What is the "learnable blur" used in fast.ai's Decrappify?

Hello. Does anyone know what the “learnable blur” used in fastAI’s decrappify is? It’s mentioned at about 9:30 in the video at the top of this page:


I’ve tried looking through the course notes, and I could never find an explanation there, either (they touch on something called blur, but as far as I can tell they never explain it).

I think it’s related to Super-Resolution using Convolutional Neural Networks without Any Checkerboard Artifacts according to the comments from the code.

However, I’m not really sure of what it does and why it is used…

1 Like

I had the same question again and looked further at the implementation. It is actually implemented as a simple average from a 2x2 kernel (a padding is added to keep same image dimensions).
It could be of any size but I guess 2x2 works pretty well.