[PYTHON] [Edge2Imori] Epi.2 pix What if you add Gradient Penalty to pix? ??
Struggle to make Edge2Imori that converts a line into a newt image after inputting
Previous article I tried to make edge2imori by applying pix2pix
Created using pix2pix.
This time, I applied Gradient-Penalty of WGAN-GP.
First of all, collecting images
Images of newts were collected using iCrawler.
The story that iCrawler was the strongest in image crawling
This has collected 630 images of newts.
Add Gradient Penalty to pix2pix
I did the manmakoto of the title.
Gradient Penalty was proposed in WGAN-GP and is a term that plays a role of regularization to bring the gradient closer to 1.
Code created this time >> https://github.com/yoyoyo-yo/DeepLearningMugenKnock/blob/master/Question_imageGenerate/answers/pix2pixGP_pytorch.py
result
The result looks like this
The input was an image in which edges were extracted from the correct image by the canny method.
There are quite a few images of newts. It seemed that the part where the line was thin was a newt, and when the line was dense, it was a grass.
Left: Input, Center: pix2 pix-GP result, Right: Correct answer
Recommended Posts