CASIA OpenIR  > 模式识别国家重点实验室  > 视频内容安全
Diagnosing deep learning models for high accuracy age estimation from a single image
Xing, Junhang1; Li, Kai2; Hu, Weiming1,2; Yuan, Chunfeng1; Ling, Haibin3
Source PublicationPATTERN RECOGNITION
2017-06-01
Volume66Issue:1Pages:106-116
SubtypeArticle
AbstractGiven a face image, the problem of age estimation is to predict the actual age from the visual appearance of the face. In this work, we investigate this problem by means of the deep learning techniques. We comprehensively diagnose the training and evaluating procedures of the deep learning models for age estimation on two of the largest datasets. Our diagnosis includes three different kinds of formulations for the age estimation problem using five most representative loss functions, as well as three different architectures to incorporate multi-task learning with race and gender classification. We start our diagnoses process from a simple baseline architecture from previous work. With appropriate problem formulation and loss function, we obtain state-of-the-art performance with the simple baseline architecture. By further incorporating our newly proposed deep multitask learning architecture, the age estimation performance is further improved with high-accuracy race and gender classification results obtained simultaneously. With all the insights gained from the diagnosing process, we finally build a deep multi-task age estimation model which obtains a MAE of 2.96 on the Morph II dataset and 5.75 on the WebFace dataset, both of which improve previous best results by a large margin.
KeywordAge Estimation Deep Learning Multi-task Learning
WOS HeadingsScience & Technology ; Technology
DOI10.1016/j.patcog.2017.01.005
WOS KeywordFACE IMAGES
Indexed BySCI
Language英语
Funding Organization973 Basic Research Program of China(2014CB349303) ; Natural Science Foundation of China(61472421, ; CAS(XDB02070003) ; U1636218 ; 61672519 ; 61303178)
WOS Research AreaComputer Science ; Engineering
WOS SubjectComputer Science, Artificial Intelligence ; Engineering, Electrical & Electronic
WOS IDWOS:000397371800012
Citation statistics
Cited Times:32[WOS]   [WOS Record]     [Related Records in WOS]
Document Type期刊论文
Identifierhttp://ir.ia.ac.cn/handle/173211/15074
Collection模式识别国家重点实验室_视频内容安全
Affiliation1.Chinese Acad Sci, Inst Automat, Natl Lab Pattern Recognit, Beijing 100190, Peoples R China
2.Univ Chinese Acad Sci, Chinese Acad Sci, Inst Automat, CAS Ctr Excellence Brain Sci & Intelligence Techn, Beijing 100190, Peoples R China
3.Temple Univ, Dept Comp & Informat Sci, Philadelphia, PA 19122 USA
First Author AffilicationChinese Acad Sci, Inst Automat, Natl Lab Pattern Recognit, Beijing 100190, Peoples R China
Recommended Citation
GB/T 7714
Xing, Junhang,Li, Kai,Hu, Weiming,et al. Diagnosing deep learning models for high accuracy age estimation from a single image[J]. PATTERN RECOGNITION,2017,66(1):106-116.
APA Xing, Junhang,Li, Kai,Hu, Weiming,Yuan, Chunfeng,&Ling, Haibin.(2017).Diagnosing deep learning models for high accuracy age estimation from a single image.PATTERN RECOGNITION,66(1),106-116.
MLA Xing, Junhang,et al."Diagnosing deep learning models for high accuracy age estimation from a single image".PATTERN RECOGNITION 66.1(2017):106-116.
Files in This Item: Download All
File Name/Size DocType Version Access License
xing_Diagnosing deep(1247KB)期刊论文作者接受稿开放获取CC BY-NC-SAView Download
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[Xing, Junhang]'s Articles
[Li, Kai]'s Articles
[Hu, Weiming]'s Articles
Baidu academic
Similar articles in Baidu academic
[Xing, Junhang]'s Articles
[Li, Kai]'s Articles
[Hu, Weiming]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[Xing, Junhang]'s Articles
[Li, Kai]'s Articles
[Hu, Weiming]'s Articles
Terms of Use
No data!
Social Bookmark/Share
File name: xing_Diagnosing deep learning models for high accuracy age estimation from a single image.pdf
Format: Adobe PDF
All comments (0)
No comment.
 

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.