Marc Wossner wrote:

[]

> So only raw data should be considered. I found a procedure to do that

> but do not know if its correct:

>

> - Take a pair of images of a uniformly lit, slightly out of focus

> colorchecker chart.

> - Take the sum of the two images, and separate the result into its

> individual color channels. Measure the average raw value of each patch

> in one of the two green channel subarrays of the Bayer color filter

> array (or if desired, for each color channel). Divide by two (and, for

> Canon cameras, subtract the bias offset) to get the average signal.

> - Take the difference of the two images, split into color channels,

> and measure the standard deviation in each color patch. Divide by

> sqrt[2] to get the combined shot+read noise for that patch. (question:

> If the signal is 12-bit do I have to take the signal as 4096 and

> divide that figure by the number for the combined shot+read noise?)

> - Plot the noise^2 vs. the signal, fit to a straight line. The

> intercept is the square of the read noise, the slope is the inverse of

> the gain.

>

> Best regards!

> Marc Wossner
Marc,

I haven't done such detailed tests as this, so I'm not really in a

position to comment. Best to ask someone like Roger Clark.

http://www.clarkvision.com/imagedetail/index.html
I can see that if the camera has fixed-pattern noise, taking the

difference of two frames will reduce the effect considerably. However,

after taking the difference of the images, I would further adjust the

difference so that its mean value was zero, if necessary. Adding (or

subtracting) two images with Gaussian noise will cause the noise to

increase as a sum-of-squares, so dividing the difference by sqrt (2) to

get the per-image noise also sounds valid. If the signal is 12-bit, then

you could divide both noise and signal by 4096 to produce a normalised

result, but it shouldn't affect the slope of your line. It will affect

the units in which the offset is reported, of course.

Have you checked out:

http://www.clarkvision.com/imagedeta...1d2/index.html
He uses DN = digital number, i.e. signal levels expressed as numbers.

Cheers,

David