Image Cartoonizer

Abstract

This report mainly focuses on the problem of transferring the style of real-life scenes into cartoon images by implementing previous work done by CartoonGAN and proposed a lighter and faster model, which is a very challenging and meaningful task in artistic style transfer and computer vision area. There are several existing methods have been proposed. There are several reasons that: 1) the transferred images don’t have really obvious cartoon features; 2) the transferred images usually have a great loss of content of original images; 3) the required network to do the transformation needs lots of memory capacity. Thus, the CartoonGAN have been proposed, a generative adversarial network (GAN) framework for cartoon stylization. This framework uses unpaired real-life images and cartoon images to do the training, which is very convenient. CartoonGAN also proposed two different loss functions 1) A formula which is a sparse regularization in the high-level feature maps of VGG network, also called semantic loss. This loss functions can deal with great style variation between real-life images and cartoon images. 2) An adversarial loss function for preventing loss of clear edges. By our implementation, this network needs very little memory capacity and can be able to generate quite good cartoon images, which outperforms state-of-arts methods.