Per Image Normalization

Could you explain, what this method is doing, as the docs don’t seem to give much information:

Linearly scales each image in  `image`  to have mean 0 and variance 1.

For each 3-D image  `x`  in  `image` , computes  `(x - mean) / adjusted_stddev` , where
* `mean`  is the average of all values in  `x`
* `adjusted_stddev = max(stddev, 1.0/sqrt(N))`  is capped away from 0 to protect against division by 0 when handling uniform images
  * `N`  is the number of elements in  `x`
  * `stddev`  is the standard deviation of all values in  `x`

I don’t understand what the “per image” part of this normalization is, if the mean is the "average of all values in x and the stddev also seems to use all elements.