Sony A900 noise, dynamic range and noise reduction

[Guillermo Luijk]

Regarding the noise reduction at low ISO I do have 100-1600 ISO shots like the ones Guillermo analyzed. So I plan to go ahead and check FFTs on all those images, but I don't have image analysis software on my primary workstation (Apple Imac using Tiger), so I need to revert to another platform, or find good software for the Mac. The need is essentially to be able to read unprocessed images in TIFF or similar format and perform FFT on them.

Erik I use a PS plugin for FFT by Alexandre V. Chirokov. You can download it from here. Just save it in the usual path for plugins and apply it to the noise image having been converted to 8 bit. Channel G of the resulting image is the FFT amplitude.

For extracting the RAW data from each channel, you can use DCRAW in Mac. I would recommend:

dcraw -v -d -r 1 1 1 1 -T -4 -t 0 file.cr2

The resulting TIFF will be a grayscale RGGB pattern with RAW data. It was simply rescaled and fitted in the 16-bit range.

To isolate each of the R, G1, G2 and B channels you don't need to leave PS, just rescale to 50% using the 'Aproximation' (no idea if the name changes for the English versions of PS) method. This performs a nearest neighbour rescaling which will pick only one out of each 4 channels:





For a RGGB pattern:
1. Straigt rescaling provides the B channel (like in the previous image).
2. If you add 1 pixel of canvas on both sides prior to rescaling it provides the G2 channel
3. If you add 1 pixel of canvas up and down prior to rescaling it provides the G1 channel
4. If you add 1 pixel of canvas on both sides so as up and down prior to rescaling it provides the R channel

Once you have the desired channel, crop an area with uniform noise (no texture), converto to 8 bit and apply the FFT as explained above.

I would like to see the FFT at ISO100 with NR OFF and HIGH. Remember after FFT, the G channel of the image contains the FFT modulus, which can be enhanced using a contrast curve.

BR

[Iliah]

For extracting the RAW data from each channel, you can use DCRAW in Mac. I would recommend:

dcraw -v -d -r 1 1 1 1 -T -4 -t 0 file.cr2

The resulting TIFF will be a grayscale RGGB pattern with RAW data. It was simply rescaled and fitted in the 16-bit range.

To isolate each of the R, G1, G2 and B channels you don't need to leave PS, just rescale to 50% using the 'Aproximation' (no idea if the name changes for the English versions of PS) method. This performs a nearest neighbour rescaling which will pick only one out of each 4 channels:





For a RGGB pattern:
1. Straigt rescaling provides the B channel (like in the previous image).
2. If you add 1 pixel of canvas on both sides prior to rescaling it provides the G2 channel
3. If you add 1 pixel of canvas up and down prior to rescaling it provides the G1 channel
4. If you add 1 pixel of canvas on both sides so as up and down prior to rescaling it provides the R channel

Once you have the desired channel...

4channels is easier, takes only file name or a wildcard as a parameter and you will get 4 totally unmodified raw planes as TIFF files.

[ErikKaffehr]

Guillermo and Iliah,

Thanks for all good suggestions. I finally used 4Channels and ImageJ under Leopard. 4Channels would SIGV under Tiger.

I tested three different images from the same series that were analyzed by Guillermo, ISO 100 with NR OFF and HIGH and ISO 1600 with NR HIGH.

The result was that I could not see anything significant in the FFTs of the 100 ISO-images while there were patterns in the 1600 ISO image. According to the manual high ISO NR is off from 100 ISO to 800 ISO.

Left image ISO 100, NR HIGH. Right image ISO 1600, NR HIGH.

Best regards
Erik

[Panopeeper]

Sorry for the delay, but this has proven to be much more laborous than I had thought. I have been experiencing with different ways of demonstrating that, what is obvious for me from analysis conducted by Rawnalyze.

Sony A900 - the effect of noise reduction on the raw data

[Plekto]

Is there any way to be 100% positive that it's not one of the converters or programs that's doing this?  If so, this is a huge problem with the Sony.       RAW should be exactly that and not what appears to me to be JPEG-like NR and compression in a RAW format.

