Feature selection for support vector machine-based face-iris multimodal biometric system

Abstract

Multimodal biometric can overcome the limitation possessed by single biometric trait and give better classification accuracy. This paper proposes face-iris multimodal biometric system based on fusion at matching score level using support vector machine (SVM). The performances of face and iris recognition can be enhanced using a proposed feature selection method to select an optimal subset of features. Besides, a simple computation speed-up method is proposed for SVM. The results show that the proposed feature selection method is able improve the classification accuracy in terms of total error rate. The support vector machine-based fusion method also gave very promising results.

Introduction

Biometric system based on single biometric trait (unimodal system) suffers from limitation such as lack of uniqueness, non-universality and noisy data. For example, the performance of face recognition is easily degraded by variations in term of pose, illumination and expression. Although different approaches (Belhumeur & Kriegman, 1996; Belhumeur and Kriegman, 1996, Georghiades et al., 2001) with high computation complexity have been proposed to counter the challenges, but face recognition under unconstrained condition still does not provide satisfactory result. In the case of voice recognition, the performance is easily affected by noisy environment. Multimodal biometric system operates based on the information extracted from multiple biometric traits. Generally, multimodal biometric can overcome the limitations possessed by single biometric trait and give better classification accuracy (Ross and Jain, 2003, Jain and Ross, 2004). For example, face-voice multimodal biometric system can complement each others because voice recognition is not affected by the varying lighting conditions while face recognition is immune to noise. Hence, the reliable performance of the system is guaranteed for real-world application.

The key to the success of multimodal biometric system is information fusion. In the case of biometric system, fusion of information can be done at four different levels: sensor level, feature level, matching score level and decision level. Fusion at different levels has advantages and disadvantages. The most popular choice is fusion at matching score level due to the ease in accessing and combining the scores generated by different matchers (Ross and Jain, 2003, Jain et al., 2005). For matching score level fusion, it can be further divided into three major categories: transformation-based score fusion, density-based score fusion and classifier-based score fusion (Nandakumar, Yi, Dass, & Jain, 2008).

For transformation-based score fusion, the matching scores have to be normalized to transform these scores into a common domain because the feature set from different modalities may be incompatible. and then combined. Jain et al. (2005) have studied different types of normalization technique on multimodal biometric system based on face, fingerprint and hand-geometry. Their study showed that z-score, min–max and tanh gave very good result. tanh method gave the best result but it involved a lot of parameters. Z-score and min–max are simple but they are insensitive to the presence of outliers. Alsaade, Ariyaeeinia, Malegaonkar, Pawlewski, and Pillay (2008) introduced unconstrained cohort normalization (UCN) into the score-level fusion process for face and voice verification. According to the author, if the system only considers a single genuine user, the imposter may easily match one of genuine users. Instead of taking only one of the genuine users, the method considers multiple top K closest images. Hence, the system is more robust and has lower EER rate. Besides that, the normalization method can also be used for face and voice recognition. In terms of combined method, method such as sum-rule, weighted sum rule, mean, product and etc. have been studied (Jain et al., 2005). Results showed that weighted sum-rule gives the best outcome under most of the situations.

Density-based score fusion is based on the likelihood ratio test. It requires explicit estimation of the genuine and imposter similarity score densities. Nandakumar et al. (2008) modeled the genuine and imposter similarity scores distribution as finite Gaussian mixture model. Their method gives very good result in NIST, XM2VTS and WVU database. Classifier-based score fusion method treats the matching score from each biometric trait as a feature vectors. Each matching score is considered as an element of the feature vector. The optimal classifier is build based on the training set that is available. Wang, Tan, and Jain (2003) combined the matching scores of face and iris recognition as a two dimensional feature vector. Linear discriminant analysis (LDA) and neural network based on radial basis function (NNRBF) are employed as the classifiers. Results showed that NNRBF outperforms LDA and weighted sum-rule method. Chen and Chu (2006) employed wavelet probabilistic classifier for face-iris multimodal biometric system. Recently, support vector machine (SVM) has been a popular technique for classification motivated by the results of statistics learning theory. Unlike traditional methods such as neural network that minimize the empirical training error, SVM is designed based on structural minimization principal. SVM maps the training data into a higher dimensional feature using kernel trick, in which an optimal hyperplane with large separating margin between two classes of the labeled data is constructed. Gutschoven and Verlinde (2000) first proposed multimodal biometric system based on SVM. Wang and Han (2009) employed SVM which incorporated radial basis function as kernel for face-iris biometric system. Their result showed that SVM-based score level fusion method outperforms LDA and weighted sum-rule.

Face recognition is the most unobtrusive biometric system. The recognition device does not require high degree of cooperation from the user. However, the face recognition has moderate classification accuracy and cannot be operate at high security area. In contrast, iris recognition is reliable and provides high classification accuracy (Daugman, 1993, Daugman, 2004). Face and iris have many similar characters, for example, they have similar image acquisition device, and they are both non-invasive and are relatively friendly. With the existence of high resolution camera, the iris information that is contained in face image is enough to perform recognition process. Several face-iris biometric systems (Chen and Chu, 2006, Wang and Han, 2009, Wang et al., 2003) have been proposed and have showed promising results.

In this paper, we propose a simple and fast face-iris multi biometrics system based on matching score level fusion. In real-world application, the feature set is generally large in terms of dimensionality. The features may be noisy and may contain irrelevant or redundant information about the target concept. This may cause performance degradation on classifiers. Besides that, large feature set also increases the storage cost and requires more computation time to process it. Feature selection is crucial to select an “optimized” subset of features from the original feature set based on certain objective function. In general, feature selection removes redundant or irrelevant data while retaining classification accuracy. The feature selection is performed to remove the redundant and irrelevant information that cause performance degradation of face recognition. For the case of iris recognition, the iris pattern is always occluded by eyelash and eyelid. The conventional iris recognition method generates a mask in order to mask out corrupted regions within the iris. However, locating the eyelashes and eyelids is a complex process and has high computation load. Furthermore, the global information that is contained in each track of a 1D iris track maybe corrupted due to eyelids and eyelashes removing process which breaks the 1D iris track to multiple small regions. The feature selection is performed to select tracks that most likely contain the discriminant information without detecting the eyelid and eyelashes. The matching scores are integrated to become a 2D feature vector and SVM is employed as classifier. Besides that, a simple method is proposed to select a fraction of the training data to train the SVM classifier to reduce the training time. ORL (ORL face database, http://www.cl.cam.ac.uk/research/dtg/attarchive/facedatabase.html) and CASIA (CASIA, http://www.cbsr.ia.ac.cn/IrisDatabase.htm) database are chosen to construct the multimodal biometric database to evaluate the performance.