I have some doubts on how to use the mean of a pre-trained CNN.
For example, if I use VGG-16 net, before the classification of an image I have to subtract the mean of the dataset (ImageNet in this case) to every pixel, am I right? And after this, do I have to normalize the pixel values in range [0,1] too?
For the fine tuning of a pre-trained network, how can I use the mean? Do I have to perform the same process described above for the images in the new dataset before starting the training?