly yet and am happy to be corrected or informed. (I did do one quick test using equal exposure at 160 and 1250 and there was no obvious difference in optimised files at 100% on screen, but there was plenty of light for the 1250 exposure and the scene was no high in contrast)

Mike

[Brian Sweeney]

From a firmware point of view, the "native" ISO equivalent of a sensor is the sensitivity when the gain of the pre-amp between the detector and A/D convertor is set to 1x. The A/D circuits that i have used are programmable, and you can set the gain as a factor. Higher ISO settings were attained by setting the gain of the pre-amp higher, to boost the signal before it hit the A/D convertor. Of course, this can vary based on the overall system design. I can imagine systems using a 22-bit A/D convertor and throwing away bits instead of adjusting analog gain. Higher gain settings usually means boosting noise. Noise in a digital image can be processed out. The M8 and M9 have a minimum of signal processing algorithms in the camera, DSLR's using CMOS sensors apply more signal processing to the image before it is stored. Images can always be post-processed for noise reduction outside of the camera. Such processing will yield more benefit to images from the M8 and M9 as they do not apply it within the camera.

[sojournerphoto]

Originally Posted by Brian Sweeney

From a firmware point of view, the "native" ISO equivalent of a sensor is the sensitivity when the gain of the pre-amp between the detector and A/D convertor is set to 1x. The A/D circuits that i have used are programmable, and you can set the gain as a factor. Higher ISO settings were attained by setting the gain of the pre-amp higher, to boost the signal before it hit the A/D convertor. Of course, this can vary based on the overall system design. I can imagine systems using a 22-bit A/D convertor and throwing away bits instead of adjusting analog gain. Higher gain settings usually means boosting noise. Noise in a digital image can be processed out. The M8 and M9 have a minimum of signal processing algorithms in the camera, DSLR's using CMOS sensors apply more signal processing to the image before it is stored. Images can always be post-processed for noise reduction outside of the camera. Such processing will yield more benefit to images from the M8 and M9 as they do not apply it within the camera.

Hi Brian,

Of course you're right and if the readout, amplification and AD conversion were noise free then chopping bits would be equivalent to increasing agin by factors of 2.

In the (14 bit) Canon's case, the noise floor is well above the resolution of the ADC and the well capacity is such that increasing gain does not reduce dynamic range on a factor for factor basis, i.e. the dynamic range curve is downward sloping with increasing iso, but the loss is much less than one stop per doubling of iso for the range we are talking about.

The DXo curve for the M9 looks, at a glance, to be pretty well 'linear' at 1 stop for each doubling of iso (yes onlog scales I know!), hence my suggestion that iso may be implemented by bit chopping. Of course there are other possible explanations, and I should have a think about whether they make a practical difference in real world shooting and processing.

I would be interested in yuor thoughts though?

Mike

[jaapv]

I have always wondered what method DxO uses to obtain their data. I find it scarcely credible that they would have taken all these cameras apart to get at the pure sensor readout. If the values are taken from the raw files considerable in-camera processing has already taken place, making their conclusions less valid.

[jaapv]

Originally Posted by sojournerphoto

.



One corollary of different approaches (and their effectiveness) to implementing high iso is that sometimes it's better or no worse from a noise perspective to underexpose a lower iso and push in processing than to shoot at a higher camera iso. This always helps to protect highlights and may reduce noise (Canon 1Ds3 above iso 800 for example). I suspcet that the M9 is a bit like this and underexposing at 160 or 320 may not look much or any worse than exposing at 1250 or 2500, and will give you highlight protection. However, before I'm flamed, I haven't tested this thoroughly yet and am happy to be corrected or informed. (I did do one quick test using equal exposure at 160 and 1250 and there was no obvious difference in optimised files at 100% on screen, but there was plenty of light for the 1250 exposure and the scene was no high in contrast)

Mike

It might work that way for uncompressed M9 files, as you say it needs testing. On the compressed files of M9 and M8 I doubt the results would be worthwhile, as the compression is non-linear.

[sojournerphoto]

Originally Posted by jaapv

It might work that way for uncompressed M9 files, as you say it needs testing. On the compressed files of M9 and M8 I doubt the results would be worthwhile, as the compression is non-linear.

Indeed. I'd forgotten the compressed option and just considered the uncompressed dng files.

[fates]

Originally Posted by Brian Sweeney

From a firmware point of view, the "native" ISO equivalent of a sensor is the sensitivity when the gain of the pre-amp between the detector and A/D convertor is set to 1x. The A/D circuits that i have used are programmable, and you can set the gain as a factor. Higher ISO settings were attained by setting the gain of the pre-amp higher, to boost the signal before it hit the A/D convertor. Of course, this can vary based on the overall system design. I can imagine systems using a 22-bit A/D convertor and throwing away bits instead of adjusting analog gain. Higher gain settings usually means boosting noise. Noise in a digital image can be processed out. The M8 and M9 have a minimum of signal processing algorithms in the camera, DSLR's using CMOS sensors apply more signal processing to the image before it is stored. Images can always be post-processed for noise reduction outside of the camera. Such processing will yield more benefit to images from the M8 and M9 as they do not apply it within the camera.

