Research Of Face Sketch-to-Photo Synthesis Methods Based On GAN

Face image synthesis is a very challenging research topic in the field of computer vision,and it has great application value in law enforcement and digital entertainment.Compared with the traditional method,the synthesis method based on Generative Adversarial Networks(GAN)has the advantages of more realistic synthetic image.In this paper,face photo synthesis and face sketch sequence synthesis based on Generative Adversarial Networks(GAN)are studied.The main work includes two aspects:(1)Face sketch to photo synthesis based on GAN.A high-fidelity face sketch generation method based on generative adversarial network(GAN)was proposed.U-Net is used as a generator and Patch-GAN with residual blocks is used as a discriminator.An effective loss function combination is designed by constraining the pixels,edges,and highlevel features of the composite photo.In addition,CUHK's sketch dataset was augmented with effective sampling methods.Qualitative and quantitative experiments verify that this method is superior to other methods.Developed a sketch-based photo editing application.(2)Face photo to sequence sketches synthesis based on GAN.A sequential enhanced face sketch dataset was first constructed,called Ord-Sketch.The dataset contains 400 photos of faces,each with 25 sketches,and the fineness of each sketch is increasing,providing researchers with support for face synthesis.Then,based on the constructed OrdSketch dataset,a multi-stage cascade network called SO-Net is proposed.Each stage uses a GAN architecture,of which 5 intermediate sketches are used as prior constraints.The experimental results show that this method surpasses other methods in a variety of evaluation indicators and visual effects.