COLOR-COMICS-IMAGE SKETCH-STYLE TRANSFORMATION BASED ON CONDITIONAL GENERATIVE ADVERSARIAL NETWORK

In this paper, a comics-sketch-style transformation algorithm basing on Generative Adversarial Network (GAN) is proposed. The goal of this paper is to improve Pix2Pix network so that it can automatically generate sketch image from comics image. All original comics images are selected from famous comics website to ensure the diversity and randomness of the data. Then they are processed by Photoshop to generate the sketch-style training samples as the training set for network. This paper improves Pix2Pix network by introducing the LBP (Local Binary Pattern) algorithm as the pre-process to extract the texture features and then reduce the network level from five layers to three layers to improve the accuracy of Generator in U-Net. The experimental results indicate that our proposed algorithm is superior to Pix2Pix in generating the comics-sketch-style images both subjectively and objectively.