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.
| Selfie | Numbers |
|---|---|
| celebA | 350 |
| identification photo | 340 |
| Style | Numers |
|---|---|
| Hand-painted | 311 |
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.