Is there a reason for the blue/cyan bias on the Sony A7RII images?

[ErikKaffehr]

Hi Jim,

Thanks for posting a good link!

For me, the separation of purplish colours is a bit of a mythical area, and I would guess the L (red) channel sensitivity that Jack mentions may play a role in that.

My guess is that spectral response has very little to do with Sony's spectral characteristics, which are probably the same as on the Phase One Q-250 or the Pentax 645Z, but very much depending on camera profiles. Still, I think that camera white balance plays the uttermost role. Number two suspect is camera DCP profile. There are plenty of those, all implementing a different look, and yes I would guess some are better than others. But, to say better, you need define good.

Best regards
Erik

Whenever someone quotes W&S, you know they're a force to be reckoned with. Well played, Jack.

I remember a meeting with a Kodak Research color scientist at Almaden Research Center. First thing he dds was pull a copy of W&S out of his briefcase and set it on the table. Then we talked.

A little background:

http://blog.kasson.com/?p=12462

The big reason that cameras don't use overlapping red and green filters like the rho and gamma cones is SNR. Takes a lot of subtraction to get from there to, say, Adobe RGB.

Jim

[eronald]

All work and no play makes Jack a dull boy   Ok, so on the work side and for my edification:  since we know that the a7RII has good - in fact class-leading - DR after the CFA at 'ISO push' compared to other current advanced DSCs, does it just boil down to the orthogonality of the CFA curves?  How does one define orthogonality of the curves?  Assuming their shapes are pretty typical would it be related to the position and distance of the three peaks, and their half power bandwidths?  If so and since we are at it, how does the human visual system discriminate green so well with the apparent lack of orthogonality in B)?  Any help by anyone would be appreciated.

Jack

Jack,

 Everyone here is smarter than me these days, and any *scientifically*reasoned debate would require me doing at least one week of reading source material. Also, I don't think most of the issues here are about physiological color, assuming there actually exists one unique average individual. The issues arise because of the desire of using three-filter-Bayer which is a tech decision which will doubtless go the way of the 78rpm record in due course.

 If you actually want to discuss Bayer camera color, then concerning DR, greenery will absorb most of the reds (chlorophyll) and if you are already standing under a forest canopy under a blue or white sky high CT sky as illumination there will simply not be a lot of red signal for discrimination. Instead of trying to show the world how dumb I am, why don't you use the spectral and camera data you have at your fingertips and a Raw image to see how much actual red light is left and what the data looks like before in-camera multipliers are applied? My guess is that most of the "real" red signal is lost in the noise, although camera "R" whose filter probably looks orange  will register a decent reading.

 Sony had a camera with decent landscape color, I think it was the A950 or something, it did badly commercially because the filters cut down the ISO too strongly, but may have been the best landscape SLR of recent times.

 One can see this game being played in churches - go to a medieval european church and it will be dark and the stained glass will make you swoon - go to a modern church and a lot more light gets in the windows, which convey little emotion.

Edmund

[Jack Hogan]

These curves would look quite different if weighted by the population of each type of cone.

Ah, is that built into the luminosity function?  And would you simply multiply the curves above by V(lambda) before normalization in order to obtain what the human visual system measures?

The gamma cones contribute a lot to our perception of color. They are one axis of three. The beta cones contribute little to luminance, but a great deal to the opponent color (red/green and blue/yellow) results. We wouldn't have a red/green axis without the gamma cells.

BTW, with a lens that has lousy LoCA, it makes a lot of sense for the eye to put most of the luminance sensing in one part of the spectrum.

Jim

Makes sense, thanks Jim.  So what's your take on the a7RII's color discrimination compared to others out there?

Jack

[Jack Hogan]

Jack,

 Everyone here is smarter than me these days, and any *scientifically*reasoned debate would require me doing at least one week of reading source material. Also, I don't think most of the issues here are about physiological color, assuming there actually exists one unique average individual. The issues arise because of the desire of using three-filter-Bayer which is a tech decision which will doubtless go the way of the 78rpm record in due course.

