Learning semantic priors for texture-realistic sketch-to-image synthesis

Zeyu Li,Cheng Deng,Kun Wei,Wei Liu,Dacheng Tao
DOI: https://doi.org/10.1016/j.neucom.2021.08.085
IF: 6
2021-11-01
Neurocomputing
Abstract:Sketch-to-image synthesis is a challenging task in the field of computer vision that generates photo-realistic images from given sketches. Existing methods of this kind are unable to discover the inherent semantic information contained in an image and use it to guide the synthesis process, substantially reduce their capacity to generate photo-realistic images. Accordingly, in this paper, we propose a novel framework that explores and leverages semantic information to generate realistic textures in synthesized images for this task. More specifically, the segmentation maps generation network is designed to learn the relationships between sketches and segmentation maps in order to obtain the semantic segmentation maps from the sketches. Taking semantic segmentation maps as the condition, a feature-wise affine transformation is then executed to change the feature maps of intermediate layers in the network, which can efficiently generate the texture required to synthesize more photo-realistic images. Extensive experiments demonstrate that when compared to other state-of-the-art sketch-to-image synthesis methods, our approach can not only synthesize images with significantly superior visual quality but is also able to achieve better results on quantitative metrics.
computer science, artificial intelligence
What problem does this paper attempt to address?