TY - JOUR
T1 - Discriminability and reliability indexes
T2 - Two new measures to enhance multi-image face recognition
AU - Zou, Weiwen
AU - YUEN, Pong Chi
N1 - Funding Information:
This project was partially supported by the Faculty Research Grant of Hong Kong Baptist University and NSFC-GuangDong Research Grant U0835005. The authors would like to thank CMU for providing the CMU-PIE database, Yale University for providing the YaleB database and University of Notre Dame for providing the FRGC database.
PY - 2010/10
Y1 - 2010/10
N2 - In order to handle complex face image variations in face recognition, multi-image face recognition has been proposed, instead of using a single still-image-based approach. In many practical scenarios, multiple images can be easily obtained in enrollment and query stages, for example, using video. By assessing these images, a good "quality" image(s) will be selected for recognition using conventional still-image-based recognition algorithms so that the recognition performance can be improved. However, existing methods do not fully utilize all images information. In this paper, two new measurements, namely discriminability index (DI) and reliability index (RI), are proposed to evaluate the enrolled and query images, respectively. By considering the distribution of enrolled images from individuals, the discriminability index of each image is calculated and a weight is assigned. For testing images, a reliability index is calculated based on matching quality between the testing images and enrolled images. If the reliability index of a testing image is small, the testing image will be discarded as it may degrade the recognition performance. To evaluate and demonstrate the use of DI and RI, we adopt the combining classifier method with eigenface representations in input and kernel feature spaces. CMU-PIE, YaleB and FRGC V2.0 databases are used for experiments. Experimental results show that the recognition performance, with three popular combination rules, can be increased by more than 10% on average using DI and RI.
AB - In order to handle complex face image variations in face recognition, multi-image face recognition has been proposed, instead of using a single still-image-based approach. In many practical scenarios, multiple images can be easily obtained in enrollment and query stages, for example, using video. By assessing these images, a good "quality" image(s) will be selected for recognition using conventional still-image-based recognition algorithms so that the recognition performance can be improved. However, existing methods do not fully utilize all images information. In this paper, two new measurements, namely discriminability index (DI) and reliability index (RI), are proposed to evaluate the enrolled and query images, respectively. By considering the distribution of enrolled images from individuals, the discriminability index of each image is calculated and a weight is assigned. For testing images, a reliability index is calculated based on matching quality between the testing images and enrolled images. If the reliability index of a testing image is small, the testing image will be discarded as it may degrade the recognition performance. To evaluate and demonstrate the use of DI and RI, we adopt the combining classifier method with eigenface representations in input and kernel feature spaces. CMU-PIE, YaleB and FRGC V2.0 databases are used for experiments. Experimental results show that the recognition performance, with three popular combination rules, can be increased by more than 10% on average using DI and RI.
KW - Discriminability index
KW - Face recognition
KW - Image assessment
KW - Multiple images
KW - Reliability index
UR - http://www.scopus.com/inward/record.url?scp=77953649646&partnerID=8YFLogxK
U2 - 10.1016/j.patcog.2010.05.024
DO - 10.1016/j.patcog.2010.05.024
M3 - Journal article
AN - SCOPUS:77953649646
SN - 0031-3203
VL - 43
SP - 3483
EP - 3493
JO - Pattern Recognition
JF - Pattern Recognition
IS - 10
ER -