Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (10): 117-124.DOI: 10.3778/j.issn.1002-8331.2002-0266

Previous Articles     Next Articles

Research on Face Recognition Technology of DFM-GAN Network in Cross-Age Simulation

WU Jie, DUAN Jin, DONG Suoqin, LI Yingchao   

  1. 1.College of Electronic Information Engineering, Changchun University of Science and Technology, Changchun 130022, China
    2.Basic Technology Laboratory, Institute of Space Optoelectronic Technology, Changchun University of Science and Technology, Changchun 130022, China
  • Online:2021-05-15 Published:2021-05-10



  1. 1.长春理工大学 电子信息工程学院,长春 130022
    2.长春理工大学 空间光电技术研究所 基础技术实验室,长春 130022


Aiming at the problem of the effect of age change on the face recognition rate, a cross-age face generation method DFM-GAN(Depth Feature Migration GAN) is proposed by combining Generative Adversarial Network(GAN) and deep feature migration, and cross-age experimental research on simulated face verification is carried out. Firstly, the real sample is mapped to the feature vector by a convolutional encoder, and then the vector is projected to the face set under the one-hot coding age condition by using a deconvolution generator. By migrating the database sample face texture style, semantic features and other attributes simulate the facial images of people at different ages to be tested, reducing the differences with database samples. At the same time, the sample data set is pre-processed by Gaussian edge blurring method, and the edge enhancement counter-loss function is introduced to make the generated image have clearer edges. Local color histogram matching is performed on the generated image to increase the contrast and increase the face recognition performance of cross-age person. A single-sample experiment at different ages and a multi-sample experiment at specified ages are carried out. Experimental measurements are performed using two indicators of face similarity and face distance. The results show that the cross-age data samples generated by DFM-GAN face images are similar. The average degree has been increased by 19.24 percentage points, and the average face distance has been reduced by 0.451, which has better feasibility and certain practical significance in the direction of face recognition across ages.

Key words: face verification, Generative Adversarial Network(GAN), depth feature migration, face simulation, cross-age recognition


针对年龄变化对人脸识别率影响的问题,结合生成式对抗网络(Generative Adversarial Network,GAN)与深度特征迁移提出一种跨年龄人脸生成方法DFM-GAN(Depth Feature Migration GAN),并进行跨年龄模拟人脸验证实验研究。首先通过卷积编码器将真实样本映射到特征向量,然后利用反卷积生成器将向量投影到独热编码年龄条件下的人脸集合,通过在特征空间中迁移数据库样本人脸纹理风格、语义特点等属性,模拟生成待检人员在不同年龄段的面部图像,减少与数据库样本之间的差异性。同时通过高斯边缘模糊的方法对样本数据集做预处理,引入边缘提升对抗损失函数,使生成图像具有更为清晰的边缘,对生成图像进行局部颜色直方图匹配,增加对比度,达到提高跨年龄人脸识别性能的目的。进行了单样本不同年龄实验与指定年龄多样本实验,以人脸相似度与人脸距两项指标进行实验测量,结果表明:跨年龄数据样本经过DFM-GAN生成后的人脸图像,相似度平均提高了19.24个百分点,人脸距离平均减少了0.451,在跨年龄人脸识别方向具有较好的可行性和一定的实际意义。

关键词: 人脸验证, 生成式对抗网络(GAN), 深度特征迁移, 人脸模拟, 跨年龄识别