For image pre-processing what does Normalization correspond to

In Keras API 1.2.2 there is to type of normalization in ImageDataGenerator:

  • featurewise_std_normalization: Boolean. Divide inputs by std of the dataset, feature-wise.
  • samplewise_std_normalization: Boolean. Divide each input by its std.
    Which one is used in DLS image pre-processing?

None of the above. We are dividing the inputs by std of batch data.

Thanks for the info. Is it done “feature-wise” ? Is it your experience that it work better than doing it on the entire dataset ?