Face localization using template matching method based on new statistical metrics

Brahim Belhaouari Samir*, Amal Seralkhatem Osman Ali, Nadir Nourain

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

Over the past decade face recognition has emerged as an active research area and template matching approaches have been widely used in this context. Normalized Cross-Correlation (NCC) is a measurement method normally utilized to compute the similarity between face templates and rectangular blocks of an input image in order to locate the face position. However, the NCC metric has always been criticized in locating faces especially images including illumination, scale, pose and shape variations. To address this problem, a novel template-matching techniques based on Chi-Square metric (CHI2) and Sum Square of Student's t-distribution metric (SST) are developed in this paper. When using Chi-Square metric; in order to locate the face exactly, the template size should have the same size as the face. For that we calculate p-value instead of chi-square metric for different sizes. This proposed concept can be implemented through considering both template and dynamic windows as probability density functions, so all pixel values are positive and their sum is equal to one. To expand the Sum Square of Student's t-distribution metric, we used the Central Limit Theorem and the fact that t-distribution square of freedom n-1 is a Fisher distribution of freedom 1 and n-1, the Sum Square t-distribution Normalized (SSTN) converges in distribution to standard normal distribution. Then we calculate SSTN for all dynamic windows to extract all suspected windows as faces corresponding to the location where the lower local minimums approach the global minimum. Using face datasets (Yale, MIT-CBCL, Caltech and BioID), extensive experiments are performed to verify the efficiency of the proposed techniques and to compare them with other techniques of face localization. The results show the superiority of our proposed approaches over other techniques in terms of accuracy of face localization.

Original languageEnglish
Title of host publicationInternational Conference on Fundamental and Applied Sciences 2012, ICFAS 2012
Pages85-90
Number of pages6
DOIs
Publication statusPublished - 2012
Externally publishedYes
Event2nd International Conference on Fundamental and Applied Sciences 2012, ICFAS 2012 - Kuala Lumpur, Malaysia
Duration: 12 Jun 201214 Jun 2012

Publication series

NameAIP Conference Proceedings
Volume1482
ISSN (Print)0094-243X
ISSN (Electronic)1551-7616

Conference

Conference2nd International Conference on Fundamental and Applied Sciences 2012, ICFAS 2012
Country/TerritoryMalaysia
CityKuala Lumpur
Period12/06/1214/06/12

Keywords

  • Face and gesture recognition
  • Pattern analysis
  • Pattern matching
  • Statistical computing
  • similarity measures

Fingerprint

Dive into the research topics of 'Face localization using template matching method based on new statistical metrics'. Together they form a unique fingerprint.

Cite this