Modular Biometric System

Modular Biometric System a

Charles Hsu , Michael Viazankoa, Jimmy O'Looneya, Harold Szub a Trident Systems Inc., Fairfax Virginia 22030 b The George Washington University, Washington DC 20052 ABSTRACT Modularity Biometric System (MBS) is an approach to support AiTR of the cooperated and/or non-cooperated standoff biometric in an area persistent surveillance. Advanced active and passive EOIR and RF sensor suite is not considered here. Neither will we consider the ROC, PD vs. FAR, versus the standoff POT in this paper. Our goal is to catch the “most wanted (MW)” two dozens, separately furthermore ad hoc woman MW class from man MW class, given their archrivals sparse front face data basis, by means of various new instantaneous input called probing faces. We present an advanced algorithm: mini-Max classifier, a sparse sample realization of Cramer-Rao Fisher bound of the Maximum Likelihood classifier that minimize the dispersions among the same woman classes and maximize the separation among different man-woman classes, based on the simple feature space of MIT Petland eigen-faces. The original aspect consists of a modular structured design approach at the system-level with multi-level architectures, multiple computing paradigms, and adaptable/evolvable techniques to allow for achieving a scalable structure in terms of biometric algorithms, identification quality, sensors, database complexity, database integration, and component heterogenity. MBS consist of a number of biometric technologies including fingerprints, vein maps, voice and face recognitions with innovative DSP algorithm, and their hardware implementations such as using Field Programmable Gate arrays (FPGAs). Biometric technologies and the composed modularity biometric system are significant for governmental agencies, enterprises, banks and all other organizations to protect people or control access to critical resources. Keywords: Cooperative/Non-cooperative Standoff Biometrics, Fingerprints, Vein-Map Recognition, Voice Recognition, Face Recognition, and FPGA. 1.

INTRODUCTION

Following the events of 9/11, commercial biometric systems have become highly-demanded and developed methodology typically for the access security control. We will not address this matured area of biometrics. Rather, we consider the specific need of Combat Search and Rescue (C-SAR) that demands the standoff Biometrics to authenticate the downpilot, for example, and its physiology state as well as access the surrounding hostile environment, before the mission planning for C-SAR. The technologies might take advantage of physical physiology characteristics such as facial vein maps, eye, body, motion gait, or voice language pattern since the last decade; however, these existing biometric technologies usually require a cooperative individual under a controlled environment, and they must be extended to a distance having limited capabilities of identifying non-cooperative objects for surveillance applications, where the observed individuals are non-cooperative, not willing or cannot cooperate, and non-habituated. Our approach is to build an infrastructure platform and a combined classifier AiTR, like WiKi frame work, which allows each researcher to plug in their favorite successful sensory data, so obtained with different biometric modalities already in different scenario, and test on the ROC vs. POT and SNR parameters with the same uniform AiTR algorithm. To illustrate such a goal, we arbitrarily take Petland eigenfaces for the FBI Most Wanted list. In order to achieve desired performance, a biometric system with various modalities is presented here and it is integrated with the existing cooperative and non-cooperative biometric technologies to raise the awareness levels and provide a more secured public safety. This innovative model configures multi-biometric solutions to evaluate accuracy of distinguishing physiological or behavioral traits. Technically all biometric techniques are implemented using a sensor, to acquire raw biometric data from an individual; feature extraction, to process the acquired data to develop a feature-set that represents the biometric trait; pattern matching, to compare the extracted feature-set against stored templates residing in a database; and decision-making, whereby a user’s claimed identity is authenticated or rejected. Unique aspect of this approach include normalization of system-specific matching scores, allowing disparate outputs to be analyzed across a standard scale; availability of operator-configurable system logic, including combinatory and weighted, to enable detailed analysis of multi-system and multi-modal biometrics; highly efficient wireless mesh network; and operator-triggered data capture functions, allowing acquisition of low-quality data as might be present in field applications. This modularity biometric approach embedded with a number of biometric technologies: fingerprints, vein-mapping authentication, face recognition, and voice recognition is a full-scale biometric solution to completely support warfighters accessing and operating machines, Independent Component Analyses, Wavelets, Neural Networks, Biosystems, and Nanoengineering VII, edited by Harold H. Szu, F. Jack Agee, Proc. of SPIE Vol. 7343, 734317 · © 2009 SPIE · CCC code: 0277-786X/09/$18 doi: 10.1117/12.820297 Proc. of SPIE Vol. 7343 734317-1 Downloaded From: http://spiedigitallibrary.org/ on 07/08/2015 Terms of Use: http://spiedl.org/terms

equipment, materials, and communications devices. The potential benefits of this modularity approach can also be extended to both physical access to controlled areas and logical access to sensitive data. This paper is organized as follows. Section 2 introduces the top-level system design of the modularity biometric system. Section 3 describes the functions of biometric modules and their technologies. Section 4 provides a high-level hardware implementation using FPGAs.