 If you actually want to discuss Bayer camera color, then concerning DR, greenery will absorb most of the reds (chlorophyll) and if you are already standing under a forest canopy under a blue or white sky high CT sky as illumination there will simply not be a lot of red signal for discrimination. Instead of trying to show the world how dumb I am, why don't you use the spectral and camera data you have at your fingertips and a Raw image to see how much actual red light is left and what the data looks like before in-camera multipliers are applied? My guess is that most of the "real" red signal is lost in the noise, although camera "R" whose filter probably looks orange  will register a decent reading.

 Sony had a camera with decent landscape color, I think it was the A950 or something, it did badly commercially because the filters cut down the ISO too strongly, but may have been the best landscape SLR of recent times.

 One can see this game being played in churches - go to a medieval european church and it will be dark and the stained glass will make you swoon - go to a modern church and a lot more light gets in the windows, which convey little emotion.

Edmund

Edmund, I mean all this in jest, as a playful learning opportunity.  I apologize in advance if my questioning style came across as trying to demonstrate anything other than I do not know how one goes about figuring out if a camera has good color discrimination or not.  So when you mention that the A950's filters 'cut down the ISO too strongly' do you mean that the normalized individual channel curves have narrower bandwidths compared to newer designs?  Here for instance are the cone sensitivity curves above (dashed lines) resized to the same height and shown on top of the CFA curves from the sensor in the paper linked by Brian, which is used in the M9 I understand (solid lines)



Is discrimination here all about slimming down and shifting left that big fat blue?

Jack

[AlterEgo]

True enough.

However, I got the same results with ASP"

http://blog.kasson.com/?p=12524

and C1 generic:

http://blog.kasson.com/?p=12573

Jim

you mean same results for greyscale patches when you set WB using one of them in both converters...

[Jim Kasson]

you mean same results for greyscale patches when you set WB using one of them in both converters...

No, I meant no systematic cyan cast to the other patches after the image was WB'd to one of the neutral patches.

Jim

[BrianVS]

With regard to the spectral response curves in the Sony links that I posted: I expect a manufacturer to post manufacturers specifications for the product and not some randomly chosen unit.  Numbers given should be representative of production with published numbers for tolerances and variations. The spectral curves given should be for the average production sensor. Kodak, OnSemi, and other manufacturers that target the scientific and technical market tend to publish these numbers. Typical, minimum, and maximum values are in the spec sheet for a given product. Some samples are better than others, and cherry-picking the best of the best usually yields better performance. The KAF-18500 (sensor in the Leica M9) specs indicate a nominal 1.8% red/green and green/blue hue shift, and the note in the spec states that this parameter is measured for every sensor during production testing. The spec also states a Max shift of 12%- hopefully those go into the "engineering samples". If QE peaks really wandered around by as much as 20nm, each sensor would have to be carefully calibrated to yield consistent colors. I believe the spec is much tighter. I used to measure such things, a very long time ago. Multi-Spectral digital imagers, 1980s.

[eronald]

With regard to the spectral response curves in the Sony links that I posted: I expect a manufacturer to post manufacturers specifications for the product and not some randomly chosen unit.  Numbers given should be representative of production with published numbers for tolerances and variations. The spectral curves given should be for the average production sensor. Kodak, OnSemi, and other manufacturers that target the scientific and technical market tend to publish these numbers. Typical, minimum, and maximum values are in the spec sheet for a given product. Some samples are better than others, and cherry-picking the best of the best usually yields better performance. The KAF-18500 (sensor in the Leica M9) specs indicate a nominal 1.8% red/green and green/blue hue shift, and the note in the spec states that this parameter is measured for every sensor during production testing. The spec also states a Max shift of 12%- hopefully those go into the "engineering samples". If QE peaks really wandered around by as much as 20nm, each sensor would have to be carefully calibrated to yield consistent colors. I believe the spec is much tighter. I used to measure such things, a very long time ago. Multi-Spectral digital imagers, 1980s.

Brian,

 Too many of us here are now dinosaurs, with knowledge of what went on before, but little knowledge of what is done now in the industry.

One of the differences between cheap and pricey camera product lines is how much calibration and testing goes into individual product samples. I believe it is usual to test in order to establish/confirm parameters for the whole production batch of a sensor.

