非受限条件下多级残差网络人脸图像年龄估计  

Age Estimation with Multilevel Residual Networks in Unconstrained Conditions

作　　者：张珂[1] 高策[1] 郭丽茹 苑津莎[1] 赵振兵[1] 李保罡[1] Zhang Ke;Gao Ce;Guo Liru;Yuan Jinsha;Zhao Zhenbing;Li Baogang(Department of Electronic and Communication Engineering,North China Electric Power University,Baoding 071000)

机构地区：[1]华北电力大学电子与通信工程系,保定071000

出　　处：《计算机辅助设计与图形学学报》2018年第2期346-353,共8页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61302163;61302105;61401154;61501185);北京市自然科学基金(4164101);河北省自然科学基金(F2015502062;F2016502062;F2016502101);中央高校基本科研业务费专项资金(2018MS094)

摘　　要：年龄是人的固有属性,在人的社会交往中起到了基础性作用,因此人脸图像的年龄估计是人工智能领域的重要问题之一.为了解决非受限条件下人脸图像年龄估计困难的问题,提出一种非受限条件下的多级残差网络人脸年龄估计方法.首先针对高分辨率图像数据集构建多级残差神经网络模型;然后采用Image Net数据集对多级残差网络进行预训练,以获得图像的基本特征表达;最后在非受限人脸年龄数据集上结合随机深度算法对网络模型进行微调.在非受限的Adience人脸年龄分类数据集上进行年龄分类对比实验的结果表明,该方法能够明显地提高非受限条件下人脸年龄估计的准确率,并在提高网络学习能力的同时有效地抑制小规模数据集带来的过拟合问题.Age is the inherent attribute of human,and plays a fundamental role in human social interaction.Therefore,age estimation is one of the important problems in the field of artificial intelligence.In order to solve the problems of age estimation in unconstrained conditions,a new CNN based method for age estimation in unconstrained conditions leveraging multilevel residual networks is proposed.Firstly,the multilevel residual neural network is constructed for the high-resolution image datasets.Secondly,the multilevel residual network is pre-trained by ImageNet dataset to obtain the basic features of the image.Thirdly,the multilevel residual network model is fine-tuned by combining the stochastic depth algorithm on the unconstrained face dataset.Finally,the comparative experiments for age estimation are performed on the unconstrained Adience dataset.Excellent experimental results show that this method can significantly improve the accuracy of age estimation in unconstrained conditions,and effectively alleviate the over-fitting problem caused by small-scale datasets while improving the learning ability of network.

关 键 词：多级残差网络 年龄估计 非受限条件 随机深度算法 ImageNet和Adience数据集 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]