耦合关系自学习的人脸年龄估计研究

PDF下载HTML

Title:

Research on the coupled-relationships self-learning human facial age estimation

作者:

田青^1,2, 毛军翔^1,3, 曹猛¹

1. 南京信息工程大学 计算机与软件学院，江苏 南京 210044;
2. 南京信息工程大学 数字取证教育部工程研究中心，江苏 南京 210044;
3. 东南大学 计算机科学与工程学院，江苏 南京 210096

Author(s):

TIAN Qing^1,2, MAO Junxiang^1,3, CAO Meng¹

1. School of Computer and Software, Nanjing University of Information Science & Technology, Nanjing 210044, China;
2. Engineering Research Center of Digital Forensics, Ministry of Education, Nanjing University of Information Science & Technology, Nanjing 210044, China;
3. School of Computer Science and Engineering, Southeast University, Nanjing 210096, China

关键词:

人脸年龄估计; 耦合关系; 特征关系; 编码关系; 输入输出关系; 关系自学习; 交替优化; 深度架构

Keywords:

human facial age estimation; coupled relationship; feature relationship; coding relationship; input-output relationship; relationships self-learning; alternating optimization; deep architecture

分类号:

TP391

DOI:

10.11992/tis.202101020

摘要:

在目前已提出多种人脸年龄估计(age estimation， AE)潜在关系挖掘的工作，绝大多数工作仅局限于挖掘单一层面的潜在关系，极少考虑多层面耦合关系的挖掘。因此，本文提出一种耦合关系自学习的AE模型CRSAE，以此挖掘输入特征关系、输出编码关系以及输入输出关系3种耦合关系，提高AE模型的泛化能力。首先对投影矩阵的行列协方差矩阵建模，构建输入特征关系与输出编码关系正则项。其次，本文通过引入一个结构矩阵，发掘输入输出关系。随后，为有效求解CRSAE模型，本文构建一种交替优化方法。鉴于面部特征具有高度非线性的特征，本文在所提出模型的基础上引入深度架构进一步提升模型的泛化能力。最后，通过在多个人脸图像数据集上的年龄评估实验，验证了所提模型的有效性和性能优越性。

Abstract:

Although a variety of human facial age estimation (AE) potential relationship-exploiting works have been proposed, most of them are limited to exploiting one-sided potential relationships, rarely considering multi-sided coupled relationships. Therefore, we propose a coupled relationships self-learning age estimation model—CRSAE, which can exploit three kinds of potential relationships, i.e., input feature relationships, output coding relationships, and input-output relationships, to improve the generalization of AE models. Specifically, the row and column covariance matrices of the projection matrix are modeled to construct the regularizer of the feature and coding relationships. The input-output relationships are then exploited through a structure matrix. To solve our proposed CRSAE model effectively, we present an alternating optimization algorithm. In view of the highly nonlinear characteristics of facial features, we also extend our proposed model with a deep architecture to further enhance its generalization. Finally, evaluation experiments are conducted to demonstrate the effectiveness and superiority of our proposed methods on multiple human facial datasets.