DeCoGAN: Photo Cartoonization Based on Deformation Consistency GAN

Cartoon style transfer via deep learning techniques, which can be widely used in games and comics, attracts potential attention in the field of computer vision. Considering the lack of paired-images, previous methods conventionally measure the semantic and style distinction by high-level features, resulting in unsatisfactory semantic preservation and texture description. To offer a more effective supervision, we propose a novel deformation consistency GAN (DeCoGAN) in the absence of paired training data for photo cartoonization, where we take up a deformation consistency assumption for an ideal style transfer model that the processing sequence of image style transfer and spatial deformation should be exchangeable. Besides, we propose to make use of average pooling to construct the color constraint for color extraction and style description. Experiment results validate that our proposal can dramatically improves the semantic stability and the detail preservation.