A view-based statistical system for multi-view face detection and pose estimation

Ju Chin Chen, James Jenn-Jier Lien

Research output: Contribution to journalArticle

15 Citations (Scopus)

Abstract

This study develops a novel statistical system for automatic multi-view face detection and pose estimation. The five-module detection system is based on significant local facial features (or subregions) rather than the entire face. The low- and high-frequency feature information of each subregion of the facial image are extracted and projected onto the eigenspace and residual independent basis space in order to create the corresponding PCA (principal component analysis) projection weight vector and ICA (independent component analysis) coefficient vector, respectively. Therefore, the proposed system has an improved tolerance toward different facial expressions, wide viewing angles, partial occlusions and lighting conditions. Furthermore, either projection weight vectors or coefficient vectors in the PCA or ICA space have divergent distributions and are therefore modeled by using the weighted Gaussian mixture model (GMM) rather than a single Gaussian model. The GMM weights and parameters of the GMM are estimated iteratively using the Expectation-Maximization (EM) algorithm. Face detection is then performed by conducting a likelihood evaluation process based on the estimated joint probability of the weight and coefficient vectors and the corresponding geometric positions of the subregions. The use of subregion position information can reduce the risk of false acceptances. Moreover, simple cascaded rejecter module is employed to exclude 85% of the non-face images in order to enhance the overall system performance. The computational overhead is further reduced by eliminating the requirement for a residual image reconstruction process in the ICA process. Finally, the performance of the proposed system is evaluated using challenging databases. The results not only demonstrate the ability of the system to automatically identify facial images with a high degree of accuracy, but also verify its ability to estimate the fine pose angles with 5° precision and an over 90% accuracy rate.

Original languageEnglish
Pages (from-to)1252-1271
Number of pages20
JournalImage and Vision Computing
Volume27
Issue number9
DOIs
Publication statusPublished - 2009 Aug 3

Fingerprint

Face recognition
Independent component analysis
Principal component analysis
Image reconstruction
Lighting

All Science Journal Classification (ASJC) codes

  • Signal Processing
  • Computer Vision and Pattern Recognition

Cite this

@article{30ccb8010ac64756828be236341cc1be,
title = "A view-based statistical system for multi-view face detection and pose estimation",
abstract = "This study develops a novel statistical system for automatic multi-view face detection and pose estimation. The five-module detection system is based on significant local facial features (or subregions) rather than the entire face. The low- and high-frequency feature information of each subregion of the facial image are extracted and projected onto the eigenspace and residual independent basis space in order to create the corresponding PCA (principal component analysis) projection weight vector and ICA (independent component analysis) coefficient vector, respectively. Therefore, the proposed system has an improved tolerance toward different facial expressions, wide viewing angles, partial occlusions and lighting conditions. Furthermore, either projection weight vectors or coefficient vectors in the PCA or ICA space have divergent distributions and are therefore modeled by using the weighted Gaussian mixture model (GMM) rather than a single Gaussian model. The GMM weights and parameters of the GMM are estimated iteratively using the Expectation-Maximization (EM) algorithm. Face detection is then performed by conducting a likelihood evaluation process based on the estimated joint probability of the weight and coefficient vectors and the corresponding geometric positions of the subregions. The use of subregion position information can reduce the risk of false acceptances. Moreover, simple cascaded rejecter module is employed to exclude 85{\%} of the non-face images in order to enhance the overall system performance. The computational overhead is further reduced by eliminating the requirement for a residual image reconstruction process in the ICA process. Finally, the performance of the proposed system is evaluated using challenging databases. The results not only demonstrate the ability of the system to automatically identify facial images with a high degree of accuracy, but also verify its ability to estimate the fine pose angles with 5° precision and an over 90{\%} accuracy rate.",
author = "Chen, {Ju Chin} and Lien, {James Jenn-Jier}",
year = "2009",
month = "8",
day = "3",
doi = "10.1016/j.imavis.2008.11.004",
language = "English",
volume = "27",
pages = "1252--1271",
journal = "Image and Vision Computing",
issn = "0262-8856",
publisher = "Elsevier Limited",
number = "9",

}

A view-based statistical system for multi-view face detection and pose estimation. / Chen, Ju Chin; Lien, James Jenn-Jier.