2.

TOP-LEVEL SYSTEM DESIGN

The modularity biometric system consists of a number of cooperative and non-cooperative biometric technologies to collaboratively operate and extract identities' physical characteristics. The modularity biometric system is decomposed into several operative modules to be mapped to a specific hardware & software architecture in the implementation phase. The top-level architecture shown in Fig. 1 enables flexible feature extraction processing of biometric data capture, complex modularity biometric processing, and an intuitive user interface that provides access to all system resources. The modularity biometric processes include the vein-map authentication module, the fingerprint module, the voice password module, the face recognition module, the data compression module, wireless communication module, the system control design and the user interface design.

Voice Password Module Face Recognition Module Data Compression Module Wireless Communication Module

UserInterface Design

Fingerprint Module

System Control Design

Vein-Map Authentication Module

Fig. 1. Top-Level Design Architecture of Modularity Biometric System

Vein patterns are unique and difficult to recreate because they are inside the human body. The vein-map authentication module is to capture the vein patterns using a specific IR sensor, and extract the vein signature. The vein-map authentication module consists of a sensor device and a DSP with feature extraction capability. The sensor device uses infrared radiation source to extract an image of small blood vessel. A near-IR vein scanning device cannot penetrate very deep under skin, the device will recognize the superficial veins and rarely the deep veins. A "smart" algorithm is needed to preserve the connectivity of the vein features and extract vein structures. Fingerprint has been systemically used for individual identification since 1902, and the growing need and demand by US police officials for a national repository and clearinghouse for fingerprint records led to an Act of Congress on July 1, 1921, establishing the Identification Division of the FBI. The fingerprint module is to digitalize the minute ridge formations or patterns found on the fingertips to provide an effective biometric identification. There are three major operations in the fingerprint module: fingerprint acquisition, fingerprint edge extraction, and fingerprint recognition. In our study, an efficient edge extraction algorithm such as a Sobel-based extraction or extraction using a Mexican hat filter or Morlet filter is needed to effectively extract the fingerprint patterns. Voice recognition and authentication is fast becoming an integral part of security worldwide. Voice biometrics is being used in information security, access restriction, home and property security and in products marketed for home consumption. The architecture of a speech recognition engine can be broken down into a feature extraction module, a pattern matching algorithm, and a hypothesis block. The feature extraction block transforms the input speech into a set of spectral components, which will be compared to a number of known patterns. Then the “hypothesis engine” decides which utterance was recognized or informs the user of failure to recognize anything. Facial recognition is a form of computer vision that uses faces to attempt to identify a person or verify a person’s claimed identity. Facial image can be captured by digitally scanning an existing photograph or by using an electrooptical camera. Second, the facial information will be analyzed using the spatial geometry of distinguishing features of the face. Principle Components Analysis (PCA) is an effective solution to distinguish the facial features. The wireless communication module is to transport the high-quality captured and processed sensor data rapidly across a wireless network in real time. Mesh is a type of network architecture. A single node is connected to at least one other node in a mesh network, it will have full connectivity to the entire network because each mesh node forwards packets to

Proc. of SPIE Vol. 7343 734317-2 Downloaded From: http://spiedigitallibrary.org/ on 07/08/2015 Terms of Use: http://spiedl.org/terms

other nodes in the network as required. In order to enhance the communication efficiency, data compression technology will be applied to effectively transfer gigantic data over a wireless distributed network. The MBS can be a custom-built application to enables biometric features matching through multiple systems and modality that can reduce false rate, combat attempts to spoof biometric systems through non-live data sources, and strictly control the classified or sensitive data access. The strong security and auditing capabilities of biometrics are combined in this approach to collaboratively provide the exact biometric identification. The potential number of combinations in biometric technologies and the concomitant deployment options for military applications are considerable.

