Research On Sketch Face Synthesis Technology Based On Deep Learning

Face sketch synthesis technology is a branch of image synthesis,and it has been widely used in the field of criminal investigation security and digital entertainment,and has received more and more attention.Although many scholars have proposed various face sketch synthesis methods in recent years,the final synthesis results still have some problems.The main challenges in face sketch synthesis technology are how to ensure the integrity of facial details and the authenticity of textures.With the rapid development of the deep learning,face sketch synthesis technology has made new breakthroughs.The face sketch synthesis method based on deep learning improves the texture effect of synthesizing sketch and the authenticity of sketch,but the synthesis results still have some problems,such as the losing facial details and blurring issues.This article analyzes the problems of existing face sketch synthesis methods,and proposes some solutions to the challenges in face sketch synthesis.The main contribution contents are summarized as follows:1.A face sketch synthesis method based on double layer generative adversarial networks is proposed.Traditional face sketch synthesis methods have tedious image blocking and segmentation steps the synthesized result is prone to appear artifacts and lacks the texture of sketch.Firstly,the method learns the mapping relationship between facial photos and face sketch images by two-layer network.Then,the method constrains the generate network by reconstruction loss and improves synthesis ability.Finally,the method optimizes network parameters and synthesizes sketches by adversarial training between generative network and discriminant network.Experiments in the standard photo-sketch face datasets show that this method has better image quality and more realistic sketch texture than the traditional sketch face synthesis method.2.A face sketch synthesis method based on Feature Learning Generative Adversarial Network(FL-GAN)is proposed.The generative methods prone to neglect detailed information and thus lose some individual specific features,such as glasses and headdresses.The proposed FL-GAN consists of one Feature Learning(FL)module which contains a facial feature extraction network and one Adversarial Learning(AL)module which contains a generator and a discriminator.The FL module aims to learn the detailed information of the image,and guide the AL module to synthesize detail-preserving sketch.The AL Module aims to learn the structure and texture of sketch and improve the quality of synthetic sketch by adversarial learning strategy.In the training phase,the combination of an adversarial loss function with a control factor and a facial detail loss function is used to optimize the entire model,and improve the synthesis quality of the model.Extensive experimental results on three public photo-sketch face datasets confirm that the FL-GAN method synthesizes sketch is more detail features and sharper than the existing sketch face image synthesis method based on the generative model.