基于深度流形学习的人脸年龄识别
Facial Age Recognition Based on Deep Manifold Learning
张会影 1圣文顺 1金鑫2
作者信息
- 1. 南京工业大学浦江学院,江苏南京 211200
- 2. 蚌埠学院艺术设计学院,安徽蚌埠 233030
- 折叠
摘要
现有的人脸年龄识别方法大多利用深度学习框架提取人脸特征来识别年龄,但深度学习方法提取的高维人脸特征往往包含大量的冗余信息,不利于人脸年龄的识别.为了提高人脸年龄识别算法的精度和鲁棒性,提出了一种基于深度流形学习(Deep Manifold Learning,DML)的算法,采用深度学习提取人脸特征,通过流形学习选择具有判别性的人脸特征,将深度学习提取的高维人脸特征嵌入到低维的判别子空间上识别年龄.在公开的人脸数据库MORPH和FG-NET上对DML算法进行了实验,结果表明DML方案平均绝对误差(Mean Absolute Error,MAE)大幅度降低,不同误差值下识别累积评分(Cumulative Score,CS)明显提高,显著优于当前流行的人脸年龄识别方法.
Abstract
Most of existing face age recognition methods use deep learning framework to extract face features to identify age,but high-dimensional face features extracted by deep learning methods often contain a lot of redundant information,which is not conducive to face age recognition.In order to improve the accuracy and robustness of face age recognition algorithm,an algorithm based on Deep Manifold Learning(DML)is proposed.DML first uses deep learning to extract face features,and then selects discriminative face features through manifold Learning,that is,high-dimensional face features extracted by deep learning are embedded into a low-dimensional discriminant subspace to identify age.Experiments on the DML algorithm are carried out on the public face databases MORPH and FG-NET.Experi-ment results show that the Mean Absolute Error(MAE)of DML is significantly reduced,and the Cumulative Score(CS)is significantly improved under different error values,which is significantly superior to current popular face age recognition methods.
关键词
年龄识别/流形学习/深度学习/卷积神经网络/特征提取/平均绝对误差Key words
age recognition/manifold learning/deep learning/convolutional neural network/feature extraction/MAE引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据