Senyao-Han/Outline_colorization

Using GAN for Anime style colorization

"# Outline_colorization"

Paper

TensorFlow implementation of U-net | Resdiual U-net | DeeplabV3+ that compare the differences between difference structures for Outline colorization works.

We also use the following algorithm:

• Spectral Normalization for Generative Adversarial Networks

• Squeeze-and-Excitation Networks

Here are some result in our test:

Setup

• Windows
• Tensorflow 1.6.0
• Python 3.6.4
• NVIDIA GPU 1080ti

Dataset

We got 110000 anime-style images from the internet.(We will upload later)

Parameter

• epoch 20
• Learning rate 0.0005
• Batch size 10
• Image size 256x256

How to use

• For training

python main.py --network (U_net | Res_U_net | DeeplabV3_Plus)

• For testing

python main.py --training False --network (U_net | Res_U_net | DeeplabV3_Plus)

Train

Sample will save in the ./sample and each time you will get an number of images according to the bacth size.

Test

Put the images which you want to test in the ./test and you will get the result in the ./result.