Sony A900 noise, dynamic range and noise reduction

[Panopeeper]

In any case I love the image quality I am getting from my A900, and the noise cleans up very well in Neat Image

Doug,

let's make something clear: I have nothing for or against the A900; I am neutral even towards my own camera (in fact the only reason I have that is, that I have a much larger investment in lenses). I have no reason to make the A900 appear better or worse than it is; my interest is purely professional. I am analyzing raw images from dozens of camera models several hours long a day, and the A900 is something special.

Now back to the question if this is done in the A/D circuitry or later. Again this question is irrelevant regarding the validity of my assertions, but it is 9an interesting question.

It is simple to say that this is done by the A/D converter, but you have to think about details: what exactly is done.

1. Action on five levels, selectable from outside.

2. Performing a different action (this is the sixth version, much different from the others) on every second pixel: the A/D is column oriented, and every second pixel is green in every column. The green pixels are treated very differently from the others.

3. Watching the neighboring column's A/D converter: the blobs are not column or row oriented but twodimensional, context sensitive.

This is such an honorous task that Sony could have implemented a much better electronics, like some MFDBs have, still much cheaper. On the other hand, this is nothing special for the on-board processor. This may be hardware (i.e. not replacable by firmware version), but that is a far cry from being done by the A/D converter.

Those knowing how exactly such circuitries are implemented may have a different opinion, and I would not dispute that after they have considered the above points.

[Panopeeper]

We read your paper carefully, several times, and spent 2 hours discussing it

Well... One of the statements in the abstract is

The green channel is affected to a much lesser degree than the other channels, with all ISOs and all NR settings.

Thus there is no contradiction between the paper and the statement

The green channels would be equally affected (namely if qbic's speculation were fitting).

[douglasboyd]

Doug,

Now back to the question if this is done in the A/D circuitry or later. Again this question is irrelevant regarding the validity of my assertions, but it is 9an interesting question.

It is simple to say that this is done by the A/D converter, but you have to think about details: what exactly is done.

1. Action on five levels, selectable from outside.

2. Performing a different action (this is the sixth version, much different from the others) on every second pixel: the A/D is column oriented, and every second pixel is green in every column. The green pixels are treated very differently from the others.

3. Watching the neighboring column's A/D converter: the blobs are not column or row oriented but twodimensional, context sensitive.

Dear Gabor,

Yes, I agree that the kind of processing I suspect would be done by software.  It is after all an algorithm.  But I can see two dimensional blobs forming even if the processing algorithm is applied to columns only.

By the way, the black dots seen in Canon5DII would be easily removed by this kind of processing....

This may also be the reason that the blotchiness is prominently seen in GREEN grass (blue and red values low).

==Doug

[Plekto]

I really don't know what you are referring to. I have not observed anything like that on D200 raw data, but my collection is quite limited.

I think it was the D20.  One of the other cameras also was visibly clipping its histogram.


EDIT:

The green channel is affected to a much lesser degree than the other channels, with all ISOs and all NR settings.

If the camera is trying to overcompensate for low levels(ie - "I see black - oh no! not enough photons!") and failing miserably, then it would stand to reason that because there are as many green pixels as the blue and red combined, it would mess up less often on the green channel.  The explanation Doug gave makes a lot of sense.  What happens if the sensor has almost no data on one of the channels - to the point where the random noise itself is considerably brighter?

I could see it getting confused.  It looks to me like it might be trying to NR the black areas and treat the noise as real data.

[Panopeeper]

This may also be the reason that the blotchiness is prominently seen in GREEN grass (blue and red values low).

The blotches appear the best (though this should be calle the "worst") in areas, where

a. the green is relative low (this does not cause any trouble but weekens the mitigating effect of the healthy green),


b. either red or blue is not low, i.e. not causing trouble,

c. the other one is very low; those blobs will be translated by the rendering directly in blotches, if not sppressed by Blacks.

