Score Fusion and Decision Fusion for the Performance Improvement of Face Recognition.

摘要

To improve the performance of a face recognition system, we propose a fusion solution consisting of score fusion of multispectral images and decision fusion of stereo images. Score fusion combines several scores from multiple matchers and/or multiple modalities, which can increase the accuracy of face recognition. In a face recognition system, low false accept rate (FAR) is as important as high accuracy rate. The FAR can be reduced by using decision fusion of stereo images. The stereo face images are taken with two identical cameras aiming at a subject, where each camera is built in two spectral bands, visible and thermal. Specifically, the score fusion combines the face scores from three matchers (Circular Gaussian Filter, Face Pattern Byte, Linear Discriminant Analysis) and from two-spectral bands (visible and thermal). The decision fusion combines the score-fusion results (genuine or impostor) from left faces and right faces in stereo imaging. We present three score-fusion results using k-Nearest Neighbor fusion, binomial logistic regression, and Hidden Markov Model fusion, meanwhile two decision-fusion results using logical AND and OR. Our experiments are conducted with the Alcorn State Univ. MultiSpectral Stereo face dataset that currently consists of the stereo face images of two spectral bands from 105 subjects. The experimental results show that score fusion can significantly improve the accuracy, whereas decision fusion (with AND rule) can reduce the FAR with a slight decrease in accuracy.