[Iliah]

Is there any way to be 100% positive that it's not one of the converters or programs that's doing this?  If so, this is a huge problem with the Sony.       RAW should be exactly that and not what appears to me to be JPEG-like NR and compression in a RAW format.

It is not noise reduction, it is a crude analog to digital conversion that drops levels and clips to black forming clusters and causing posterization in shadows. That, indeed, is some problem; and Sony is well aware of it. By the way, the problem exists not only in Sony cameras.

[Ray]

It is not noise reduction, it is a crude analog to digital conversion that drops levels and clips to black forming clusters and causing posterization in shadows. That, indeed, is some problem; and Sony is well aware of it. By the way, the problem exists not only in Sony cameras.

Is this not a case of 'you get what you pay for'? When you say that the problem is due to a crude A/D conversion, that implies that the solution to the problem is a more sophisticated and therefore more expensive A/D converter, or a completely redesigned A/D converter costing lots of research dollars which would have to be recouped with a higher price for the camera.

[ErikKaffehr]

Hi,

The problem with this discussion is that it develops to a lot of Sony bashing. The problems that Panopeeper see show up in well designed tests, but not really in normal pictures. That said there are probably other cameras offering better noise performance like the Nikon D3X or the Canon 5DII. Some of he problems get exaggerated if the ACR engine is used for RAW-conversion. Panopeeper and Iliah don't agree on the cause of the problem. In my view Gabor did an excellent work in analyzing the images, but the interpretation of his results is subject to discussion. It seems that the Canon images are void of this problem. I obviously cannot compare the three top tier FF cameras, owning only one.

The two images below are from a test shoot of my new Sigma 12-24/4.5-5.6 lens, surprisingly good, BTW. One is the full image, the other one a detail from the shadows with "Light and Shadows" applied over part of the image. In my view this image is quite demanding regarding DR. The highlight include sun visible trough clouding and it is of course clipped, recovery turns it into a grey area.

Sorry for having problems with putting attachements in the right place...

Erik

[attachment=12203:20090315...748_Edit.jpg]

It is not noise reduction, it is a crude analog to digital conversion that drops levels and clips to black forming clusters and causing posterization in shadows. That, indeed, is some problem; and Sony is well aware of it. By the way, the problem exists not only in Sony cameras.

[douglasf13]

Is this not a case of 'you get what you pay for'? When you say that the problem is due to a crude A/D conversion, that implies that the solution to the problem is a more sophisticated and therefore more expensive A/D converter, or a completely redesigned A/D converter costing lots of research dollars which would have to be recouped with a higher price for the camera.

It's a little more complicated than that, because there are many ADCs, and they're all on the chip.  The EXMOR is a completely different design than what we've seen from other cmos makers, and there are good and bad things about it. It seems that, however the D3x is attaining it's 14bit processing, it partly shows why the camera is so expensive. We've known all along that the D3x is better in the shadows.

[Ray]

It's a little more complicated than that, because there are many ADCs, and they're all on the chip.  The EXMOR is a completely different design than what we've seen from other cmos makers, and there are good and bad things about it. It seems that, however the D3x is attaining it's 14bit processing, it partly shows why the camera is so expensive. We've known all along that the D3x is better in the shadows.

Can you expand upon that? Are you saying that each of the 24m pixels has its own A/D converter? If that's the case, presumably they are all the same on the same design of chip. The D3X presumably has a different design and quality of A/D converter because it's 14 bit instead of 12, which presumably contributes to the extra cost of the D3X.

Are you implying that the problem with the A900 is basically due to its 12 bit processing?

