Hey guys!
First, I want to thank Jeremy, Rachel and everyone else involved for these fantastic lessons.
I had a chance to view lesson 14 on superresolution and noticed that the bluebird image still suffers from the checkerboard artifacts. It’s quite apparent if you zoom in a bit.
I worked with superresolution for some time on astronomical images (which are apparently quite different from natural images) and noticed that using feature maps in the perceptual loss BEFORE the relu helps to alleviate this issue. I’m using Keras and can’t try it with the fastai library. Can you guys try it to confirm this?
A few additional notes. I also noticed that using anything besides block1_conv1 and block1_conv2 in VGG16 perceptual loss also tends to add artifacts (at least on astronomical images). Can be also worth checking.
Best regards,
Dennis
