This is our project which can convert a selfie to a hand-painted cartoon image by GAN. This work is motivated by CartoonGAN. However, CartoonGAN cannot avoid the influence of background if background doesn’t belong to style. We add a segmentation network(Unet) to solve the problem. More information can be seen in the slide.

We build two new datasets

• a hand-painted cartoon image dataset;

  

  

• an identification photo dataset.

The following are sources of training image

• Baidu image

• Google image

• Taobao comments

The following are dataset information.

The following are our results.

You may see that the background of the 1st row is red. To make it more interesting, I changed the background of the 1st row to red color by photoshop. However, others are indeed directly generated by our model.