GAL-GAN: Global styles and local high-frequency learning based generative adversarial network for image cartoonization

Abstract

The style transfer of cartoon images has always been a challenging problem in computer vision. Currently, there are still two aspects that need to be improved in this field: (1) existing methods can only perform simple domain-to-domain cartoon style transfer, ignoring the global style information of the image, and (2) the neglect of local features in image style transfer, such as edge information and texture information, leads to lower quality of stylized images. To alleviate these two issues, we propose a novel global styles and local high-frequency learning based generative adversarial network (GAL-GAN) for image cartoonization. Specifically, the feature information of each channel is weighted by cartoon feature mapping to improve the quality of the global cartoon style of the generated image. In order to enrich the local feature information of generated images, we introduce a high-frequency learning strategy to reduce noise and enhance texture and detail extraction. Experiments reveal that GAL-GAN can generate high-quality stylized images with a specific style and have advantages over current state-of-the-art models.