3.

BIOMETRIC TECHNOLOGIES

The term Biometrics is derived from the Greek words bio (life) and metric (to measure). Biometrics is basically a technology that measures and analyzes human physiological and behavioral characteristics for personal identification. The main reason for the acceptance of the biometrics as a tool for security is its universality, distinctiveness, permanence and collectability. Main issues to be considered when implementing a biometric system is performance, acceptability, and circumvention. Origin of biometrics dates back to the 14th century in China [1], which was reported by the explorer Joao de Barros. He wrote the Chinese merchants were stamping children's palm prints and footprints on paper with ink to distinguish the young children from one another. Technically, a biometric system typically consists of a sensor, a feature extraction unit, a matching unit and a decision making unit [2]. The functional block diagram is depicted in Figure 2. Trained Templates Sensor Device

Feature Extraction

Feature Map

Matching Process

Decision Process

Fig. 2. Functional Block Diagram of a biometric system

The sensor device is to collect the biometric data. It is the most important stage because the accuracy of the entire biometric system ultimately depends on the quality of data. The objective of feature extraction and feature map is to extract the significant features that can be used to identify a person. The matching process is to find the best match of the extracted features to the trained templates. If the best match is found with a specific distance threshold, the match is effective and the individual is identified. Otherwise, the identification will be denied and an appropriate will be executed in the decision process. A number of biometric methods have been used in the matching process to provide a better biometric feature extraction and identification. The comparison charts presented by National center for State Courts (NCSC) [1] is listed in Table 1. Table 1. Comparison charts of Biometric technologies presented by NCSC

3.1

Biometrics

Fingerprint

Iris

Face

Voice

Verification

Yes

Yes

Yes

Yes

Identification

Yes

Yes

No

No

Reliability

High

High

Medium

Low

Error Rate

1 in 500+

1 in 131,000

No data

1 in 50

User Acceptance

Medium

Medium

Medium

High

Intrusive

Somewhat

None

None

None

Ease of Use

High

Medium

Medium

High

Standards

Yes

Not yet

Not yet

Not yet

Fingerprint Technology [3,4]

Fingerprint patterns are divided into three main groups consisting of: Arches, Loops and Whorls shown in Fig.3. Approximately 5% of all fingerprints are Arches, 30% are Whorls and 65% are Loops. Fingerprint scan quality can affect the reliability of any electronic fingerprint system. A good fingerprint system should be able to avoid dry prints


scan and wet prints scan. A dry print can appear broken or incomplete to the electronic imaging system that can result in inferior model construction during a registration process or inconsistent matching during a look-up process. A wet print can cause line-type features in the print to blend together during the registration or look-up process resulting in inferior model constructs or inconsistent look-ups.

Fig.3. Three main fingerprint patterns: Arches, Loops, and Whorls. 3.1.1

A Fingerprint De-Smearing Wavelet Algorithm

Biometric-based personal identification techniques such as DNA and fingerprints are becoming popular compared to the traditional token-based or knowledge-based techniques such as identification card (ID), password and etc. One anticipates the availability of next generation combatant devices (CD) with augmented keypad fingerprint acquisition systems (KFAS), which can actively opto-electronically scan fingerprints without inks. The CD identifies legitimate users without the need for short half-life passwords. The fingerprint is acquired by touching the keypad or by an input from an attached surrogates KFAS. If a higher security concern is necessary, KFAS needs longer time measurement of live finger temperature and even blood pulse that may generate inadvertent smearing. Then, the dynamic de-smear wavelet algorithm overcomes the robustness problem associated with either fast fingers striking the smart keyboards or held motionless on a surrogate KFAS. Both may create a time-integration dynamic smear due to the vertical or lateral minutia motion either by means of a line-by-line scanning mechanism or a long exposure time. Another application could be that crime scene fingerprints (latent fingerprints). These are usually obtained under uncontrolled and uncooperative conditions when a person touches a material or fabric surface with natural perspiration on fingers, oils, blood, or dirt. Even a small movement while handling an item can cause the image deposited to be blurred and unusable. Fingerprints were obtained before the recent advent by inking a person's fingers and rolling them on a white card. They can also be captured using "live-scan" computer equipment. Using live-scan, a person's fingers are scanned one-by-one and digitally recorded by computer. Both fingerprint acquisition mechanisms require an effective device to avoid blurring and provide useful fingerprints. An interesting de-smearing algorithm by means of Quadruture Mirror Filter (QMF) Bank [5] can effectively measure the degradation and retrieve the spread-out information to recreate sharp-ridge fingerprints. The de-smearing QMF algorithm is fast and relatively insensitive to noise in the fingerprint image. While the algorithm does require an estimate of the point-spread function, we have found that even if the estimate is inaccurate useful results can still be obtained. Analysis Input F(z)

