Staged Sketch-to-Image Synthesis via Semi-supervised Generative Adversarial Networks

Sketch-based image synthesis is a challenging problem in computer graphics and vision. Existing approaches either require exact edge maps or rely on the retrieval of existing photographs, which limits their applications in real-world scenarios. Accordingly in this work, we propose a staged semi-supervised generative adversarial networks based method for sketch-to-image synthesis, which can directly generate realistic images from novice sketches. More specifically, we first adopt a conditional generative adversarial network (CGAN) to extract class-wise representations from unpaired images. These class-wise representations are then exploited and incorporated with another CGAN, which are used to generate realistic images from sketches. By incorporating the class-wise representations, our method can leverage both the general class information from unpaired images and the targeted object information from input sketches. Additionally, this network architecture also enables us to take full advantage of widely available unpaired images and learn more accurate class representations. Extensive experiments demonstrate, compared with state-of-the-art image translation methods, our approach can achieve more promising results and synthesize images with significantly better Inception Scores and Fréchet Inception Distance.