Human Identity Verification Using Multispectral Palmprint Fusion ()
1. Introduction
There exists a large number of computational approaches in intra-modal fusion [1,2] at different levels of human recognition. However, there are some incapable constraints in mono-modal biometric as well as intra-modal biometric systems, such as lack of accurate image registration methods [2], template matching with loss of complementary information [2], and association of redundant adaptive parameters [2]. These factors make the poor performance of the system. Intra-modal biometric image fusion can remove some of the limitations of uni-biometric systems [3] because the uni-modal biometric system usually compensates for the inherent limitations of the secondary sources. Intra-modal systems have the following advantages [4-6] over uni-modal biometric systems.
• Fusion of the evidence obtained in different form from the same or different sources can significantly improve the overall accuracy of the biometric system.
• Intra-modal biometric can address the problem of nonuniversality which often occurs in uni-modal system.
• Intra-modal systems can provide certain degrees of flexibility.
• The availability of multiple sources of information can reduce the redundancy in uni-modal system.
Biometric image fusion at sensor level/low level refers to a process that fuses multispectral biometric images captured by identical or different biometric sensors. This fusion produces an image in spatially enhanced form which contains richer, intrinsic and complementary information. Biometric verification systems seek considerable amount of improvement with respect to their reliability and accuracy.
Automatic authentication of users by their respective characteristics plays an important role in security. A biometric system recognizes the identity of a person with certain physiological/behavioral characteristics, such as fingerprints, face, iris, speech, hand geometry, etc. Biometric systems based on palmprint have been proposed in [7-13]. The palmprint recognition system has many advantages over other biometric systems in respect of reliability, low cost and user friendly. Palmprint is one of the most reliable means in personal identification because of its stability user friendliness, acceptability and uniqueness [7,12,13].
Palmprint image consists of wrinkles and creases along with three principal lines, namely, heart line, headline and life line. These lines vary little over time while wrinkles are much thinner than the principal lines and much more irregular. Creases which are detailed textures, like the ridges in a fingerprint, found all over the palmprint, can only be captured using high-resolution cameras. With the low-resolution palmprint image, the principal lines and thick wrinkles can be used for recognition.
Variations in different palmprint images of an individual can be combined to produce a fused palm image. Images acquired through different imaging sensors have been fused using various techniques discussed in [4-6]. The necessity for fusion techniques is increased with the inception of new image acquisition devices. By fusing images, it is possible to discern the useful information from the input images. However, biometric fusion using multiple palm images [5,6] at low level is expected to produce more accurate results than the systems that integrate information at later stages, namely, feature level, score level, etc. [2]. This is because of the availability of more relevant and precise raw information. Apart from integrating the contributive features to other levels of fusion, an image fusion scheme of a higher abstraction suppresses inconsistencies, artifacts and noise in the fused images.
Another problem often occurred in biometric applications is the selection of a set of features [14]. Feature selection is the most important step that can affect the performance of recognition system. It is often necessary to select that set of features which reflects the relevancy among the features.
Review of Some State-of-the-Art Systems
There exist many palmprint authentication systems which exhibit encouraging results. But there is a need to improve the performance of the existing systems. To cope up with spoof attacks and to make it tamper proof, multispectral algorithms can be used to high security zones where vulnerability often happens. In the recent years, a few multispectral palmprint systems have been developed for reliable means of authentication. In this section we briefly describe some well known multispectral palmprint authentication systems.
A multispectral palmprint recognition system using wavelet based image fusion has been proposed in [15]. It uses a multispectral capture device to sense the palm images under different illumination conditions, including red, blue, green and infrared. Further wavelet transform is used for combining the palmprint images obtained from different channels. During image acquisition the situation of hand movement is also considered. Finally, competitive coding scheme has been adopted for matching. It uses Wavelet based image fusion as data-level. Again this system has been further extended in [16] where features extraction and matching have been made of red, green, blue and NIR bands of a multispectral palm image. Finally these matching scores obtained from matching against different bands are fused using simple sum rule.