Very informative. Thanks Brian! So what I'm gleeming from this is that 500 ISO is less noisy from 640 ISO as the gain applied to reach the 500 ISO is linear and not a processed step down from analogue gain applied to the sensor at the 640 ISO level. Am I correct on this?

I've noticed while shooting with the auto ISO on for a bit that the camera (almost) always preferred increments of 160 (160, 320, 640, 800 for some reason, and 1250)... this is why I started questioning the possibility of this sensor having a step processing system like the 5D does...

[Brian Sweeney]

In order to answer this question in more detail, I will simply have to buy an M9 to conduct further tests. I am looking forward to the 16-bit uncompressed raw images and manual lens selection.

I worked on a digital sensor in 1983 that used 14-bit A/D converters, spec'd with 1/2 bit of noise. That was very tricky to implement "way back then" and the first electronics box induced 9 bits of noise. Lots of rework to beat that down. It was easier to comment on these things when you were in on the design. I still have my FFT code on the computer.

[sojournerphoto]

Originally Posted by Brian Sweeney

In order to answer this question in more detail, I will simply have to buy an M9 to conduct further tests. I am looking forward to the 16-bit uncompressed raw images and manual lens selection.

I worked on a digital sensor in 1983 that used 14-bit A/D converters, spec'd with 1/2 bit of noise. That was very tricky to implement "way back then" and the first electronics box induced 9 bits of noise. Lots of rework to beat that down. It was easier to comment on these things when you were in on the design. I still have my FFT code on the computer.

Brian, I think the M9 has a 14 bit ADC, but delivers files in 16 bit format. Photoshop truncates to 15 bits in any case

Mike

[Brian Sweeney]

Well then, I am glad to know that the system we built 28 years ago has as much depth to the pixels as my new M9. "Bleeding Edge" of technology.

So are the upper two bits set to zero? I guess Photoshop uses 15bits with one sign bit, 16-bit Signed Word data format. INTEGER* 2.

[sojournerphoto]

Originally Posted by Brian Sweeney

Well then, I am glad to know that the system we built 28 years ago has as much depth to the pixels as my new M9. "Bleeding Edge" of technology.

So are the upper two bits set to zero? I guess Photoshop uses 15bits with one sign bit, 16-bit Signed Word data format. INTEGER* 2.

To be honest I don't know which bits they set to zero. I suspect you could glean it from the dng spec, which is based on tiff of course. PS may well use 16 bit signed word format, though a negative value wouldn't make much sense

Mike

[Brian Sweeney]

On the KAF-10500 long sheet, the Saturation per pixel is 60,000 electrons and the nominal total noise figure (includes amplifier and system noise) is given as 18 electrons. So I make out a 3000:1 ratio between saturated pixel and a dark pixel.

http://www.kodak.com/global/plugins/...00LongSpec.pdf

We used to do a lot of frame-to-frame subtraction for image processing. I can see why Photoshop would use a signed value.

Just to add: I have read that the KAF-18500 used in the M9 picks up about a stop improvement. With that in mind, it seems well-matched to a 14-bit A/D.

[fates]

So with all this science, I went ahead and conducted a very 'detailed' test. I shot every ISO with the lens cap on at 1/30 (so there would be no post-process Jaapv) and looked at the pixel noise. There is a linear ramp up in noise with each ISO and gain in color noise and red color bleed. It gets ugly past 1250, but nothing that jumps out at me as a lower value like 1000 being worst than 1250 like it is on a 5D, so it would stand to reason that there is no secondary processing of the signal, as the gain is increased in real increments.

[Brian Sweeney]

That sounds good- allows the user to select the noise reduction algorithms best suited for the individual image rather than rely on a one-size-fits-all in-camera algorithm. It's a feature, not a flaw.

I did a really nice noise-reduction algorithm for my wife. She was taking an image processing class for medical imaging at the time. My "Honey-Do" list usually meant writing code.

[jaapv]

As a matter of fact this confirms leica's statements on noise reduction, i.e. none except for minor corrections at 2500.

[back alley]

i want to change the title of this thread to...when camera geeks talk shop...or something like that!!

[Brian Sweeney]

That would be "when LEVEL 6 camera geeks talk shop".

I could list assembly language source code for changing the pre-amp gain on the A/D that I used. I cached the settings and speeded up the conversion rate dramatically. But it could result in spontaneous non-geek combustion.

[back alley]

Originally Posted by Brian Sweeney

That would be "when LEVEL 6 camera geeks talk shop".

I could list assembly language source code for changing the pre-amp gain on the A/D that I used. I cached the settings and speeded up the conversion rate dramatically. But it could result in spontaneous non-geek combustion.

my deepest apologies...as i understood absolutely nothing said it was impossible for me to accurately gauge what geek level you chaps were at...

[Brian Sweeney]

I rely on that fact. I make this stuff up as I type.

[jaapv]

We must consider one thing, On the M9 Leica has shifted the definition of the black point in the DNG parameters to get a noise reduction (just like Canon does I believe) so the output is biased.

[Brian Sweeney]

Do they truncate the output of the A/D converter, or just set a parameter in the DNG header to define a noise floor? If it the latter, the parameter can be ignored. A noise-reduction algorithm can simply operate on the raw values.