Image style transfer based on decoupled representation
Linke Liu, Xiaoming Guo, Zejun Dai · 2022 Global Conference on Robotics, Artificial Intelligence and Information Technology (GCRAIT) · 2022
The present methods of image style transfer lack the ability to express the image content, which leads to the poor quality and lack of diversity of stylized images. This paper proposes a style transfer method based on image decoupled representation. Generator network is composed of content encoder and style encoder for feature extraction. Decoder generates target image; The discriminator enhances the rationality of the target image. Define two mapping functions to map content features in domain-crossed shared content space to domain independent content spaces. Using Photo2Monet, Cat2Dog, and photographic 2portrait data sets, we found that the domain independent content spaces better represents the content, and style transfer effect is better. The average FID index value was 56.59. The average value of LPIPS index was 0.3749.