Using Score Fusion for Improving the Performance of Multispectral Face Recognition
Score fusion combines several scores from multiple modalities and/or multiple matchers, which can increase the accuracy of face recognition meanwhile decrease false accept rate (FAR). Specifically, the face scores are generated from two-spectral bands (vi
- PDF / 874,534 Bytes
- 22 Pages / 439.37 x 666.142 pts Page_size
- 67 Downloads / 172 Views
Using Score Fusion for Improving the Performance of Multispectral Face Recognition Yufeng Zheng
Abstract Score fusion combines several scores from multiple modalities and/or multiple matchers, which can increase the accuracy of face recognition meanwhile decrease false accept rate (FAR). Specifically, the face scores are generated from two-spectral bands (visible and thermal) and from three matchers (circular Gaussian filter, face pattern byte, elastic bunch graphic matching). In this chapter, we first review the three face recognition algorithms (matchers), then present and compare the fusion performance of seven fusion methods: linear discriminant analysis (LDA), k-nearest neighbor (KNN), artificial neural network (ANN), support vector machine (SVM), binomial logistic regression (BLR), Gaussian mixture model (GMM), and hidden Markov model (HMM). Our experiments are conducted with the Alcon State University Multispectral face dataset that currently consists of two spectral images from 105 subjects. The experimental results show that all score fusions can improve the accuracy meanwhile reduce the FAR, and the KNN score fusion gives the best performance.
5.1 Introduction Human identification with biometrics has many applications, for examples, authentification of entering secured areas, authorization of using restrictive devices, and surveillance of crossing borders. In fact, biometrics can be implemented by companies, governments, border control to either verify a person’s identity for limiting or allowing access to a secured facility (like computer files, building area, border crossings), or to identify individuals (like suspects, criminals, terrorists) in order to record and track their information. Basically biometrics uses the digital signatures of physiological characteristics such as faces, fingerprints, irises, voice, palms to automatically Y. Zheng (B) Alcorn State University, Lorman, MS, USA e-mail: [email protected] J. Scharcanski et al. (eds.), Signal and Image Processing for Biometrics, Lecture Notes in Electrical Engineering 292, DOI: 10.1007/978-3-642-54080-6_5, © Springer-Verlag Berlin Heidelberg 2014
107
108
Y. Zheng
identify a person. Compared to face recognition, the recognition techniques using fingerprints and irises are well developed and widely applied. However, the data collection procedures for both fingerprint and iris are quite time consuming, and require closely contacting the imaging devices. These techniques cannot satisfy the needs in highly-populated sites such as airports or customs. Tight security-imposed applications (at customs or borders) also require a passive and contactless data acquisition, and a fast identification process. Face imaging is such a passive data collection that does not require much subject collaboration. The performance of face recognition can be improved utilizing multispectral imagery. Chang et al. [1] demonstrated image quality enhancement of fusing multispectral face images (e.g., visible and thermal) but did not report any recognition performance.
Data Loading...
