An RCCL-based GAN for Illustrative Sketch Generation from Game Scenes
GAN, illustrative sketch, game scene, relaxed cycle consistency loss, attention map
Jihyeon YEOM, Heekyung YANG, Kyungha MIN
We present a GAN-based framework for producing illustrative sketches from game scenes. This framework includes a relaxed cycle consistency loss (RCCL) module that estimates the difference between the edges of game scenes and resulting sketches extracted by the holistically-nested edge detection (HED) scheme. We compare the two edges using the learned perceptual image patch similarity (LPIPS) metric to focus on clear line representations. We also employ an attention map that focuses on semantic areas to identify areas that need to be intensively transformed. Our framework consists of a style extraction module, a generator module, a discriminator module, and a relaxed cycle consistency loss module. First, We extract the styles of illustrative sketch images using the style extraction module. Next, we generate illustrative sketches via style attention maps extracted from game scenes using the generator module. We then process the generated sketch images to the discriminator module to obtain the probability that the generated sketch satisfies the quality of illustrative sketch. We also apply RCCL to maintain the structure of the game scenes in the generated illustrative sketches. We demonstrate the superiority of our framework by illustrative sketches from various game scenes.
KCI Link : https://www.kci.go.kr/kciportal/ci/sereArticleSearch/ciSereArtiView.kci?sereArticleSearchBean.artiId=ART002697933