2

I'm wondering about data normalization in CNN, how can we do it for the input images?, what can it add to the model's performance? and what are the main pre-processing techniques before doing the convolutional layer?

Happy to get extra-reading, thanks.

UPDATE:

I'm looking for how to practically normalize the data inputs, I didn't get it! . Additional reading (explanation, papers) will be so helpful.

Khalil Meg
  • 121
  • 4
  • 1
    Also https://stats.stackexchange.com/questions/336458/what-is-the-recommended-way-to-normalize-data-to-a-nn-or-cnn and https://stats.stackexchange.com/questions/185853/why-do-we-need-to-normalize-the-images-before-we-put-them-into-cnn – Sycorax Apr 26 '19 at 20:16
  • I have updated the question, I didn't understand the normalization procedure, thanks for help – Khalil Meg Apr 27 '19 at 21:17
  • The three questions that I've linked to appear to address this question. Can you elaborate what you don't understand? "I don't get it!" is not an answerable question. – Sycorax Apr 27 '19 at 23:31

1 Answers1

1

Normalizing the data prior to CNN will create more spherical error surfaces which will yield faster convergence of the gradient-based optimization procedure. If this step is skipped, some axes (determined by the eigenvectors of the loss function) will have steeper/flatter structures and consequently, the travel of our optimizer on the surface will get troubled.

You can standardize the images. However, generally, we map the images into 0-1 interval (by dividing to 255 for gray images) so that the values are bounded.

Monotros
  • 742
  • 4
  • 10