Dietmar Wueller's Image Engineering make a gadget that spits out a spectral sensitivity curve from one camera click - and computes an ICC profile.  The first version of this was based on a slide projector with a stabilised power supply and a bunch of narrowband filters, and needed several clicks. As you know one can also aim a monochromator at the camera and get the same result, this is done at universities routinely, and some do it at home. Jack Holm when he was at at HP had a patent on an active colorchecker thingy based on a collection of LEDs, I'm sure one could make one of those at home too.

I think we could have an effective and productive discussion of what is going on with current cameras, if we chose to get around the showoff factor, and made an effort to get some hard data. As I said before this means collecting real *experimental* spectral curve data, agreeing on some measure of how bad a given Luther-Ives failure is, and validating this metric with real world setting descriptions like "under forest canopy" or  "2700K incandescent", and the spectral and S/N data of known cameras AND SCENES.

Any discussion without agreed observer functions, knowledge of how people evaluate scenes, scene measurements, and hard camera data is IMHO a waste of our times; in fact I am sure that  if you ask a quadrichromat human she will laugh at your evaluations - but is she representative?

Edmund

[BrianVS]

Brian,

 Too many of us here are now dinosaurs, with knowledge of what went on before, but little knowledge of what is done now in the industry.


Edmund

I work in the scientific/technical field, manufacturers that cater to that market publish data sheets that are representative of actual sensor performance. It's different from consumer grade sensors.  CCD's are still used, the last camera chosen was based on the increased response in the near IR- as per the spectral response curve given by the manufacturer.

https://www.hamamatsu.com/us/en/C3077-80.html

Compare the response with the monochrome Nikon microscope camera that uses the same sensor in the Df. The Nikon was rejected due to poor performance in the NIR.

http://www.nikon.com/products/instruments/lineup/bioscience/camera_microscopy/dsqi2/index.htm

The spectral response curve gives a very good indication as to the performance of a detector. Detector geometry and chemistry affect spectral response. BSI changes the geometry and more blue light will be absorbed at the photosensitive portion of the detector.

A monochrome BSI sensor aimed at the scientific/technical market should have some good spectral response data available, should make it easier to compare the new technology with FSI sensors. I suspect that BSI sensors will have increased response at the short end of the spectrum. It will have very poor IR response. Good for photography, not so good for NIR applications.

[mcbroomf]

I enjoy threads like this even if I don't follow all of the tech.  I wanted to comment on the way FSI chips are made.  I may be mistaken but I think there are assumptions that the light has to pass through some silicon on the way to the detector (which would mean absorption of blue more than green, then red.  In fact the wiring layers above the sensor surface are suspended in silicon oxide.  The oxide will most likely be fluorine or carbon doped to reduce the capacitance between metal lines that run close to each other.  Of course there will still be some light absorption and scatter, but nothing like having to pass through silicon, and I don't know the spectral qualities of these oxides to know if blue is affected more.  On top of the last oxide layer before the CFA there will be a silicon nitride passivation layer to seal the surface from moisture and although the RI is higher than oxide it will be relatively thin.

[BrianVS]

I looked it up- a layer of silicon dioxide (SiO2) is used. Slightly better transmission at longer wavelengths.

I also found some papers on a hybrid CCD/CMOS sensor aimed at better Near Infrared response.

Some articles state that thinned sensors offer higher QE in the visible band at the expense of longer wavelengths. Thicker photo-sensitive sites are subject to recombination, loss of signal. QE peaks shift to shorter wavelengths as the photosensitive layer is thinned. IR response is reduced, which is considered an advantage for photography.

This company gave some numbers for absorption of light as it passes through silicon. I did not realize the difference in penetration depth varied to this degree. I worked with Midwave and Longwave IR, 3um~5um and  9um~12um for the most part, Visible+NIR CCD's.
 
http://www.aphesa.com/documents.php?class=wp

"Spectral Response of Silicon Sensors" available for download.

