imnoise does not work as expected

imnoise(a, 'gaussian', mean, var) seems simple enough in concept: add a Gaussian distribution of noise to an image. However, it does not recover the values expected.

Try: >> a = zeros(16, 16); b = imnoise(a, 'gaussian', 0, 0.1); >> mean(b(:)) ans = 0.1262 >> std(b(:))^2 ans = 0.040

The mean is not zero! The variance is not 0.1!

I can't figure out why the imnoise code produces this result. The code snippet is below.

switch code case 'gaussian' % Gaussian white noise b = a + sqrt(p4)*randn(sizeA) + p3;

If a = 0 and p3 = 0, then the result should have a mean value of zero.

Answers?

Answer 1

[Steve Eddins]

imnoise clips the output to the range [0,1].

>> min(b(:))

ans =

       0

[Brian]

Ok. Now try something that should address this:

>> a = ones(1024); b = imnoise(a, 'gaussian', 0, 0.1); >> mean(b(:) - 1) ans = -0.1261 ??? shouldn't this be zero?

>> std(b(:) - 1) ans = 0.1844 ??? shouldn't this be sqrt(0.1)?

[Brian]

You really need to add this to your documentation of your demos. I did not expect

I = im2double(imread('cameraman.tif'));

to normalize the image [0,1]. If you look at the documentation for im2double, it states "rescaling the data if necessary". That is subjective. Necessary for what?

The best documentation has mathematical formulas and references.