R3 NR analysis

Started 6 months ago | Discussions thread
EthanCam Regular Member • Posts: 111
Re: R3 NR analysis

Thanks JACS for your ongoing analysis on the R3, and Hoka Key for the files.

For the R3 I find particularly interesting the apparent "triple" base ISO, as I am not aware of other cameras behaving like this:

bclaff made the point that the apparent third ISO range might actually be an effect of stronger noise reduction, but only NR analysis will tell.

It's already interesting that you see no sign of NR at higher ISOs.

Some time ago I started a NR analysis on my R6 but I have not yet taken the initiative to share it for some reason. But now I guess it is a good time.

So I go straight to the results, such that you have something to compare at. This are however for my R6.

3x3 Blur Kernels are applied on each of the Red, Blue and Green=G1+G2 channels separately. All three blur kernels are the same. The intensity of the applied Blur Kernels decreases with increasing Noise at the pixel level.

- It is a blur kernel (normalized to 1), and not a median filter, star-eater filter, ... as others have suggested. Indeed, such filters have a FFT signature that is different from the FFT we get from the raw data.

- The Red, Blue and Green kernels are the same. They are 3x3 kernels with 1 middle value and 8 identical outer shell ("first onion ring") values.

The kernels can be parametrized by a NRscaling argument, where we have:
outhershell_values = (16-sqrt(256-288*(1-NRscaling^2)) ) / 144
middle_value = 1-8*outhershell_values

In turn, NRscaling can be roughly parametrized from the noise at the pixel level PixelNoiseDN:
NRscaling = min(1,1./(1+exp(-(+0.15+(log2(PixelNoiseDN))).^1.15)))

- How to get the noise at the pixel level?

In principle, the noise at the pixel level could be estimated from the signal S at the pixel level, the read noise RN from the optical black pixels, and the gain g of the sensor for a particular ISO value:
PixelNoiseDN = sqrt(g*S+RN^2)

- Note: the intensity of the blur kernels has a well defined "y vs. x" relationship with noise values only. The relationships with Signal value, ISO value or SNR value alone are not well defined...

- As you mentioned, the 3x3 kernels result in 5x5 auto-correlation kernels.

- A direct consequence of this: at ISO 100, the "true" PDR of the R6 is about 2/3 stops less than the 11.16 stops reported on https://www.photonstophotos.net/Charts/PDR.htm#Canon%20EOS%20R6

Another consequence: details are a bit smeared out in the darkest part of the image at low ISO values, but difficult to notice unless one lifts the shadows by several stops.

- Why Canon is doing this? No idea.

Below I show my data from which I estimated the parametrized formula and a "quick and dirty" visualization of the pixels involved in the blur kernels (tilting the head 45° might help visualize how the green blur kernel is in fact a 3x3 square kernel involving the most direct green neighbors, just like the blue and red kernels).

Intensity of R6 blur kernel vs. Noise at the pixel level. Estimated with dashed black line. Methodology: Mechanical shutter. Pair of exposures over the whole signal range (addition to estimate signal, subtraction to estimate noise, see for example www.strollswithmydog.com/digital-sensor-iq-model/). Noise analysis of central 512x512 sections: find kernels matching the observed FFT and auto-correlation kernel of the raw data.

Source Image: www.strollswithmydog.com/hv-spectrogram/. Yellow dots form a 3x3 set of pixels forming a red blur kernel . Black dots form a 3x3 set of pixels forming a green blur kernel (affecting the green value with a black circle). White dots form a 3x3 set of pixels forming a green blur kernel (that affects the green value with a white circle).

Post (hide subjects) Posted by
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow