基于双层生成对抗网络的素描人脸合成方法  

FACE SKETCH SYNTHESIS METHOD BASED ON DOUBLE LAYER GENERATIVE ADVERSARIAL NETWORKS

作　　者：李凯旋 曹林[1] 杜康宁 Li Kaixuan;Cao Lin;Du Kangning(Department of Communication Engineering,Beijing Information Science and Technology University,Beijing 100101,China)

机构地区：[1]北京信息科技大学通信工程系

出　　处：《计算机应用与软件》2019年第12期176-183,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61671069)

摘　　要：为解决传统素描人脸合成方法中素描人脸图像细节模糊和清晰度低的问题,提出一种基于双层生成对抗网络的素描人脸合成方法。该方法学习面部照片与素描人脸图像之间的映射关系,并通过双层网络将映射关系限制为一对一映射;利用重建损失函数约束生成网络,提高合成能力;通过生成网络与判别网络的对抗训练,优化网络参数,合成最终素描人脸图像。通过在CUHK素描人脸库上的对比实验,证明该方法合成的素描人脸图像质量明显优于其他传统素描人脸合成方法,其合成的素描人脸图像面部细节更完整,清晰度更高。Traditional face sketch synthesis methods have blurred details and low definition in face sketch images.To solve this problem,we propose face sketch synthesis method based on double layer generative adversarial networks.The method learned the mapping relationship between facial photos and face sketch images,and limited to one-to-one mapping by two-layer network.It constrained the generative network by reconstruction loss function so as to improve synthesis ability.The method optimized network parameters and synthesized face sketch images by adversarial training between generative network and discriminant network.The experiments on the CUHK face sketch database prove that the quality of the face sketch image synthesized by this method is better than other traditional face sketch synthesis methods.The synthesized face sketch image has more complete facial details and higher definition.

关 键 词：素描人脸合成 生成对抗网络 深度学习 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]