Synthesis

H0(z)

2

2

G0(z)

H1(z)

2

2

G1(z)

Output F^(z)

Fig. 4. DWT by using lossless QMF

The principle behind the Quadrature Mirror Filter (QMF) is to hierarchically decompose the input signals into a series of successively lower resolution reference signals and their associated detail signals. At each level, the reference signals and detail signals contain the information needed to be reconstructed back to the next higher resolution level. The lossless QMF is composed of the analysis and synthesis processes. We denote H0 as a low pass filter and H1 as a high pass filter in the “Analysis” process. In the “Synthesis” operations, G0 is denoted as a low pass filter and G1 as a high pass filter. An input signal F(z) is input to the analysis low pass filter H0(z) and the analysis high pass filter H1(z). The odd samples of the filtered outputs may be discarded, corresponding to decimation by a factor of two. For the synthesis process (reconstruction), interpolation by a factor of two is performed, followed by filtering using the low pass and high pass synthesis filter G0(z) and G1(z) as shown in Fig. 4. Constraints on filter design include perfect reconstruction (lossless in terms of image quality), finite-length (finite number of taps in the filter with no feedback), and regularity (the


filter convolved upon itself will converge) so that the iterated low pass filters may converge to continuous functions. To provide a lossless QMF design, we apply the discrete filter bank theory to the scaling H0 filter via the standard ztransform: H 0 ( z ) = ∑ ck z − k k =0

; H ( z) = d z −k ∑ k 1

(1); (2)

k =0

From the theory of filter banks, one should note that in order to eliminate aliasing, the following relations must be satisfied. n (3), (4) G1 ( z ) = ( −1) n+1 H 0 ( z ) ; H1 ( z ) = (−1) G0 ( z ) We may obtain equation (5). If equation (3) and (4) are satisfied, F(z) = F^(z). F ^ ( z) =

1 {F ( z ) H 0 ( z ) + F (− z ) H 0 (− z )}G0 + 1 {F ( z ) H1 ( z ) + F (− z ) H1 (− z )}G1 2 2

(5)

Discrete Wavelet Transform (DWT) can be described in terms of the filter bank that is related to Sub-band Coding & Quadrature Mirror Filter (QMF). The DWT and Inverse DWT with QMF can be applied to de-smear the fingerprint. After a pass through the bi-modal histogram check, the fingerprint input is assumed to have a limited smear as if it were passed through the lowpass (LP) filter (1, α, α, 1) of arbitrary strength α to be determined by the lossless reconstruction. The decomposition algorithm and the a-band filter, which was studied in optical engineering SPIE July 1994 [6], are used to compute the original imagery. The smeared fingerprint can be considered as the imagery after passing the low pass filter L, which is (1, α, α, 1) in our case. Given any value of α, the pseudo-original imagery (I) can be reconstructed through the de-convolution of the filter and smeared imagery (S). Then this pseudo-original imagery must be verified through 2-channel DWT with QMF. If the pseudo-original imagery is our original imagery (O), the ratio between the output imagery (O) and the input imagery (I) should be consistent. If the ratio is not consistent, then we should choose a different value of α, and repeat the above process again, until the consistent ratio is found. The desmearing fingerprint using the QMF de-smear algorithm is illustrated in Fig. 5. The image size is 512 x 512 pixels with 8-bit grayscale.

