基于CycleGAN的非配对人脸图片光照归一化方法  

An Unpaired Face Illumination Normalization Method Based on CycleGAN

作　　者：曾碧[1] 任万灵 陈云华[1] Zeng Bi;Ren Wan-ling;Chen Yun-hua(School of Computers,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学计算机学院,广东广州510006

出　　处：《广东工业大学学报》2018年第5期11-19,共9页Journal of Guangdong University of Technology

基　　金：国家自然科学基金资助项目(61703115);广东省自然科学基金资助项目(2016A030313713);广东省重大科技专项项目(2016B010108004);广东省产学研合作专项项目(2014B090904080)

摘　　要：针对人脸识别过程中光照对识别结果的影响问题,提出了一种基于CycleGAN的光照归一化方法.使用了生成对抗式的网络结构,利用图像翻译的原理,将较亮图片的光照风格迁移至较暗图片,同时保持原人脸表面平滑且结构基本不变.使用非配对的数据集,无需人工标注标签,简化了数据准备阶段的工作,达到了利用无监督的深度学习方法去除图片光照影响的目的.最后用训练好的模型处理CroppedYale测试集,比较处理前后的人脸识别准确率.实验证明,本文方法具有较强的降低人脸光照对识别率影响的能力且基本不改变人脸结构,有利于提高人脸识别的准确率.Aiming at the influence of illumination in face recognition process,a method of illumination normalization based on CycleGAN is presented.By using the Generative Adversarial Nets and the principle of image translation,the illumination style of the darker image was shifted to the brighter image,while the surface and the structure of the face kept smooth at the same time.Unpaired data sets without labels are used,in order to achieve unsupervised removal of illumination,and greatly simplify the work of data preprocessing.Finally,the CroppedYale data set is used to train a deep learning face recognition model,using this CycleGAN model to process the test set,and comparing the accuracy before and after processing.Experiments show that this method has a strong ability to reduce the influence of human face illumination on recognition rate while basically not changing the face structure,and therefore is helpful to improve the accuracy of face recognition.

关 键 词：生成对抗网 深度学习 人脸识别 光照归一化 人脸光照处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]