Bilkent University
Department of Computer Engineering


Multi-modal Biometric Fusion Using Soft Information From Correlation Output Planes


Mehmet Keskinoz

Assistant Professor of Telecommunications Engineering Program
Sabancı University

Advanced correlation filters have been shown to be successful in various biometric feature verification applications. Traditionally, correlation filters are designed to produce high peak-to-sidelobe-ratios (PSR) for authentic images at their output in order to have a good verification performance. In this work, we propose a soft fusion method using correlation plane outputs obtained from multiple biometric features rather than using PSR to improve the verification accuracy over the individual biometrics. The fusion of multiple correlation plane outputs is accomplished using a Support Vector Machine. Using the traditional method, we obtain the average equal error rate (EER) of 6.2% and 2.3% for the CMU PIE illumination face dataset and the NIST 24 plastic distortion fingerprint dataset respectively whereas fusion of face and fingerprint correlation planes using the proposed method reduces the EER to 0.7% showing that multi-modal biometric fusion can improve verification accuracy.

Mehmet KESKINOZ: Mehmet Keskinoz is an assistant Professor at Telecommunications Engineering Program of Sabancı University Istanbul, Turkey since 2001. He received his B.S. degree from Boğaziçi University Electrical and Electronics Engineering Department, İstanbul/Turkey. After that, he obtained his M.S. and Ph. D. degrees from Department of Electrical and Computer Engineering, Carnegie Mellon University (CMU), Pittsburgh/ PA, in 1997 and 2001, respectively. During his stay at CMU, he was a signal processing/communication research assistant in Data Storage Systems Center (DSSC). In the fields of communications, statistical data processing and pattern analysis, his research interests include robust receiver algorithms development for ultra wide band (UWB) PAN communications, Multi-band-OFDM based UWB, information fusion in wireless sensor networks, software radio, multiple input multiple output (MIMO) Wireless Channels, 1-D and 2-D inverse problems, hard/soft decision fusion, linear and non-linear correlation filters based biometric identification and pattern recognition , multi-model person verification, data modeling, parameter estimation and time series analysis, support vector machines (SVM), independent component analysis (ICA), sub-space methods, polynomial signal processing, turbo codes, low density parity check codes (LDPC), Bayesian networks, belief propagation approach to decode error control coding, multimedia signal processing.


DATE: December 17,2004, Friday @ 13:40