基于图像块比较网络的素描人脸合成方法  

Sketch face synthesis based on the comparison network of image similarity blocks

作　　者：司淑狄 罗倩[1,2] 郭亚男 杜康宁 曹林[1,2] SI Shudi;LUO Qian;GUO Yanan;DU Kangning;CAO Lin(School of Information and Communication Engineering,Beijing Information Science&Technology University,Beijing 100101,China;MOE Key Laboratory for Optoelectronic Measurement Technology and Instrument,Beijing Information Science&Technology University,Beijing 100101,China)

机构地区：[1]北京信息科技大学信息与通信工程学院,北京100101 [2]北京信息科技大学光电测试技术及仪器教育部重点实验室,北京100101

出　　处：《北京信息科技大学学报（自然科学版）》2022年第2期62-68,共7页Journal of Beijing Information Science and Technology University

基　　金：国家自然科学基金项目(62001033,U20A20163);北京市教委面上基金项目(KM202011232021);北京信息科技大学“勤信人才”培育计划(QXTCP C202108,QXTCP A201902)。

摘　　要：针对传统素描人脸合成方法生成的素描肖像清晰度低且风格特征不显著的问题,提出了一种基于图像块比较网络的素描人脸合成方法,通过内容编码器和风格编码器自适应融合生成素描肖像。此外,引入多层次图像块比较模块来丰富素描人脸细节,加入预先训练的VGG16模型提取特征并使用均方误差约束使得素描肖像风格突出。在公共素描人脸数据库上与其他传统方法进行对比实验,结果表明该方法生成的素描肖像清晰度更高、风格更显著,各定量指标值最优。Aiming at the problems of low definition and insignificant style features of sketch portraits generated by traditional sketch face synthesis methods,a sketch face synthesis method based on image block comparison network is proposed,which is generated by adaptive fusion of content encoder and style encoder.In addition,a multi-level image patch comparison module is introduced to enrich sketch face details,a pre-trained VGG16 model is added to extract features and the mean square error constraint is used to make the sketch portrait style stand out.A comparative experiment is carried out on the public sketch face database with other traditional methods.The results show that the sketch portrait generated by this method has higher definition,more prominent style,and the best quantitative index values.

关 键 词：深度学习 生成对抗网络 素描人脸合成 图像块比较 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]