A700 terrible red channel noise problem

[agorabasta]

Solution is simple - don't do channel shadow clipping (black level subtraction) before demosaicing, then demosaic at a WB settng that fills channels histograms the best possible way, then only apply WB and black per-channel subtraction on your 16bit/channel demosaiced image.

I didn't do any of those things when I did the raw conversion. See my OP. I set everything to zero in ACR. I don't know what ACR is doing under the covers though and I don't have much control over it.

The problem is that ACR does all those bad things for you... And you have near-zero control.

[agorabasta]

Here's a link to the junk album containing the full size developed image (and some other irrelevant junk) - http://picasaweb.google.ru/AgoraBasta/T ... directlink - quite open for dissection. (As it's a JPEG, use an appropriate space, like YIQ or any YCrCb)

[agorabasta]

And I thought everyone was able to press the 'download' link in Picasa album...
The humankind never stops to disappoint me

[agorabasta]

Hmmmm.......

[bakubo]

There is an easier way!

I am certainly not averse to easier!

Now we are talking a quick and sometimes usable technique for reducing blue sky noise. I own Neat Image Pro+ and CS3 and all you need to do is select the channel you want to NR and then click reduce noise (further setting proper parameters of course, or using pre-saved presets and/or profiles). This can be a somewhat effective technique for reducing blue sky noise and, if needed, you can actually cut new profiles based on just the offending channel(s) if you like. Be careful though as it is still very easy to soften or quack up an image even when working on just one channel so you often times you will need to apply some NI sharpening....which is what I am almost always doing anyway when using NI. To be honest, while reducing the red channel alone will usually improve the noise, often times IQ affecting noise is still visible in the sky (depending on the output size and/or medium) that was derived from the other channels, especially the green.

I think I follow you. In NI on the Device Noise Profile tab over on the top right corner there is a pulldown menu called Working color space. The default for that is YCrCb, but I can choose RGB. Then on the Noise Filter Settings tab in the Noise Reduction Amounts I can set the amount of noise reduction for the Red, Green, and Blue channels separately. Is this what you are talking about?

I have been out a lot today, but I will try to give this a try tonight.

[bakubo]

Thanks, Don! The part I didn't know was the PS Window|Channels thing. I assume that after running NI on the red channel that when I then turn off PS Window|Channels it recombines the R, G, and B channels using the new noise reduced red channel? For the last hour I have been playing around and I will continue. If I get something looking pretty good I will post it.

In PSP X there is a Split channels command that creates an R, G, and B window. You can then do what you want to each of them and then use the Combine channels command to put them back together. See these screen shots to see how it works in PSP X.

[bakubo]

I took one of the ISO 160 Grand Teton images I made in May and worked on it. I did as Don suggested on the red channel and here are the results. The first one is a 100% crop before using NI on the red channel and the second one is after using NI with 100% for Y and Very low freq turned on. I also tried it on the full image and using the 100% noise reduction setting on the red channel had very little effect on the detail of trees and other detail in other parts of the image. By the way, both of these crops have no sharpening done to them.

[stevecim]

Looking at lot of my photos, same red channel noise is happening with the a350, so it's not just limited to the CMOS based sensors also affects the the CCD sensor. Since it's happening on both type of sensor is it safe to say it's due to how the images are processed and not some limitation of the sensors?

[agorabasta]

The image like this is not really so good an example of noisy sky, simply because it's pretty uniformly noisy throughout:



Yet it's possible to use it and demonstrate how a cyan boost can sway the 'Sony skies' towards 'Canon skies'. Below are two samples of a 300% magnified non-filtered crop of that image before and after modifying it with black'=black-0.1cyan




The trick is modifying the reds so as to effectively lower the colour separation in the red channel, it makes the blues slightly brighter while lowering their saturation. Subjectively, though, the saturation of bright parts seems higher and the saturation of darker parts seems lower. Noise is quite clearly lower in result.

If the trick is applied to all channels, like black'=black-0.1cyan-0.1magenta-0.1yellow, then adjusting relative gamma to about 0.9, you get essentially Canon-like look from Sony pix.

Then if reversed proc is applied to a Canon pic, i.e. black'=black+0.1cyan+0.1magenta+0.1yellow and gamma 1.1, you get a close mockup of 'Minolta colours' from a Canon.

(link to Picasa album with all those samples - http://picasaweb.google.com/AgoraBasta/ ... directlink - the page should appear in English)

[agorabasta]

Chances are, Sony is not doing any in-cam raw processing at all. It's the Apical doing it for them...
One thing for sure is that Apical is fed raw data whenever they do the DRO. I'd guess it would be stupid to have two separate routes for raw processing in one cam, and it would be verrrry difficult to have them meet each other seamlessly - hence the above conjecture.

[agorabasta]

I have no big problems with a700 v04. It's got the most pliable, the most true 'raw' of all comparable cameras out there in existence - just ask the astrophotographers. The only cam that's got a more 'linearizable' sensor raw data is the a900. The seemingly better results from the others are really a result of cheating and corner-cutting.
I know how to get what I need out of its raw, and how to set up the internal conversion to get the very top quality JPEG out of it.
So far, the only small complaint is that there's no in-cam CA correction, so I really can't use the internal conversion with the higher CA lenses.
That's about it.

[agorabasta]

Of all current cropped cams there's only the D300 that has the same high ISO performance. You need an FF cam to beat A700.
Sony colours may appear odd compared to others, but you can reduce Sony colours to the exact looks of others as it's exactly the Sony that is registering them with most fidelity, which is not always pleasing all.
Things like AF fine-tuning, AWB fine-tuning and automatic CA/geometry correction would be certainly highly welcome, now I have to hand-pick lenses and normally avoid AWB. But Sony's AWB is really just as poor as any other AWB.
Their raw software is worse than nonexistent - it's destructive to image quality.
Sony needs to replace the mechanic dials with Hall-effect switch dials, those mechanic dials don't really last over one year.
Channel clipping, as any other unnecessary data reduction, should always be avoided by all means.
Default in-cam image processing is truly poor, but there's a great freedom for adjustments and then the results are truly great.

[agorabasta]

D5000 has its own shortcomings. Though it's got seemingly the same sensor, the mods to that have been done.
It appears they simply lowered the photosite well depth without changing anything else. That killed low-ISO DR, while adding some at higher ISOs. This way they linearized the otherwise very exponential response of the sensor that was the reason for bit-slicing/noise artifacts with earlier firmware versions of a700. The D5000 sensor modding trade-off shows itself in that the low-ISO pix from D5000 bear a clear P&S-ish look to them.

Then there was the D90 which was also crippled, that time it was mostly optical crippling with LPF blur that helped them hide the NR blurring...

So no, the only APSC equivalent to a700 with current firmware is the D300. And you still get more pliable colour data with the a700.

[agorabasta]

That DxO puts D300 at the same levels as a700 and well below the D5000.

To me that means that they really do measure something quite consistently. It's when they turn to interpreting their measured data that they turn delusional.

Having their 'color sensitivity' evaluation so greatly noise-dependent is a stark example of deceptive nature of their interpreted results. And at the same time they don't even try to evaluate the coverage of sRGB/aRGB spaces, leave alone measuring the colour fidelity within those and/or having their data normalised to faithful colour response.

They are ridiculous.

[agorabasta]

May them D5000 fanboys enjoy the green skies till kingdom come!