12-11-2012, 04:32 PM
Score Level Fusion of Fingerprint and Finger Vein Recognition
Score Level Fusio.pdf (Size: 392.93 KB / Downloads: 76)
Abstract
Unimodal biometric recognition is not able to meet the performance requirements in most cases with its application
becomes more and more broadly. Recognition based on multimodal biometrics represents an emerging trend recently. In
the paper, we propose multimodal biometrics recognition based on score level fusion of fingerprint and finger vein, since
fingerprint recognition and finger vein recognition are complementary in several aspects. Experimental results based on
homologous biometrics database demonstrate that the fusion of fingerprint and finger vein leads to a dramatically
improvement in performance.
Introduction
Biometric systems make use of physiological and/or behavioral traits of individuals, for recognition
purposes [1]. The physiological traits include fingerprint, face, finger vein, iris, hand geometry, palm print,
retina, etc, and the behavioral traits include gait, voice, signature, etc. Compared with the traditional
authentication such as key, password, IC card, biometric systems have higher scientific value and wider
application prospects, which have advantages in the following aspects: will not be forgotten to carry, will
not be lost and will be more secure.
Unimodal biometric systems which use single trait for recognition are often affected by various practical
problems such as noisy sensor data, unacceptable error rate, spoof attacks and non-universality [2]. And the
application enviroment has too many restrictions, for example, the deaf can not provide sound information,
the man who often engaged in manual work may not provide clear fingerprint texture. Hence each
biometric can not have a true sense of universality.
Fusion Levels of Multimodal Biometrics
In a multimodal biometric system based on multiple biometric traits, various levels of fusion are possible
[6]: fusion at feature level, matching score level, and decision level.
Feature Level Fusion
Feature level fusion means the fusion of feature vectors obtained from several feature sources. There are
serveral feature sources: (i) feature vectors obtained from different sensors based on a single biometric, (ii)
feature vectors obtained from different entities based on a single biometric, like iris feature vectors obtained
from left and right eyes, (iii) feature vectors obtained from multiple biometric traits.
Feature vectors which are compatible can be combined to form high-dimensional feature vectors when
several feature vectors belong to different types. For example, Ross and Govindarajan [15] fused face
features with hand geometry features, Kumar [16] proposed fusion of hand geometry features and
palmprint features, X. Zhou and B. Bhanu [17] proposed feature fusion of face and gait. They trained the
reconstructed faces with high resolution and gait energy images by PCA and MDA respectively in order to
get feature vectors, then combined them to a fusing vector after normalization. But it is difficult to
consolidate information at feature level because feature sets from different biometric modalities may
neither be accessible nor compatible [6].
Socre Level Fusion
Score level fusion is the fusion in matching score level. For this reason, it is also called matching level
fusion. Different matching scores got by different classifiers or from different biometrics can be fused to
match at this level. Fusion at matching level can be approached in two distinct ways [6]. One is viewed as a
classification problem; the other is viewed as an information combination problem. In the classification
approach, a feature vector is reconstructed using matching scores output by individual matchers. Then these
feature vectors are classified into “Accept” (genuine user) or “Reject” (impostor). In the information
combination approach, individual matching scores are fused to generate a single scalar score that is used to
make final decision. Note that the individual matching score should be normalized to an uniform field
before fusion.
X. Zhou and B. Bhanu [11] proposed fusion of side face and gait at matching score level. They got side
face features (EFSI) and gait features (GEI) from video, then fused at matching level using three strategies:
Sum, Product and Max. Ross et al. [18] proposed matching level fusion of fingerprint verification using
minutiae matching algorithm and texture matching algorithm.
Decision Level Fusion
Fusion at decision level refers that matching results outputted from a different set of matchers fuse to make
final decision. There are several fusion strategys at decision level fusion. For example, majority voting,
Bayesian inference, weighted voting based on Dempater-Shafer theory, AND or OR logical rules, etc.
In theory, the earlier information combines, the better results achieved. However, there are many
difficulties to achieve wonderful results from feature level fusion in practice. The main difficulties are: (i)
feature spaces of different biometric features are unknown in most cases, (ii) data sets at feature level may
be incompatible, (iii) connecting two feature vectors may lead to dimension redundancy as the substantial
increase of the fusion feature vectors. Consequently, only a few of fusion are studied at feature level.
Majority of fusion studies are at matching level or decision level. Table 1 shows comparision of the three
fusion levels. Fusion in this paper is at score level due to the ease in accessing and combining the scores
generated by fingerprint and finger vein.
Fingerprint Recognition
Fingerprint recognition mainly contains image preprocessing, feature extraction, matching. In order to
extract fingerprint minutiae accurately, we do a series of pre-process including calculation of orientation
field [19], enhancement [20], fingerprint thinning [21], etc. Fingerprint segmentation of this paper uses
variance method which is the simplest method, while the calculation of fingerprint orientaion field,
enhancement, and finger thinning use methods in the literatures.
Fingerprint matching based on point pattern can be approached in two stages. In the first stage, two
matching finerprints should be aligned by finding rotation and translation parameters of the two
fingerprints. In the second stage, aligned fingerprints will be matched. This paper used minutiae matching
algorithm based on ternary vector [22].
Dataset and Experiment Setting
A homogenous database including 2880 fingerprint and finger vein images collected from 80 fingers was
adopted in the experiments. Fingerprint images were acquired using an optical fingerprint scanner
developed by Zhongzheng Inc. Finger vein images were acquired using a device designed by Joint Lab for
Intelligent Computing and Intelligent Systems of Wuhan University. Examples of acquired fingerprint
images and finger vein images are shown in Figure2, and they are homogeneous.