Jointly learning multi-instance hand-based biometric descriptor

Abstract

Multibiometric recognition has become one of the most important solutions for enhancing overall personal recognition performance due to several inherent limitations of unimodal biometrics, such as nonuniversality and unacceptable reliability. However, most existing multibiometrics fuse completely different biometric traits based on addition schemes, which usually require several sensors and make the final feature sets large. In this paper, we propose a joint multi-instance hand-based biometric feature learning method for biometric recognition. Specifically, we first exploit the important direction data from multi-instance biometric images. Then, we simultaneously learn the discriminative features of multi-instance biometric traits and exploit the collaborative representations of multi-instance biometric features such that the final joint multi-instance feature descriptor is compact. Moreover, the importance weights of different biometric instances can be adaptively learned. Experimental results on the baseline multi-instance finger-knuckle-print and palmprint databases demonstrate the promising effectiveness of the proposed method.

Introduction

Biometric recognition has become one of the most important and effective personal authentication technologies due to its superior user acceptance over conventional token- and password-based authentication mechanisms [33], [43]. However, unimodal biometrics, which is based on a single modality, suffer from several inherent problems, such as nonuniversality, intraclass variations and unacceptable reliability [13], [20]. To address these issues, multibiometrics, which use multiple biometric traits to improve recognition performance, have drawn increasing research attention [26].

To date, there have been numerous multibiometric methods that fuse different kinds of biometric traits based on various strategies, such as sensor-, feature-, score- and decision-level fusions [26]. For example, Raghavendra et al. [31] fused visible and near infrared face images at the sensor level to improve the face verification performance of their model. Jing et al. [17] fused multiple modalities of biometrics, such as faces and palmprints, for personal recognition. Wang et al. [36] first captured palmprint and palm-vein images by using general color and near-infrared cameras and then fused them at the feature level. Nigam et al. [29] proposed a hand-based biometric system by fusing finger-knuckle-print (FKP) and palmprint traits at the matching score level. Monwar et al. [27] first extracted face, ear and signature features and then fused them at the rank-based decision level. Yang et al. [41] proposed a multimodal personal identification by fusing fingerprints and finger-veins at the feature level. Zhang et al. [44] and Li et al. [21] proposed a finger-based trimodal biometric recognition method by extracting and fusing fingerprint, FKP and finger-vein features. In addition, Gupta et al. [11] proposed a palm-dorsa vein-based multimodal authentication method by fusing multiple features extracted by different algorithms. Additional multibiometric fusion strategies and applications can be found in a conducted multibiometric survey [26]. It is noted that most existing multibiometrics usually fuse completely different biometric traits, such as by combining faces and fingerprints and combining palmprints and palm-veins. These multibiometrics usually require multiple sensors and different feature extraction methods, making them difficult to use in practical applications. In addition, most fusion strategies, such as feature-, score- and decision-level fusions, need to extract features from individual biometric traits and then fuse them by an addition scheme, making the dimensionality of the final multibiometric feature large. Moreover, the correlation of multiple biometric traits cannot be effectively exploited [26] by these fusion-based multibiometric technologies. How to effectively exploit and combine multibiometric features for biometric recognition remains a central and challenging problem.

In this paper, we propose a new multibiometric recognition method by combining multiple similar hand-based biometric traits. Specifically, we jointly learn the collaborative feature representations of multi-instance hand-based biometric traits, such as the multiple FKPs of the same hand and the left and right palmprints of the same subject. We first form direction data vectors to specifically sample the important direction information in hand-based biometric images. Then, we jointly learn a multi-instance feature projection function that can convert the multi-instance biometric data vectors into compact binary feature codes, where the weights of different biometric instances are also adaptively learned. Finally, we form a histogram representation of the compact binary codes as the multi-instance biometric feature descriptor. Fig. 1 shows the basic idea of the proposed method. We conduct comparison experiments on the widely used multi-instance FKP and palmprint databases with state-of-the-art methods to demonstrate the effectiveness of the proposed method.

The main contributions of this paper can be summarized as follows.

• We propose a new multi-instance hand-based biometric feature descriptor that jointly learns multi-instance biometric discriminative features and their common collaborative representations.

• Unlike in most existing addition-based multibiometric fusions, we exploit the collaborative representations of multi-instance biometric traits, so that the final learned feature descriptor is compact. In particular, our method automatically learns the weight settings of multiple instances of biometric traits.

• We conduct extensive experiments on both the FKP and palmprint databases, and the experimental results show that the proposed method outperforms the state-of-the-art feature descriptors on multi-instance biometric recognition tasks. The simple yet efficient obtained multibiometric representation demonstrates the potential of the proposed method to serve as a baseline for multi-instance biometric recognition.

The rest of this paper is organized as follows. Section 2 briefly reviews related work. Section 3 proposes the joint multi-instance hand-based biometric feature learning method. Section 4 provides the experimental results, and Section 5 offers the conclusion of the paper.