In: Image and Vision Computing, Vol. 27, No. 9, 03.08.2009, p. 1252-1271.

Research output: Contribution to journalArticle

TY - JOUR

T1 - A view-based statistical system for multi-view face detection and pose estimation

AU - Chen, Ju Chin

AU - Lien, James Jenn-Jier

PY - 2009/8/3

Y1 - 2009/8/3

N2 - This study develops a novel statistical system for automatic multi-view face detection and pose estimation. The five-module detection system is based on significant local facial features (or subregions) rather than the entire face. The low- and high-frequency feature information of each subregion of the facial image are extracted and projected onto the eigenspace and residual independent basis space in order to create the corresponding PCA (principal component analysis) projection weight vector and ICA (independent component analysis) coefficient vector, respectively. Therefore, the proposed system has an improved tolerance toward different facial expressions, wide viewing angles, partial occlusions and lighting conditions. Furthermore, either projection weight vectors or coefficient vectors in the PCA or ICA space have divergent distributions and are therefore modeled by using the weighted Gaussian mixture model (GMM) rather than a single Gaussian model. The GMM weights and parameters of the GMM are estimated iteratively using the Expectation-Maximization (EM) algorithm. Face detection is then performed by conducting a likelihood evaluation process based on the estimated joint probability of the weight and coefficient vectors and the corresponding geometric positions of the subregions. The use of subregion position information can reduce the risk of false acceptances. Moreover, simple cascaded rejecter module is employed to exclude 85% of the non-face images in order to enhance the overall system performance. The computational overhead is further reduced by eliminating the requirement for a residual image reconstruction process in the ICA process. Finally, the performance of the proposed system is evaluated using challenging databases. The results not only demonstrate the ability of the system to automatically identify facial images with a high degree of accuracy, but also verify its ability to estimate the fine pose angles with 5° precision and an over 90% accuracy rate.

AB - This study develops a novel statistical system for automatic multi-view face detection and pose estimation. The five-module detection system is based on significant local facial features (or subregions) rather than the entire face. The low- and high-frequency feature information of each subregion of the facial image are extracted and projected onto the eigenspace and residual independent basis space in order to create the corresponding PCA (principal component analysis) projection weight vector and ICA (independent component analysis) coefficient vector, respectively. Therefore, the proposed system has an improved tolerance toward different facial expressions, wide viewing angles, partial occlusions and lighting conditions. Furthermore, either projection weight vectors or coefficient vectors in the PCA or ICA space have divergent distributions and are therefore modeled by using the weighted Gaussian mixture model (GMM) rather than a single Gaussian model. The GMM weights and parameters of the GMM are estimated iteratively using the Expectation-Maximization (EM) algorithm. Face detection is then performed by conducting a likelihood evaluation process based on the estimated joint probability of the weight and coefficient vectors and the corresponding geometric positions of the subregions. The use of subregion position information can reduce the risk of false acceptances. Moreover, simple cascaded rejecter module is employed to exclude 85% of the non-face images in order to enhance the overall system performance. The computational overhead is further reduced by eliminating the requirement for a residual image reconstruction process in the ICA process. Finally, the performance of the proposed system is evaluated using challenging databases. The results not only demonstrate the ability of the system to automatically identify facial images with a high degree of accuracy, but also verify its ability to estimate the fine pose angles with 5° precision and an over 90% accuracy rate.

UR - http://www.scopus.com/inward/record.url?scp=67349116435&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=67349116435&partnerID=8YFLogxK

U2 - 10.1016/j.imavis.2008.11.004

DO - 10.1016/j.imavis.2008.11.004

M3 - Article

VL - 27

SP - 1252

EP - 1271

JO - Image and Vision Computing

JF - Image and Vision Computing

SN - 0262-8856

IS - 9

ER -