[ErikKaffehr]

Ray,

The one you should ask is Iliah, because he discussed the raw-conversion problem. The Sony has not 24 million ADCs but well over 6000, one for each column. Both the A900 and the D3X seems to use on chip AD-converters, with the Nikon being able to use 12-bit or 14-bit ADC conversion, 12-bit conversion is 5 FPS whereas 14-bit conversion is 1.8 FPS. Now, the FPS difference doesn't count if you are shooting with self timer and mirror lock up. I don't know how Nikon achieves 14-bit conversion, could be that they have 14-bit converters on chip or that they use existing 12-bit converters in a smart way. To my understanding the D3 does have external ADCs on the circuit board but these are absent from the D3X.

If you dig into into DxO-mark you can find the following figures:

[attachment=12207:Alpha900.jpg]
[attachment=12206:Canon5DII.jpg]
[attachment=12208:Nikon3DX.jpg]

You see that Nikon and Alpha are similar and Canon is different, why I don't know.

Best regards
Erik

Can you expand upon that? Are you saying that each of the 24m pixels has its own A/D converter? If that's the case, presumably they are all the same on the same design of chip. The D3X presumably has a different design and quality of A/D converter because it's 14 bit instead of 12, which presumably contributes to the extra cost of the D3X.

Are you implying that the problem with the A900 is basically due to its 12 bit processing?

[Ray]

The Sony has not 24 million ADCs but well over 6000, one for each column.

Ah! That makes more sense. But it's still a puzzle to me how any design of a RAW converter could undo what is a fait accompli due to sloppy hardware A/D conversion.

[Iliah]

But it's still a puzzle to me how any design of a RAW converter could undo what is a fait accompli due to sloppy hardware A/D conversion.

It is not sloppy (crude and sloppy have different meanings IMHO), it is state of the art, meaning the camera successfully competes over my D3X and 1Ds MkIII, rendering sometimes colour detail separation superior to D3X but for quite a lot less money. In raw converter void shadows and void highlights in one of the channels are treated equally - through additional interpolation.

[Guillermo Luijk]

The result was that I could not see anything significant in the FFTs of the 100 ISO-images while there were patterns in the 1600 ISO image. According to the manual high ISO NR is off from 100 ISO to 800 ISO.

Left image ISO 100, NR HIGH. Right image ISO 1600, NR HIGH.

I wouldn't be so sure. There seems to be 'something' happening at ISO100:



This could agree with the soft NR Gabor observed. I'll try to read his article now but I am working on a Jurassic connection speed these days.

BR

[Iliah]

There seems to be 'something' happening at ISO100:

Something is definitely happening at low ISO, especially with those weaker channels, red and blue (daylight shooting) - no question about that. If you derive the looks of FFT for the type of ADC behavior I described you can see similarities. Furthermore, if you examine raw files taken at different ISO settings including intermediate, like 320, you will see that the effect is in fact weaker at ISO settings higher than base. Comparative study with other cameras, 5D original on the first place, will show more. Attributing the effect to pixel-level noise reduction one needs to come up with an explanation how it should work to produce the distributions we observe.

[Panopeeper]

In my view this image is quite demanding regarding DR

Erik, what about uploading this raw file?

[Panopeeper]

It is not noise reduction

1. "It" does reduce the noise,

2. "it" can be influenced by NR selection,

3. Sony calls "it" noise reduction.

Reasons enough to call it noise reduction.

it is a crude analog to digital conversion that drops levels

1. I find it highly improbable, that the analog circuitry is capable of doing this on *five levels, selectable by the user*.

2. The green channels would be equally affected.

and clips to black

It does not.

[Iliah]

> "it" can be influenced by NR selection

Please demonstrate it.

> The green channels would be equally affected.

Please demonstrate this too.

Gabor, what I read in your post is a lot of statements without any proof, model, all subjective interpretation, and you refuse to perform experiments that will show you why and where your methods and conclusions are wrong. OK, send your study to Sony and see what they will answer you.

[ErikKaffehr]

Yeah but where? I can send it to you or anyone asking for it.

Best Regards
Erik

Erik, what about uploading this raw file?

[Panopeeper]

Yeah but where?

Yousendit, as usual. If you post the link here, then everyone can download it; if you don't want that, then send the link to me and to those specifically asking for it.