Fig. 5. A one-directional smeared fingerprint (left); De-smeared fingerprint using QMF algorithm (right)

The de-smearing algorithm can be applied to solve the one-directional smearing problem. The de-convolution of the lowpass filter and the smeared imagery can be implemented by the row operation of the smeared imagery. The above de-smearing algorithm can be applied into multi-directional imagery as well. If the smear is at 45o, the solution is to repeat the de-smearing algorithm twice, one in the x-coordinate and another in y-coordinate. 3.1.2

Compression Preserve the Discontinuity Singularity Map

FBI has been collecting fingerprint cards since 1924. Over the past 78 years, their collection has grown to over 200 million cards and continues to accumulate at a rate of 30,000 ~ 50,000 new cards per day. For instance, a fingerprint image is captured by FBI at 500 dpi, 1.54 by 1.54 inches, and has about 512 x 512 pixels with 8-bit dynamic range per pixel. A huge database (512 x 512 x 10 x 285,000,000 = 7.47 x 1014 bytes) needs to be established to store the entire catalog of US fingerprint images, assuming a population of 285,000,000. It is obvious that the mission is impossible without the development of the perception-lossless fingerprint compression. Wavelet Scalar Quantization (WSQ) is the fingerprint compression standard developed by FBI. The principle behind the wavelet transform is to hierarchically decompose the input signals into a series of successively lower resolution reference signals and their associated detail signals. The block diagram in Fig. 6 gives an overview of the main steps involved in WSQ encoding/decoding. Two compression ratios among several are shown in Table 2 [3]. Although the JPEG method has detected the same number of minutiae (see the first row in Table 1), it produces more misses (not shown) due to the 8 x 8 block process producing a


check board-like image. The Best Basis Wavelet (BBW) method suggested by the Yale group indicates (in the second row of the Table) the fact that one can detect minutiae better if more computer time is used to design the mother wavelet adaptively from the library using the statistical entropy criterion. WSQ Encoder: DWT

WSQ Decoder: Compressed bit stream

Scalar Quant.

Huffman Coder

TABLE

TABLE

Huffman Decoder

Scalar Dequant.

TABLE

TABLE

Compressed bit stream

IDWT

Fig. 6. Functional Block Diagram of FBI WSQ Fingerprint Compression Table 2. Minutiae detection for different compression methods and ratios Compression Ratio

15:1

Compression Methods

20:1

JPEG

WSQ

BBW

JPEG

WSQ

BBW

Correctly detected in original & compressed

72

72

69

72

72

69

Correctly detected only in compressed

6

10

9

10

11

11

Fig. 7. Compressed result using WSQ (left); Compressed result using WSQ and SM (middle); Original fingerprint (right)

In our development experience, we pay attention to the change of intensity instead of image pixels. Thus, pixel's dynamic range compression algorithms are “overkill” due to availability of supercomputers in solving image compression problems. According to Nobel Laureates Hubel and Wiesel, the HVS has intrinsically oriented edges in several octaves multiple resolutions. To capture these multiple resolution edges, we compute for convenience the topological index set, e.g. the derivative jumps at arbitrary order and label those gray scale edges together with the jump value. Such a level set of edge contour when is labeled with topological index may be called the singularity map (SM). Obviously, the SM can be reduced to a trivial case of a constant jump in the slope across a binary edge. How does the Human Visual System (HVS), without learning the topology calculus, extract the SM? The visual neurons fire when they receive the light. Since the firing neurons require to be replenished by resources, the neurons at the light side boundary have more resources to use because the dark side neighborhood neurons do not need them. Also, once they are fired, they use up temporarily the resource and thus inhibit nearby neurons. Therefore, the dark area becomes darker, and the light area becomes lighter. The transition from dark to light appears enhanced. This is called the lateral inhibition mechanism (LIM) in biology that creates or sharpens the discontinuity in intensity slopes called the singularity in mathematics. The mathematical construction of the SM of the HVS is to reduce the redundancy of spatial and temporal data for image information. In digital image processing, the change of the intensity the so-called singular-edge is essential to our image perception. Transmission of the change of the singular map can optimize the usage of


communication bandwidth. The SM of the HVS is composed fundamentally of the general vision wavelet compression operation and the fine detail edge vision operation. The broad vision operation simulates the biological function of rods to process less significant and low-resolution imagery. Generally speaking, the wavelet compression can be used to provide the general lossy imagery without DCT block artifact. In order to reduce the transmission of the coordinate positions of the singular map, the wavelet compressed image is still sharp enough to derive the coordinate position information directly from the lossy imagery. The experimental simulation is demonstrated in Fig. 7. The left of Fig 7 shows the compression result using WSQ whereas the result of the WSQ compression with SM is illustrated in the middle of Fig. 7. It can be observed that the compressed result with WSQ and SM provides more detailed edge information than the WSQ only. The original image is shown on the right of Fig.7. 3.2