Now the effect is at the maximum. Example the blue and red patches on the color checker. See
Blotches and Blobs

Another aspect against A/D caussing this, though it is not an absolute proof: some pixel values appear only in the left half of the image. See the attached captures: the orangy vertical line is just in the middle of the sensor. The images are in pairs: the first of a pair with red, blue respectively green pixels only in the left half, the next pixel level is overall. (The cross like form is from the black separator between the color patches.)

Note, that this super A/D converter would have to take care of this separately for the red, green and blue (the "levels of skipping" are not identical).

[Panopeeper]

If the camera is trying to overcompensate for low levels(ie - "I see black - oh no! not enough photons!") and failing miserably, then it would stand to reason that because there are as many green pixels as the blue and red combined, it would mess up less often on the green channel

LOL, you are attributing to the on-sensor A/D superhuman (or superelectronics) capabilities, combined with super mistakes.

The facts are not supporting this:

1. the green does not exhibit the blobs even in a black frame, where there is no difference between the intensity of the channels,

2. the red and blue blobs appear even at -5 EV, which is about 5-6 stops brighter than the black frame.

I could see it getting confused

Yes, I see that :-)

It looks to me like it might be trying to NR the black areas and treat the noise as real data.

This "black nonsense" (like "black clipping") is somehow firmly anchored in some minds. See #2 above: there is no black anywhere.

[Panopeeper]

Still to the "black myth":

1. This raw file (36 MB) is well exposed; take a look at it. The blue channel on the red patch is blobby; that at ISO 100. Load the image in PS and take a look at the red channel (blobby on the green and blue) and the blue channel (blobby for example on the yellow).

2. Look at following histograms, with ISO 1600. The first one is NR Off, the second is NR High. The red channel is not near black, still strongly affected, while the green is only slightly. (The green is only from one channel in both histograms, so that effect is shown in more clarity.)

[Iliah]

Well... One of the statements in the abstract is

The green channel is affected to a much lesser degree than the other channels, with all ISOs and all NR settings.

Thus there is no contradiction between the paper and the statement

The green channels would be equally affected (namely if qbic's speculation were fitting).

Interesting that you do not see the contradiction between two of your statements; because I was not speculating.
Again, here you are telling me (not qbic) that your article demonstrates the green channel to be equally affected:
Your statement:
The green channels would be equally affected.
Me:
Please demonstrate this too
You:
Sony A900 - the effect of noise reduction on the raw data is doing just that.

More interesting that you ignore the fact that in daylight green channel registers much more photons. Repeat under the incandescent light and see the results.

But most telling is your refusal to send your study to Sony. One of the possibilities is that at this stage you must know that it is not worth it - the study is one big error. In this case it is hard to understand why you still you spread it. The other, even less pleasnt possibility is that you are not attempting to help Sony users here.

[ErikKaffehr]

Hi,

Just for your information, the pictures I have taken were made under  incandescent light. Two simple halogen lamps on either side.

Best regards
Erik

Interesting that you do not see the contradiction between two of your statements; because I was not speculating.
Again, here you are telling me (not qbic) that your article demonstrates the green channel to be equally affected:
Your statement:
The green channels would be equally affected.
Me:
Please demonstrate this too
You:
Sony A900 - the effect of noise reduction on the raw data is doing just that.

More interesting that you ignore the fact that in daylight green channel registers much more photons. Repeat under the incandescent light and see the results.

[Iliah]

Just for your information, the pictures I have taken were made under  incandescent light. Two simple halogen lamps on either side.

I'm not talking about your images, but about tests shots making grey step wedge. Minimum is newspaper shots as I descrbed.

[Panopeeper]

Interesting that you do not see the contradiction between two of your statements; because I was not speculating.
Again, here you are telling me (not qbic) that your article demonstrates the green channel to be equally affected:
Your statement:
The green channels would be equally affected

I really don't understand your problem.

If qbic's suggestion was right THEN the green channel would be equally affected.

However, I have shown, that the green channel's processing is vastly different from the other channels.

Thus qbic's suggestion can not be right.

