1

I am new to CNN's, and have read in multiple places that inputs should have low magnitudes such as normalized to values of 0->1, -1->+1, standardized around zero and so on. I tried all of these methods but the accuracy is never as good as when I use a simple 0->255 input. I am guessing that this is because I'm using pre-trained imagenet weights which were trained on images with 0->255 pixels, meaning similar inputs work better than standardized or normalized ones. Could someone please confirm whether this is true or not? And if not, why normalizing and standardizing can reduce accuracy? Thanks.

numunu
  • 25
  • 5
  • I do not have definite answer, but may this is an element of response: https://www.quora.com/Why-does-the-accuracy-decreases-after-I-did-data-normalization-in-CNN-model – Gregoire Cattan May 28 '22 at 20:13

0 Answers0