Feature Weight Learning CycleGAN for Stylization of Portrait-Sketch Synthesis

• 1、Laboratory of Mobile Robot and Intelligent Technology, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：In order to solve the problem of the generated image obtained by existing portrait-sketch synthesis algorithm which is not clear enough and the local area has less sketch texture, this paper proposes a cycle-consistency GAN based on feature weight learning (FWL-CycleGAN) which makes the interrelationship between channels in the neural network structure, so as to generate a sketch image with more facial details. In addition, for the sake of alleviating the contour blur caused by the pixel-to-pixel unalignment between the prediction and the groundtruth, this paper proposes feature matching loss function to present high-level expressions of generated images and real images for comparison in feature space, and finally produces a high-quality sketch image with a method of smoothing difference between adjacent pixels.

Keywords： Generative Adversarial Networks portrait-sketch feature weight learning feature matching loss function