is image normalization necessary in the preprocessing?

[Jianyu Lin]

Hi guys,

I am new to caffe and deep learning so I am confused about the normalization step for preparing the input data to caffe CNN model.

Now I want to classify grayscale images using caffe, so I think it will be nice if I start by modifying the MNIST example code (the MNIST input is also grayscale image).

But I wonder whether it is necessary to do image normalization (e.g, to mean 0, variance 1) when we prepare the data? I didn't see this step in those CAFFE self-contained examples.

And I noticed that in the CIFAR10 and imageNet examples the mean of image is computed beforehand, but it seems that there's no scaling step after this mean value computation. So I wonder is it OK if I don't do any normalization towards the original data (intensity ~-1200 - 200)?

Many thanks.

Jianyu

[Jianyu Lin]

sorry, I just check the prototxt from MNIST example, it did scale the image. But what if I want to normalize the image into (mean=0 variance = 1) case?