rico wrote:
Statistically speaking, per-pixel noise is reduced by one bit (one f-stop) when four pixels are averaged (the arithmetic mean is taken). If neighbors are averaged, the image size in pixels is then 25% of the original. Alternatively, you can average across four frames, and retain the image size: this is called "stacking" in astro. Stacking requires image registration and a static subject. Either averaging approach can be applied with no upper limit from the arithmetic, but the sample size grows exponentially: 4, 16, 64, 256, etc.
Dark-frame subtraction (erroneously called noise reduction in photog-land) is a separate concept, and involves a simple arithmetic subtraction. The target is systematic error, or bias (e.g. purple amp glow, hot pixels). Bias is not noise! The dark frame must be created with the same sampling regimen as the main image, otherwise the noise is reintroduced (although the bias will be gone).
dcmiller wrote:
Yeah, but what's noise? Is it the error in measurement? Or the crunchy stuff with off color pixels we don't like?
The latter, mostly (IMO).
Canon's enginerds got in a bit of trouble by being sincere about accurately measuring wavelength and getting the analog signal to the A/D converter unchanged.
Well, I'm mainly with Canon on this.
Nikon knew you could just blast away at unpleasant looking noise post capture. So the photos smooth out and lose detail. So what. Look at those dark areas! Smooooth.
Yes, but they're UNNATURALLY smooth, IMO. Also, when one PRINTS, the noise seems to be lessened in my experience.
So which type of noise are we talking about? Do you want accurate or pleasant looking?
Yes.
I want "pleasant looking", but my definition of pleasant-looking is closer to Canon than to Nikon.
Me, I've just attached a flux capacitor to my D60 and use the fourth dimension to average out the quantum differences. I actually finished writing this reply before I started.
David, thank you for the legwork and pulling together the explanations of what's going on with sRAW. I'm with you in expecting that full raw that's downsized in PP would give better quality than sRAW. I see no substitute for having all the original capture information, with any compression being lossless. What does make sense is that an sRAW image of x MP should have better quality than an image originally shot at x MP, since the sRAW has extra information saved from the 4x MP capture (hence the larger file size).
rico wrote:
Statistically speaking, per-pixel noise is reduced by one bit (one f-stop) when four pixels are averaged (the arithmetic mean is taken). If neighbors are averaged, the image size in pixels is then 25% of the original. Alternatively, you can average across four frames, and retain the image size: this is called "stacking" in astro. Stacking requires image registration and a static subject. Either averaging approach can be applied with no upper limit from the arithmetic, but the sample size grows exponentially: 4, 16, 64, 256, etc.
Dark-frame subtraction (erroneously called noise reduction in photog-land) is a separate concept, and involves a simple arithmetic subtraction. The target is systematic error, or bias (e.g. purple amp glow, hot pixels). Bias is not noise! The dark frame must be created with the same sampling regimen as the main image, otherwise the noise is reintroduced (although the bias will be gone).
A couple of things. Dark-frame subtraction is not a bias-only subtraction. Pure bias frame subtraction uses a bias frame of extremely short duration. Dark-frame is removing systematic error, but in this case its the error accumulated over longer exposures at a certain thermal profile for the sensor. In this sense, the use of the term "long-exposure noise reduction" is indeed correct. Additionally, this is noise. It is not the signal you intend to collect, its systematic noise.
On the topic of binning, true pixel-level binning (bin the signal and photon and shot noise of the pixels, then read them all once) should help reduce noise by reducing the read noise and shot noise, but in practice the gains aren't as great as one might hope.
I wonder what changes allowed DSLR makers to eliminate dark frame subtraction from the camera startup. I believe with Canon the D60 was the last camera to perform this function always when the camera was turned on.
I understand WHY they stop doing it at startup. But what replaced this technique?
dcmiller wrote:
I wonder what changes allowed DSLR makers to eliminate dark frame subtraction from the camera startup. I believe with Canon the D60 was the last camera to perform this function always when the camera was turned on.
I understand WHY they stop doing it at startup. But what replaced this technique?
I am not familiar with this and didn't even know it had occurred. However, it sounds like this was a bias-frame subtraction. My guess is that they got the inherent system noise under more control, so they could remove this step.
I am kinda perplexed! How can you do effective binning with an GRGB Bayer sensor? Seems to me that any binning, in particular when the down-size image is 1/2 of the full frame (such as 20mp binned down to 10mp) would be subject to accidental colour content and noise would not be "properly reduced" because of the colour filtration . What if there is no red/green/blue when the pixels are merged, what happens when the two noise components are added together, etc. etc.
I can see how binning would be good for monochrome (non Bayer) sensors.
I don't think that they do real binning--hence what I said about "true pixel-level" binning. In fact, I think that it would be much harder to do so with a CMOS sensor, due to the nature of the architecture. Not that it couldn't be done, but between that and the Bayer filter, I would guess its easier to cheat somehow. How, I'm not positive.