More interesting that you ignore the fact that in daylight green channel registers much more photons. Repeat under the incandescent light and see the results.

It is *totally* irrelevant, which pixels collect more light in which illumination generally; only the *actual light capture* counts. The attached histograms are from the red patch in an ISO 1600 shot, first NR Off, then NR High. The average of the red on that patch is 1.2 EV HIGHER than the green; still, the red undergoes a much-much greater change (between NR Off and NR High) than the green.

But most telling is your refusal to send your study to Sony. One of the possibilities is that at this stage you must know that it is not worth it - the study is one big error

I think it's time for you to get serious; joking has its limits.

Anyway, I wonder if you are really so naive as to believe, that Sony does not know this issue better. I am convinced, that this is not only known, but it is well-planned. In fact, it is a very good idea, as long as the blotches are suppressed. The customers are happy with their cameras, most reviewers have been fooled in publishing unrealistic figures, so what's the problem? Though I made a search and found an article (some printed photography related paper's site) from last year, saying that the raw data is massively NRed.

[Iliah]

However, I have shown, that the green channel's processing is vastly different from the other channels.

Not so. You ignore different effective amount of light. Results in the green channels are different (btw you show 3 histograms instead of four), but the explanation is not different processing, but different captured signal. And we are not talking about noise reduction at high ISO here.

Sony engineers and management need to have feedback on the perception of the product. That helps improving. Your position here is beyond my understanding.

[Plekto]

LOL, you are attributing to the on-sensor A/D superhuman (or superelectronics) capabilities, combined with super mistakes.

The facts are not supporting this:

1. the green does not exhibit the blobs even in a black frame, where there is no difference between the intensity of the channels,

2. the red and blue blobs appear even at -5 EV, which is about 5-6 stops brighter than the black frame.

This "black nonsense" (like "black clipping") is somehow firmly anchored in some minds. See #2 above: there is no black anywhere.

I didn't technically mean "black" so much as too low of a signal at the sensor.  Maybe  they've set the NR to a value where it's thinking "too dark - apply maximum NR" when it's still 5-6 stops before true black.  

In any case, the on-board software looks to be bugged and set up wrong.  That red/blue artifacting  shouldn't be so bad.  Maybe the green is fine but something about the sensor itself is mangling the red/blue channels and the NR is their attempt at "fixing" the problem - like adding a software micro-lens.  Since it's not applied to all three channels, of course you'd get artifacts.  Is there a way you can rule out Sony possibly applying a second level of NR to only one or two of the channels?

P.S. What this really reminds me of is over-amplification of a TV signal.  You get artifacts, ghosting, and noise because you're also amplifying the noise as well.

[Panopeeper]

I didn't technically mean "black" so much as too low of a signal at the sensor.  Maybe  they've set the NR to a value where it's thinking "too dark - apply maximum NR" when it's still 5-6 stops before true black

Fact is, that even NR Low affects already the fifth-sixth stop at ISO 1600, where the noise is only 15% (SNR > 6).

Maybe the green is fine but something about the sensor itself is mangling the red/blue channels and the NR is their attempt at "fixing" the problem - like adding a software micro-lens

Apparently this nonsense with "the sensor is doing it" won't go away.

Is there a way you can rule out Sony possibly applying a second level of NR to only one or two of the channels?

What would be the "first level"?

[Plekto]

Fact is, that even NR Low affects already the fifth-sixth stop at ISO 1600, where the noise is only 15% (SNR > 6).

If it's true, then something is badly wrong here.  

What would be the "first level"?

I meant that they are doing some fiddling and pre-processing with the data from the sensor before the NR is applied.  Kind of like a hidden extra level of adjustment that's not accessible to the user.  ie - "We screwed up with the sensor - quick - tweak it and kludge a fix in software before it ships!"  Is it the sensor/chips/hardware or the software that's causing this?  The data from the three channels is obviously not close - that green channel is just way off - like it's coming from a completely different camera.

[Panopeeper]

"We screwed up with the sensor - quick - tweak it and kludge a fix in software before it ships!"