Facial Recognition

Face recognition is an amicable alternative because the authentication can be completed in a hands-free way without stopping user activities. Technically, facial recognition records the spatial geometry of distinguishing features of the face. Different vendors use different methods of facial recognition, however, all focus on measures of key features of the face. Because a person’s face can be captured by a camera from some distance away, facial recognition has a clandestine or covert capability (i.e. the subject does not necessarily know he has been observed). For this reason, facial recognition has been used in projects to identify card counters or other undesirables in casinos, shoplifters in stores, criminals and terrorists in urban areas. The face is treated as a two-dimensional pattern of intensity variation. Principal Component Analysis (PCA) has been proven to be an effective face-based approach. PCA also known as the Karhunen-Loeve transform has proven to be an exceedingly useful tool for dimensionality reduction of multivariate data with many application areas in pattern recognition and appearance-based visual recognition. PCA aims to find the principle components (eigenvectors/eigenfeatures) of a given set of multi-modal sensor data and represent each feature in a lower dimensional space using eigenvectors. PCA is widely applied in the application area of face recognition. PCA is used to find a low dimensional representation of data. Some important details of PCA are highlighted as follows [7]. Let X = {X ,n =1,...,N}∈ Rdxd be an ensemble of vectors. In imaging applications, they are formed by row concatenation of the image data, with dxd being the product of the width and the height of an image. Let be the average vector in the ensemble as shown in equation (6). After subtracting the mean from each element of X, we get a modified ensemble of vectors depicted in equation (7). And then the auto-covariance matrix M for the ensemble X is defined in equation (8), where M is d2xd2 matrix, with elements described in equation (9). It is well known from matrix theory that the matrix M is positively definite (or semi-definite) and has only real non-negative eigenvalues [7]. The eigenvectors of the matrix M form an orthonormal basis for Rdxd . This basis is called the K-L basis [8]. Since the auto-covariance matrix for the K-L eigenvectors are diagonal, it follows that the coordinates of the vectors in the sample space X with respect to the K-L basis are un-correlated random variables. Let {Yn, n = 1,..., N} denote the eigenvectors and let K be the d2xd2 matrix whose columns are the vectors Y1,…,YN . The adjoint matrix of the matrix K, which maps the standard coordinates into K-L coordinates, is called the K-L transform. In many applications, the eigenvectors in K are sorted according to the eigenvalues in a descending order. In determining the dxd eigenvalues from M, we have to solve a d2xd2 matrix. Usually, d = 128 and therefore, we have to solve a 16x16 matrix to calculate the eigenvalues and eigenvectors. The computational and memory requirement of the computer systems are extremely high. From matrix theory that if the number of training images N is much less than the dimension of M, i.e. N < dxd, the computational complexity is reduced to O (N). Also, the dimension of the matrix M needed to be solved is also reduced to N x N. Details of the mathematical derivation can be found in [9]. Since then, the implementation of PCA for characterization of face becomes flexible. In most of the existing works, the number of training images is small and is about 200. However, computational complexity increases dramatically when the number of images in the database is large. The PCA of a vector y related to the ensemble X is obtained by projecting vector y onto the subspaces spanned by d’ eigenvectors corresponding to the top d’ eigenvalues of the autocorrelation matrix M in descending order, where d’ is smaller than d. 1 N (6) E( X ) = ∑ X n N n =1 (7) X = X n , n = 1,..., N ; X n = X n − E ( X )

{

}

M = cov( X ) = E ( X ⊗ X ) 1 M (i , j ) = ∑ X n (i ) X n ( j ),1 ≤ i, j ≤ d 2 N


(8) (9)

A computer simulation of PCA trained by 21 faces is displayed in Fig. 8. Only 5 eigen faces (features) are selected for the recognition operation which is shown in Fig. 9. Another tested image (one of the 21 people, but different image) is recognized using PCA technology and its result is depicted in Fig. 10. (The face image data can be found in http://cswww.essex.ac.uk/mv/allfaces/grimace.html.)

Fig. 8. 21 facial data used by PCA

Fig. 9. 5 eigenfaces generated using PCA.

