Originally posted by reh321 What is the real difference between “post facto denoising” and what Pentax does?
There is no difference. With "post facto denoising" I referred to the approach (practised by Pentax) of reducing noise through post-capture image processing. This kind of image processing can be done at any point in time, in-camera or in post-processing on a computer.
Originally posted by reh321 Denoising is noise reduction
I use the following definitions (perhaps debatable but definitely useful):
"noise reduction" = reduction of sensor-induced noise using in-sensor technology (e.g., dark current suppression). These techniques address shortcomings of the light capturing process and never have detrimental effects on image data integrity (due to the fact that interfering sources are attenuated, subtracted, etc.; they cannot be confused with the signal/data).
"denoising" = image processing technique for removing noise from images (e.g., by smoothing high spatial frequencies). This approach does not have access to additional information at the time of capture and will result in some loss of data integrity (some information will be lost, due to the fact that signal/data and any interferences one wishes to remove are now inextricably combined). Subjective evaluation may attest "better image quality" since the image does not look as noisy anymore, but inevitable some information will have to be sacrificed (hence Pentax's comments on the need for caution when using this approach).
It would be reasonable to use "noise reduction" as a general term and regard "denoising" as a special case of "noise reduction" but then I don't know what term I'd use for the in-sensor "noise reduction".
Originally posted by reh321 ... removing what was added to the data.
Modern sensor technology is so good that there is hardly any noise added to the data (data = photon count received by the sensor). The order of magnitude is a couple of individual electrons per pixel.
The overwhelming noise one sees in a high-ISO image is part of the data. Illumination is a stochastic process -- individual quanta (photons) randomly hit a scene -- and in very low light levels, its random nature is revealed. The random nature of illumination means that recorded intensities will fluctuate. Some pixels receive more photons than others but not because there is an actual brightness difference, but because they randomly received more photons. When the number of photons is large (i.e., light levels are higher) then two pixels recording the same brightness will statistically receive a very similar amount of photons. When the number of photons is small (i.e., light levels are low) then one pixel may receive zero photons and the other pixel two photons, even though they both should have received one photon each.
So, in summary, the salt & pepper look of high-ISO images is correctly capturing reality; that's how the scene looks like when only a very limited number of photons are around. In other words, there is "scene noise" which is part of the data.
What "denoising" does is attempting to make the image of the scene look better than the scene looked. Granted, "denoising" will also address the very low residual sensor-contributed noise, but due to the proportion between "sensor-induced noise" and "scene noise", "denoising" mainly combats "scene noise".
As you can imagine, leaving the data intact, i.e., refrain from "denoising" can have advantages, e.g., by stacking multiple noisy image one can get a clean image since one is essentially simulating a longer exposure time, which in turn means more photons are being used, which in turn increases the signal to noise ratio. If you stack denoised images, you no longer get this simulation effect, as you are now working on data that no longer represents the photons during capture. Another advantage of preserving intact data is that future image processing may do an even better job than the "accelerator unit (II)" of reconstructing a nice looking scene from noisy (here, "noisy" = "incomplete") data.
I hope this helps to better appreciate my perspective.