Performance evaluation study of data selection and matching criteria in face recognition for human surveillance

Face recognition is an important biometric in many fields, such as access control and surveillance. Currently there are many published reports in face recognition for surveillance application. However, there was no standard surveillance database, no standard evaluation criteria, and no standard meth...

Full description

Saved in:

Bibliographic Details
Main Author:	Ervin Gubin Moung
Format:	Thesis
Language:	English English
Published:	2018
Subjects:	TA1501-1820 Applied optics Photonics
Online Access:	https://eprints.ums.edu.my/id/eprint/37709/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/37709/2/FULLTEXT.pdf
Tags:	Add Tag No Tags, Be the first to tag this record!

id	my-ums-ep.37709
record_format	uketd_dc
spelling	my-ums-ep.377092023-11-29T02:21:38Z Performance evaluation study of data selection and matching criteria in face recognition for human surveillance 2018 Ervin Gubin Moung TA1501-1820 Applied optics. Photonics Face recognition is an important biometric in many fields, such as access control and surveillance. Currently there are many published reports in face recognition for surveillance application. However, there was no standard surveillance database, no standard evaluation criteria, and no standard method published for selecting training and testing data used by all researchers. Thus, a comparison between the various results reported is hard to be made. The aim of this work is to establish a standard training database selection method based on surveillance database suitable for face recognition for surveillance application and also to establish. a test bed and test criteria for evaluating common reported approach. Two surveillance databases are used; ChokePoint Pre-processed Grayscale Database (PPG) and ChokePoint Manually Preprocessed Colour Images Database (MPCI). Three sessions are used to acquire the images in the database. The images in each session were equally divided into three classes: CLOSE, MEDIUM, and FAR. A commonly used PCA-based face recognition system has been selected for this work. The effect of the distance between the subject and the camera, the effect of the number of images per class, the effect of mean image, the effect of training database size, and the effect of database sessions on face recognition have been investigated. It was found that using images from the FAR class for training gives better performance compared to MEDIUM or CLOSE class. However, it was found that using one image from each class gives better recognition performance compared to using three FAR class images for training. It was also found that as the number of images per class increases, the recognition rate increases. Finally, it was found that using one mean image per class from all the available database sessions gives the best performance. The performance of YCsCR individual channels on face recognition has been investigated and compared with grayscale. It was found that grayscale performed better than all the individual YCsCR channels because grayscale has better quality of visual features. A fusion strategy using the individual YCsCR has been presented and compared to grayscale performance. It was found out that the fusion of CsxCR with any other channel outperforms the grayscale when three images of the same class from the same database are used for training (Case 1). For YCBCR individual channels, the best performance is achieved by using the CsxCR channel. It was also found that if the differences between individual channels performance vary significantly, the individual channel performance become an important criteria when selecting channels for fusion. In general, increasing the number of fused channels increases the performance of the system. A comparison of PPG and MPCI databases performance and also a comparison of Euclidean distance (ED) and Euclidean distance with SVM (ED+SVM) performance have been made. It was found that the recognition rate pattern stays the same regardless of the training database used and similarity matching method used. Processing time wise, ED is much more efficient compared to ED+SVM. The best recognition performance is achieved by PCA-based (ED) system using the MPCI database and ED matching method with Case 6 training database criteria, giving 100% average correct recall and reject rate, and uses 5.69 seconds process time for a single test person and 26. 7 MB space for training data. 2018 Thesis https://eprints.ums.edu.my/id/eprint/37709/ https://eprints.ums.edu.my/id/eprint/37709/1/24%20PAGES.pdf text en public https://eprints.ums.edu.my/id/eprint/37709/2/FULLTEXT.pdf text en validuser dphil doctoral Universiti Malaysia Sabah Fakulti Kejuruteraan
institution	Universiti Malaysia Sabah
collection	UMS Institutional Repository
language	English English
topic	TA1501-1820 Applied optics Photonics
spellingShingle	TA1501-1820 Applied optics Photonics Ervin Gubin Moung Performance evaluation study of data selection and matching criteria in face recognition for human surveillance
description	Face recognition is an important biometric in many fields, such as access control and surveillance. Currently there are many published reports in face recognition for surveillance application. However, there was no standard surveillance database, no standard evaluation criteria, and no standard method published for selecting training and testing data used by all researchers. Thus, a comparison between the various results reported is hard to be made. The aim of this work is to establish a standard training database selection method based on surveillance database suitable for face recognition for surveillance application and also to establish. a test bed and test criteria for evaluating common reported approach. Two surveillance databases are used; ChokePoint Pre-processed Grayscale Database (PPG) and ChokePoint Manually Preprocessed Colour Images Database (MPCI). Three sessions are used to acquire the images in the database. The images in each session were equally divided into three classes: CLOSE, MEDIUM, and FAR. A commonly used PCA-based face recognition system has been selected for this work. The effect of the distance between the subject and the camera, the effect of the number of images per class, the effect of mean image, the effect of training database size, and the effect of database sessions on face recognition have been investigated. It was found that using images from the FAR class for training gives better performance compared to MEDIUM or CLOSE class. However, it was found that using one image from each class gives better recognition performance compared to using three FAR class images for training. It was also found that as the number of images per class increases, the recognition rate increases. Finally, it was found that using one mean image per class from all the available database sessions gives the best performance. The performance of YCsCR individual channels on face recognition has been investigated and compared with grayscale. It was found that grayscale performed better than all the individual YCsCR channels because grayscale has better quality of visual features. A fusion strategy using the individual YCsCR has been presented and compared to grayscale performance. It was found out that the fusion of CsxCR with any other channel outperforms the grayscale when three images of the same class from the same database are used for training (Case 1). For YCBCR individual channels, the best performance is achieved by using the CsxCR channel. It was also found that if the differences between individual channels performance vary significantly, the individual channel performance become an important criteria when selecting channels for fusion. In general, increasing the number of fused channels increases the performance of the system. A comparison of PPG and MPCI databases performance and also a comparison of Euclidean distance (ED) and Euclidean distance with SVM (ED+SVM) performance have been made. It was found that the recognition rate pattern stays the same regardless of the training database used and similarity matching method used. Processing time wise, ED is much more efficient compared to ED+SVM. The best recognition performance is achieved by PCA-based (ED) system using the MPCI database and ED matching method with Case 6 training database criteria, giving 100% average correct recall and reject rate, and uses 5.69 seconds process time for a single test person and 26. 7 MB space for training data.
format	Thesis
qualification_name	Doctor of Philosophy (PhD.)
qualification_level	Doctorate
author	Ervin Gubin Moung
author_facet	Ervin Gubin Moung
author_sort	Ervin Gubin Moung
title	Performance evaluation study of data selection and matching criteria in face recognition for human surveillance
title_short	Performance evaluation study of data selection and matching criteria in face recognition for human surveillance
title_full	Performance evaluation study of data selection and matching criteria in face recognition for human surveillance
title_fullStr	Performance evaluation study of data selection and matching criteria in face recognition for human surveillance
title_full_unstemmed	Performance evaluation study of data selection and matching criteria in face recognition for human surveillance
title_sort	performance evaluation study of data selection and matching criteria in face recognition for human surveillance
granting_institution	Universiti Malaysia Sabah
granting_department	Fakulti Kejuruteraan
publishDate	2018
url	https://eprints.ums.edu.my/id/eprint/37709/1/24%20PAGES.pdf https://eprints.ums.edu.my/id/eprint/37709/2/FULLTEXT.pdf
_version_	1794022679101046784

Performance evaluation study of data selection and matching criteria in face recognition for human surveillance

Similar Items