How does matlab perform the image zero-centering in input layer?

6 views (last 30 days)
I am using the CNN to classify the objects. In input layer matlab itself implements the training data to zero-center by default.
I am wondering how it is performing this zero centering? Is it the per-pixel mean of channel-wise mean for image data? Is it dividing the images by traing data standard deviation?
inputlayer = imageInputLayer(inputSize,Name,Value)
'Normalization' Data transformation 'zerocenter' (default) | 'none'
Data transformation to apply every time data is forward-propagated through the input layer,
specified as the comma-separated pair consisting of 'Normalization' and one of the following.
'zerocenter' The software subtracts its mean from the training set.
'none' No transformation.
Example: 'Normalization','none'
Data Types: char

Answers (2)

Arnav Mendiratta
Arnav Mendiratta on 12 Jun 2017
Zerocenter normalization means having the data dimensions of approximately the same scale. This is done by dividing each dimension (channel) by its standard deviation once it has been zero-centered. Zero-centering means subtracting the mean from each of these dimension so that "data cloud" is centered towards the origin.
This answer talks about the motivation behind using this approach for CNN.
  1 Comment
Jay
Jay on 19 Jun 2017
Hi Arnav,
I understand what do you mean by zero-centering. But there are several ways we apply zero-centering.
1) using channel mean as you explained.
2) using the image mean/std(So rather than using the channel mean we are using the mean value of the image)
3) using the third dimension e.g.(1,1,x) pixel values where x is the red channel pixel locations (or all three channel image), that actually helps when we train networks with objects centred in the image patches.
I just wanted to know which mean you are using.
1)per-channel
2)per image or
3)per pixel?

Sign in to comment.


Erik Anzalone
Erik Anzalone on 26 Oct 2018
How can I find the zero centering? How AverageImage is computed?

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!