A contact-free palmprint verification system has been presented in [5] using multispectral palm image by means of feature level registration and pixel level fusion strategies. Initially a sequence of multispectral hand images is obtained by illuminating the hand with multiple active lights. Coarse localization of ROIs is performed through preprocessing on each image and it is then further refined through feature level registration. Finally, authors integrate the multiple image sources and the fusion is performed with multi-scale decomposition, activity measure and coefficient combining methods.
Feature band selection based multispectral palmprint recognition has been proposed in [17] where the statistical features are extracted to compare each single band. Score level fusion is performed to determine the best combination from all candidates. The most discriminative information of palmprint images can be obtained from two special bands. Region of Interest (ROI) is determined from hyper-spectral palm cube using local coordinate system.
In [18], multispectral palmprint recognition has been presented where multiple information related to hand are used. Hand shape, fingerprints and palmprint modalities are used for recognition. This system shows good recognition accuracy on a medium size database while fusion is performed with multiple fingers and fusion of finger and palm.
A comparative study of several multispectral palm image fusion techniques has been presented in [6] and some well-studied criteria are used as objective fusion quality measure. However, the curvelet transform is found to be the best among others in preserving discriminative patterns from multispectral palm images.
This paper presents a novel palmprint verification method in which palm images are fused at low level by wavelet transform [4] and fused palm is then represented by Gabor wavelet transform [8-10] to capture the palm characteristics in terms of neighborhood pixel intensity changes. Gabor palm responses contain high dimensionality features and due to this high dimensionality ant colony optimization (ACO) [19] is applied to select the optimal set of distinct features. Finally, support vector machines (SVMs) [20] are used to train the reduced feature sets of different individuals and verify the identity. Proposed palmprint system is evaluated with CASIA palmprint database [5,6] and the results are also compared with other existing methods to measure the effectiveness and robustness of the system.
The paper is organized as follows. Section 2 presents some preliminaries used for the proposed system. Section 3 briefly describes the proposed model and waveletbased palm image fusion scheme. Gabor wavelet representation of fused palm image is discussed in the next section. Feature selection using ant colony optimization algorithm is presented in Section 5. Classification method is discussed in Section 6. Experimental results are analyzed in the next section. Finally, conclusions are given in the last section.
2. Preliminaries
2.1. Region of Interest (ROI) Detection from Palm Image
Major issues for the degradation of a palmprint recognition system are accurate registration, palm feature representation and redundancy exploitation. Method of ROI detection [8] is employed to reduce the error caused due to translation and rotation. This process roughly aligns the palmprint and it does not reduce the effect of palmprint distortion.
To extract the ROI of palm image, it is necessary to define a coordinate system based on which different palm images are aligned for matching and verification. Gaps between fingers have been used in [8] as reference points for determining the coordinate system. This paper also applies this technique to determine the ROI of the multispectral palm image. The following algorithm is followed to extract the central part of the palmprint image as ROI and further this ROI is used for multispectral fusion of palm images.
• Step 1: Convert the multispectral palm image to a binary image. Gaussian smoothing can be used to enhance the image.
• Step 2: Apply boundary-tracking algorithm to obtain the boundaries of the gaps between the fingers. Since the ring and the middle fingers are not useful for processing. Therefore, boundary of the gap between these two fingers is not extracted.
• Step 3: Determine palmprint coordinate system by computing the tangent of the two gaps with any two points on these gaps. The y-axis is considered as the line which joining these two points. To determine the origin of the coordinate system, midpoint of these two points are taken through which a line is passing and the line is perpendicular to the y-axis.
• Step 4: Finally, extract ROI for feature extraction which is the central part of the palmprint.
Figure 1 illustrates ROI of palm image which is cropped from palmprint image. In practice, it has been seen that principal lines do not contribute adequately to high accuracy because of their similarity amongst different palms. Although wrinkles play an important role in palmprint authentications it is still a difficult task to extract them accurately. This problem motivates to apply texture analysis to palmprint authentication. There exist many texture based palmprint verification schemes including Gabor filtering, wavelet, etc. [8-11].