Sketch-style face image synthesis via constrained least squares estimation

This paper presents a new sketch-style face image generation algorithm dealing with the problem of converting a photorealistic human face image to its corresponding sketch-style one. We prepare two images sets - one is photorealistic face images set and the other is sketch-style face images set. In our model, an input test face image can approximately be viewed as a linear combination of already known face images. We seek an identical linear combination model for both input photorealistic face image and output sketch-style face image, and we retain holistic geometric information of a face image. The parameter, which is estimated under constrained least square criterion, from input photorealistic face image can be directly used to infer the output sketch-style face image. As a result, estimation error that is showed as jarring halo artifacts in resultant image is inevitable and then we utilize bilateral image filtering to reduce unfavorable artifacts and meanwhile reserve distinctive geometric information.