search

thomaskuestner / CNNArt

Automatic and reference-free MR artifact detection
Apache License 2.0
25 stars 17 forks source link

basic VAE for patch size 40 #49

Closed TokenJan closed 6 years ago

TokenJan commented 6 years ago

I’ve built a VAE network last week and trained it with Head images with and without motion artifact in patch size 40, 40, this is the architecture of the VAE network. The default padding is same, kernel sizes are all 3x3 and activation function is LeakyRelu. VAE architecture

Following pictures show the result of this architecture. It can predict the rough structures quite well and the motion artifact is reduced to some extent but some of the details are lost during the process as well. This is just a very basic network still needs to be optimized, but I think it is time to merge it into the master first. result1

result2

thomaskuestner commented 6 years ago

yes the results look promising, see slack...