属性约束的人脸素描-照片转换与识别  

Attribute Constrained Facial Sketch Photo Transfer and Recognition

作　　者：黄婕 赵启军[1,2] 吕泽均 HUANG Jie;ZHAO Qijun;LV Zejun(College of Computer Science,Sichuan University,Chengdu 610065;School of Information Science and Technology,Tibet University,Lhasa 850000)

机构地区：[1]四川大学计算机学院,成都610065 [2]西藏大学信息科学技术学院,拉萨850000

出　　处：《现代计算机》2021年第9期45-50,共6页Modern Computer

基　　金：国家重点研发计划项目(No.2017YFB0802300);国家自然科学基金项(No.61773270、61971005)。

摘　　要：目击证人描述和嫌疑人画像是侦破刑事案件的重要依据,然而尽管人脸识别技术近年来发展迅速,但是根据嫌疑人画像自动识别嫌疑人身份仍然极具挑战。大部分现有素描人脸识别方法或者在特征空间中减小嫌疑人画像与照片之间的模态差异,或者先在嫌疑人图像和照片之间进行转换再进行人脸比对,而都忽略了目击证人对嫌疑人的属性描述。本文提出将人脸属性引入素描与照片间的转换过程,在照片域和素描域中同时进行人脸比对,并融合两者结果以识别嫌疑人身份。本文给出基于条件和环形生成对抗网络(CycleGAN)的实现方案,在香港中文大学的素描人脸库与现有方法的对比实验结果证明本文方法的有效性。Eyewitness description and suspect’s portrait are important clues for criminal investigation.Despite the rapid development of face recogni⁃tion technology in recent years,it is still challenging to automatically identify suspects based on the suspect’s portrait.Most of existing sketch face recognition methods either reduce the modality gap between the suspect's portrait and photo in the feature space,or convert the suspect's portrait into photo or vice versa and then conduct face comparison in either sketch or photo domain.They all ignore the eyewit⁃ness description of the suspect’s attributes.In this paper,face attributes are introduced into the transformation process between sketch and photo,and face comparisons are conducted in photo and sketch domains simultaneously,and their results are fused to identify the person in the facial sketch.The proposed method is implemented based on CycleGAN.Evaluation on the CUHK Face Sketch Database with compari⁃son to existing methods proves the effectiveness of the proposed method.

关 键 词：素描人脸识别 素描-照片转换 生成对抗网络 风格迁移 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]