Two datasheets for devices from the same family- both CCD's with 100% fill-factor. The shift in QE towards shorter wavelengths (and reducing NIR) and some of the work that went into achieving the results is documented:

http://www.astrosurf.com/audine/pdf/kaf1600.pdf

http://www.onsemi.com/pub_link/Collateral/KAF-1603-D.PDF

[eronald]

I think the FSI issues stem from the fact that the various orthogonal flow Al metal signal paths on top block light from reaching the gate which is below them. The sensel looks like a funnel. BSI has the added advantage that one can grab signals from the middle of the chip rather than the edges, allowing higher throughput.

I think a color discussion that went somewhere would be useful, because we could turn it into sensor validation code.

Edmund

I enjoy threads like this even if I don't follow all of the tech.  I wanted to comment on the way FSI chips are made.  I may be mistaken but I think there are assumptions that the light has to pass through some silicon on the way to the detector (which would mean absorption of blue more than green, then red.  In fact the wiring layers above the sensor surface are suspended in silicon oxide.  The oxide will most likely be fluorine or carbon doped to reduce the capacitance between metal lines that run close to each other.  Of course there will still be some light absorption and scatter, but nothing like having to pass through silicon, and I don't know the spectral qualities of these oxides to know if blue is affected more.  On top of the last oxide layer before the CFA there will be a silicon nitride passivation layer to seal the surface from moisture and although the RI is higher than oxide it will be relatively thin.

[BrianVS]

Agreed that the biggest gain in BSI vs FSI is the angle of acceptance, being able to use large aperture lenses and mounts with short registers. I'm waiting for Leica to put one into the M-Mount camera.

The original question was in regard to a color bias. I used a set of Nikon color filters ranging from Blue through to deep Red arranged on a backlit Slide Sorter to compare the M Monochrom with the DCS200 (KAF-1600). Nikon publishes the response curves for the filters, but that is not as important for detecting a difference between two cameras. I like the slide sorter as the light source is behind the filters. Such an arrangement should be useful for comparing two cameras and testing if the new camera has a bias towards Blue. Look at Raw RGB values from cropped areas within the color filters, compare histograms between the two cameras. Color filters are dirt cheap these days, as are slide sorters. It would also be interesting to put a graduated grey scale card up on a copy stand to test differences between two cameras. Again- easy DIY projects to do at home, cheap.

[jrp]

Imaging Resource test shows that

Like many cameras, the Sony A7R II shifts cyan toward blue, red toward orange, orange toward yellow and yellow toward green, but shifts are relatively minor. (The cyan to blue shift is actually fairly minor and very common among the digital cameras we test; we think it's a deliberate choice by camera engineers to produce better-looking sky colors. The yellow to green shift combined with its desaturation is however unfortunate, as it can produce somewhat dingy-looking yellows.) With an average "delta-C" color error of 5.45 after correction for saturation at base ISO, overall hue accuracy is about average, with accuracy only varying slightly at higher ISOs.

[AlterEgo]

Like many cameras, the Sony A7R II shifts cyan toward blue, red toward orange, orange toward yellow and yellow toward green, but shifts are relatively minor. (The cyan to blue shift is actually fairly minor and very common among the digital cameras we test; we think it's a deliberate choice by camera engineers to produce better-looking sky colors. The yellow to green shift combined with its desaturation is however unfortunate, as it can produce somewhat dingy-looking yellows.) With an average "delta-C" color error of 5.45 after correction for saturation at base ISO, overall hue accuracy is about average, with accuracy only varying slightly at higher ISOs.

again this just reflects a specific raw conversion with specific profile... change that and you will get different results... average delta-c = "5.45" ?


http://1.bp.blogspot.com/-H3-8D5Sdd5k/VoLFnxGMtJI/AAAAAAAAKf4/xV6N5zJbc6A/s1600/ACR20151229-01.jpg



little modded Adobe Standard and adding some red primary saturation (+20) in ACR with raw file from I-R and CC24 Classic measurements (30 samples from BabelCOlor)... process 2010 naturally

[ErikKaffehr]

Hi,

I am not sure that manufacturers tweak CFA-s a lot to improve high ISO, it can happen from time, but the only way to know is to see the actual curves.

Now, getting back to the Alpha 950, that camera did not even exist, but I am pretty sure that the A900 was meant. And yes, it was known to have good colour. And that was confirmed by some folks I am aware of, namely Tim Parkin, the publisher of On Landscape and Iliah Borg of LibRaw and RawDigger fame.

I have three interesting cameras in context, so I could make a decent test. Something along these lines:

• Build a daylight emulator and an A-illuminant simulator
• Make very good shots of a ColorChecker Passport, avoid surround light and glare with all three cameras
• Build DCP profiles for  each
• Now find an adequate target and shoot carefully under constant and repeatable conditions - somewhat problematic for a guy like me who shoots landscapes mostly - include a white balance card or ColorChecker in all shots
• Develop all images with identical setting, using the DCP profiles we generated for each camera
• Do an objective evaluation of colour

[/list

The last point is a bit tricky, unless we shoot test charts.

A year ago I did a comparison between my P45+ and my Sony Alpha 99, in part related to a discussion between me and Tim Parkin. Tim deeply dislikes the colour rendition of the P45+ while he likes that of Sony Alpha 900. What I wanted to see was to what extent raw converter profiles affected colour rendition.

For this comparison a flower with bluish purple petals was chosen. The reason was Tim stated that reproduction of chlorophyll was yellowish on the P45+ and I have previously observed that purples were problematic to reproduce.

I also used my ColorMunki to sample both the bluish purples and the blade greens, so I have a good colour reference.

Here is a link to that comparison: http://echophoto.dnsalias.net/ekr/Articles/OLS_OnColor/SimpleCase/

My surprise was that my P45+ produced nice bluish purple with Lightroom - almost identical to the measured samples - while Capture turned the petals into blue. With the Sony Alpha 99 the situation was pretty similar - Capture One turned bluish purple into blue. This really strengthens my conviction that colour profiles play a greater role than camera CFA:s.

When looking at my measured patches I have noticed that both blade and petal had extremely high infrared content. The ColourChecker patches don't have excessive IR. I guess that IR filtering on the sensor may play a role, as IR affects the red channel. We know that nature greens reflect a lot of IR, light vegetation is one of the characteristics of IR images.

Profiles also twist hues, based on luminosity, here is an article describing it: http://chromasoft.blogspot.se/2009/02/visualizing-dng-camera-profiles-part-3.html

Sandy McGuffog, the author of the above article has a command line tool to modify the way the dcp-profiles handle these twists.

Just to say, I have checked my Adobe Standard Profiles for colour shifts with varying exposure, and they are there. Sandy's tools can sanitise that, but I have not seen any real issue in my work.

Best regards
Erik

Edmund, I mean all this in jest, as a playful learning opportunity.  I apologize in advance if my questioning style came across as trying to demonstrate anything other than I do not know how one goes about figuring out if a camera has good color discrimination or not.  So when you mention that the A950's filters 'cut down the ISO too strongly' do you mean that the normalized individual channel curves have narrower bandwidths compared to newer designs?  Here for instance are the cone sensitivity curves above (dashed lines) resized to the same height and shown on top of the CFA curves from the sensor in the paper linked by Brian, which is used in the M9 I understand (solid lines)



Is discrimination here all about slimming down and shifting left that big fat blue?

Jack

[/list]

[Jack Hogan]

That would be an interesting test, Erik.  I would be interested to hear what Anders Torger thinks about all this...

Jack

[ErikKaffehr]

Hi Jack,

I am doing some preparation for this, like getting Solux bulbs and a regulated voltage power supply. Unfortunately I lost my ColorChecker Passport, so I need to order a new one.

I want to make a profile for my Sony A7rII, so looking into the other cameras is not a lot of work.

The cameras I have right now is the A900, a P45+ and the A7rII. For completeness I also have a Minolta Dimage 7D and Sony A99 and Sony A77, I plan to sell those.

Problem is the evaluation part. Technically it is easy, just shoot an IT-8 chart, but doing subjective evaluations is not easy.

Best regards
Erik

That would be an interesting test, Erik.  I would be interested to hear what Anders Torger thinks about all this...

Jack

[AlterEgo]

I want to make a profile for my Sony A7rII

I modded the new from 9.0.2 version IQ3-100 flash - flat art repro profile (replaced the original input curves inside to boost shadows) from C1 distribution and it actually looks nice for A7RII raws