I don't think the sensor is doing anything wrong (except for not being as good as Sony wants it to appear). I find the method quite smart, though it starts too "soon", i.e. at noise levels, which should be left for the later processing.

Basically, the method of keeping the green intact (well, not really, but compared to red and blue) preserves much detail, and the noise will be much less due to the red and blue adjustment. However, there are several downsides:

- loss of details, when not the green is dominant,

- color blotches (but obviously this is not a widely seen problem, I guess the "Blacks" takes care of it),

- off colors; this is very easy to see, I am surprized that we don't hear many complaints.

The attached captures show a crop from the same setting, ISO 1600, one NR Off, the other NR Normal, one pair sharpened, the other not. NR was set to 0 in ACR. Look at the red details: this is the consequence of the red mutilation (I added some extra contrast in PS to enhance the differences). Note, that this these reds are in the fifths and sixth stop of the DR, i.e. not dark.

[ErikKaffehr]

Thanks,

I'll check out a bit more! I like your processing. Highlight recovery seems to be better than in Lightroom. Would be interesting to see a full size JPEG or TIFF or some detail.

Erik

Yes, Aperture. I put your image through it and gave it some mild Photoshop touch-up afterwards. No noise reduction. I could squeeze a lot of information out of your ARW that held up pretty well when pushed a bit.

[attachment=12260:20090315_DSC01748.jpg]

[ErikKaffehr]

Gabor,

I don't really think that Sony manipulates the images just to get better reviews. For one they get bad reviews anyway. Most reviews are about JPEG and there is no need to manipulate "raw" to get good JPEGs. I see that there are some issues with Sony ARWs at high ISO with blochiness, these can be caused by the artifacts you see or something else. I think you demonstrated quite clearly that the artifacts are also present at low ISO.

I appreciate your efforts even if I don't necessarily share your opinion of Sony purposely manipulating the images just to get god DR.

Best regards
Erik

I really don't understand your problem.

If qbic's suggestion was right THEN the green channel would be equally affected.

However, I have shown, that the green channel's processing is vastly different from the other channels.

Thus qbic's suggestion can not be right.


It is *totally* irrelevant, which pixels collect more light in which illumination generally; only the *actual light capture* counts. The attached histograms are from the red patch in an ISO 1600 shot, first NR Off, then NR High. The average of the red on that patch is 1.2 EV HIGHER than the green; still, the red undergoes a much-much greater change (between NR Off and NR High) than the green.


I think it's time for you to get serious; joking has its limits.

Anyway, I wonder if you are really so naive as to believe, that Sony does not know this issue better. I am convinced, that this is not only known, but it is well-planned. In fact, it is a very good idea, as long as the blotches are suppressed. The customers are happy with their cameras, most reviewers have been fooled in publishing unrealistic figures, so what's the problem? Though I made a search and found an article (some printed photography related paper's site) from last year, saying that the raw data is massively NRed.

[Panopeeper]

I don't really think that Sony manipulates the images just to get better reviews

I don't believe that they are doing this for better reviews; that's a side effect, I would not even say that it is calculated.

As I posted, some reviewer found out the heavy NR on the raw data already last year. One does not need to go deep in the raw data; I presented samples from ACR prooving the NR effect, the details are only for the geeks. Thus the reviewer's oversight was not a calculable, except for DPReview, who don't understand anything of the issues past JPEG.

Btw, the apparent increasing of the DR through the darkening of the deepest shadows is not done in order to fool the measurements, it is inherent to the applied method. However, that should have been noted by everyone, who was *measuring* DR and noise instead of *prattling* about it.

[douglasf13]

As I posted, some reviewer found out the heavy NR on the raw data already last year. One does not need to go deep in the raw data; I presented samples from ACR prooving the NR effect, the details are only for the geeks. Thus the reviewer's oversight was not a calculable, except for DPReview, who don't understand anything of the issues past JPEG.

  Are you finding the same issues with non-Adobe RAW converters?  Could we see some examples with RAW Therapee, RPP, etc.