Biometric technologies, including fingerprint, face, iris, and palmprint recognition, are drawing increasing attention in both academic research and industrial applications.1 Palmprint identification uses the inner surface of the palm to distinguish between people2 and includes among its many merits stable features, a high level of accuracy, and user-friendliness. To the best of our knowledge, all of the current palmprint recognition methods are 2D: they employ a digital camera to capture a 2D image and then extract features such as principle lines and wrinkles. However, a 2D palmprint could be relatively easy to forge, since it only shows the texture information of the palm surface.
Recently, 3D techniques have been applied to biometric authentication, for example, in face3 and ear recognition.4 The human palm, of course, is also a 3D object and offers plenty of data that might be used for authentication. Moreover, unlike the face and ear, features of a palm such as principle lines and wrinkles are especially resistant to counterfeiting owing to variations in their depth. By the same token, they are harder to acquire in 3D mode. The most popular approach to generating the desired depth information is laser triangulation.3,4 But this method, too, falls short. To achieve more accurate and higher acquisition speed than with ordinary laser-scanning technologies, we have developed a 3D acquisition device based on a structured light method5 that uses a cost-effective gray liquid crystal display (LCD) projector with a light-emitting diode (LED) light source to project shift-structured light patterns (stripes) onto the palm surface. The reflected light is captured by a charge coupled device (CCD) camera, and a series of images are then collected that can be used for calculating the palm depth information. Figure 1 shows a 3D palmprint obtained using our device.
Figure 1. A 3D palmprint sample.
In our approach, we first obtain a 3D palmprint from which we extract the region of interest (ROI) to eliminate noisy data on the boundary and reduce computing time. Collecting the 3D information also produces a 2D palmprint. We apply 2D preprocessing2 to this palmprint to get a 2D ROI, and then map the result to 3D data to generate a 3D ROI. Figure 2 shows both the 2D and the 3D ROIs that we extracted from the palmprint.
Table 1. Surface-type labels from curvature signs.
Once the ROI has been obtained from the 3D palmprint data, the task is to find and extract stable and distinguishing 3D features from the ROI and to match them. The most common 3D biometric matching method, iterative closest point (ICP),6 is not suitable for 3D palmprint identification because the slight deformation of the palm introduces a great deal of noise. After many experiments, we have found that the local curvature of the 3D palmprint is, relatively speaking, very stable and contains most of the 3D information needed. With this in mind we have devised an effective algorithm7 that calculates the mean curvature (MC) and Gaussian curvature (GC). The MC and GC produce curvature signs that can be used as reference points for eight surface types (STs), listed in Table 1.7 Comparing and matching the MC, GC, and ST produces high recognition performance when employing the 3D palmprint approach. Using our 3D palmprint database, which contains 6000 samples from 260 individuals, to do matching experiments, the equal error rate (EER) in verification tasks is 0.63%, and the rank-one correct rate of identification is 99.38%.
Figure 2. Region of interest (ROI) extracted from a palmprint. (a) 2D ROI. (b) 3D ROI.
3D biometric identification presents a challenging yet promising direction for future research. Fortunately, mathematical theories such as differential geometry, topology, and tensor analysis provide an excellent foundation for success because they support the transformation of many surface geometric properties into useful biometric features. In the future, we will develop more valid and more general feature extraction and matching algorithms for 3D biometric recognition as well as continue to apply the strategy of fusing 3D and 2D information to improve the performance of the technology.
David Zhang, Nan Luo
The Hong Kong Polytechnic University
Hong Kong, China
David Zhang is Chair Professor at the Hong Kong Polytechnic University where he is the founding director of the Biometrics Technology Centre. He also serves as a visiting Chair Professor at Tsinghua University, and adjunct professor in Shanghai Jiao Tong University, Harbin Institute of Technology, and the University of Waterloo. His current research interests include biometrics authentication and medical diagnosis in traditional Chinese medicine.
Shenzhen Graduate School
Harbin Institute of Technology
Shanghai Jiao Tong University
3. I. A. Kakadiaris, G. Passalis, G. Toderici, M. N. Murtuza, Y. L. Lu, N. Karampatziakis, T. Theoharis, Three-dimensional face recognition in the presence of facial expressions: an annotated deformable model approach, IEEE Trans. Pattern Anal. Machine Intell. 29, no. 4, pp. 640-649, 2007.