R3 NR analysis

Started 4 months ago | Discussions thread
ForumParentFirstPrevious
Flat view
J A C S
J A C S Forum Pro • Posts: 20,292
R3 NR analysis
8

Hoka Hey was kind enough to provide a bunch of RAW files taken with his R3 for NR analysis. This would be a series of posts in which I will share my findings. If you want to short version, here it is: it looks more or less the same as the R5, there is a slight NR for low ISO at low light levels, and I see no NR at high ISO (contrary to some reports). Those are preliminary impressions; they may change while I keep posting.

For each ISO, Hoka sent me three frames: pitch black, dark, and (dark) gray. Here are some crops combined together, developed in DPP, default settings. All shot with Electronic Shutter.

I start with ISO 100. I did many tests - rows, columns, standard deviation by row and column, FT by row and column, and also overall, plus auto-correlation. Here is a plot of the St. Dev. of each row. Apparently, there is no NR on the first and on the last one (also, no NR on the first and the last column), confirmed by their FT. You see that the St.Dev. is reduced by about 1/2 stop compared to the most left and the most right rows. So 1/2 stop is my estimate of the NR at the bottom of the tonal level. This is G2, the other channels behave similarly.

St. Dev. by row on the G2 channel, ISO 100, black frame

Next, I computed the auto-correlation. For those not quite familiar with this term - it measures how related a pixel is compared to its neighbors. The bright yellow one has value 1 because each pixel is correlated (=1) with itself, and no NR would mean all others at zero. I combined G1 and G2 into one to get this:

auto-correlation of G1+G2, ISO 100, black frame

Next, the auto-correlation of R and B, they are similar:

auto-correlation of R and B, ISO 100, black frame

Here, I isolated the R, respectively the B pixels only! To get the "big" Bayer picture, you have to put gaps between the pixels mentally. So the 3x3 box is really a 5x5 one.

Now, a speculation. I do not know how this processing has been done. If it is done by local averaging, then here is my estimate of the blur kernel. It looks a bit darker that what you have above. If you look closely, you'd see above that the active pixels form a somewhat larger 5x5 box (which in the Bayer sensor is 9x9) but the border ones are very faint. The blur kernel (called "PSF" below) looks like it spreads to the neighboring pixles mostly:

blur kernel, B and R, ISO 100, black frame

The autocorrelation is just the blur kernel convolved with itself, which explains the wider spread and the brighter neighboring pixels. The blur kernel is a better representation of the NR of whatever is going on there.

Back to G1+G2: I did not try to extract the blur kernel, too much work. Enough to say that the bright blue pixels would look somewhat darker.

To be continued...

Canon EOS R5 Panasonic Lumix DMC-G2
If you believe there are incorrect tags, please send us this post using our feedback form.
ForumParentFirstPrevious
Flat view
Post (hide subjects) Posted by
ForumParentFirstPrevious
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark MMy threads
Color scheme? Blue / Yellow