Introduction

First, instead of adding textures such as brush strokes in many other styles, cartoon images are highly simpliﬁed and abstracted from real-world photos.
Second, despite variation of styles among artists, cartoon images have noticeable common appearance — clear edges, smooth color shading and relatively simple textures — which is very different from other forms of artworks.

卡通画是现实图片的高度简化抽象
卡通画有明显的边缘、平滑的彩色阴影和相对简单的材质

主要贡献
(1) We propose a dedicated GAN-based approach that effectively learns the mapping from real-world photos to cartoon images using unpaired image sets for training.
不使用成对的训练数据
(2)We propose two simple yet effective loss functions in GAN-based architecture
提出了效果不错的GAN的损失函数
(3)We further introduce an initialization phase to improve the convergence of the network to the target manifold
一种初始化方法提升收敛性

Related Work

Non-photorealistic rendering

Stylization with neural networks

NTS
Deep Analogy

Image synthesis with GANs

CycleGAN
Several works [5, 14, 16] have provided GAN solutions to pixel-to-pixel image synthesis problems. However, these methods require paired image sets for the training process.
To address this fundamental limitation, CycleGAN [38] was recently proposed

CartoonGAN

CartoonGAN architecture

下采样 - 8个残差块 - 上采样
判断任务相对简单所以用了一个比较浅的网络

Loss function

然后通过极大极小求解参数

Adversarial loss
However, we observe that simply training the discriminator D to separate generated and true cartoon images is not sufﬁcient for transforming photos to cartoons. This is because the presentation of clear edges is an important characteristic of cartoon images, but the proportion of these edges is usually very small in the whole image.