3.3

Fig. 10. Tested face (left); recognized face (right)

Vein-Map Authentication:

Vein authentication is a new biometric technology using the vein features inside human's body such as palm or finger to verify personal identification. The vein-map authentication has been proved to effectively extract a feature vector from specific physiological or behavioral characteristic [10] and it provides many important biometric features: • Uniqueness and permanence of the vein pattern: The physical shape of the subcutaneous vascular tree of the back of the hand or finger contains information that is capable of authenticating the identity of an individual reported in [1114]. • Non-contact and non-invasive detection can promote high-level of user acceptance. • Anti-forge and anti-copy: Since veins are covered and protect under the skin, they are impossible to forge and copy. • The vein pattern is intricate enough to allow sufficient criteria for detecting various subject even identical twins. Technically, the vein-map authentication is composed of a sensor device followed by a signal processing device. The sensor device will take a snapshot of the subject's vein under a source of infrared radiation at a specific wavelength. Hitachi and Fujitsu have successfully developed the vein pattern authentication technology and systems for fingers and palm respectively. Fig. 11 shows the finger vein pattern authentication system with the finger vein pattern from Hitachi (left two), and the palm vein pattern authentication system with the palm vein pattern from Fujitsu (right two).

n

Fig.11 Finger Vein Pattern Authentication System (Hitachi: left two); Palm Vein Pattern Authentication System (Fujitsu: right two).

Once the image is captured using a IR sensor, a "smart" algorithm is needed to overcome some potential disconnectivity of vein map and effectively extract the feature map. The disconnectivity of the vein map is due to that the infrared radiation can't penetrate all kinds of tissues in the same manner. The SM technology described in section 3.1.2 can be


used to preserve the connectivity of the vein. The opportunities to implement the vein-map authentication span a wide range of vertical markets, including security, financial/banking, healthcare, commercial enterprises and educational facilities. Applications for the device include physical admission into secured areas; log-in to PCs or server systems; access to POS, ATMs or kiosks; positive ID control; and other industry-specific applications [14]. 3.4

Voice Password

Currently, voice recognition and authentication is not widely used in the safe industry; fingerprint analysis is the only biometric based security system available for home safes [15]. However, voice recognition and authentication is fast becoming an integral part of security worldwide. Voice biometrics is being used in information security, access restriction, home and property security and in products marketed for home consumption. And the voice and word recognition programs are projected to be 90-95% accurate in current technology development. The voice biometric solution market produced $80 million in licensing and application revenue in 2006 and that number is expected to grow to $800 million by 2011 [15]. Voice based security systems consists of voice recognition and voice verification. In voice recognition, voice samples are obtained and features are extracted from them and stored in a database. These samples are compared with various other stored ones and using methods of pattern recognition. As the number of speakers and features increase this method becomes more taxing on the computer, as the voice sample needs to be compared with all other samples stored. A highly efficient voice recognition algorithm is needed to find unique features for each user if the voice data base is gigantic. Voice verification is a relatively easy procedure wherein a user supplies the speaker’s identity and records his voice. The goal of speaker verification is to confirm the claimed identity of a subject by exploiting individual differences in their speech. The features extracted from the voice sample are matched against stored samples corresponding to the given user, therefore verifying the authenticity of the user. In most cases a password protection accompanies the speaker verification process for added security. Feature extraction with specific transformation can effectively extract the significant signatures that will reduce the computation complexity in the recognition and verification process. Discrete wavelet transforms and the continuous wavelet transforms (CWT) together with artificial neural networks (ANN) have been employed to achieve automatic pattern recognition [16-17]. In [17], an ANN method was developed to construct an optimum mother wavelet that can organize sensor input data in the multi-resolution format that seems to become essential for brain style computing.

4.

HARDWARE IMPLEMENTATION

For real-time biometric security processing as required by the aforementioned parallel processes, the hardware design and overall architecture will require acceleration in processing power; thus, algorithms can be implemented in embedded processors and/or Field Programmable Gate Arrays (FPGAs) to achieve these high recognition rates and high throughput requirements. FPGA-based embedded systems are emerging as an alternative to traditional microprocessors and DSP processors. Their lower clock speeds minimize power consumption and performance is maintained through application specific hardware implementation. Current FPGAs permit extreme parallel processing capabilities and contain hard IP core embedded processors for providing an integrated hardware/software solution as a system on a chip (SoC). For this paper, we propose a highly-parallel biometric system implementation using an FPGA. 4.1

FPGA–based Reconfigurable Computing

FPGAs provide a flexible architecture that comprises of an array of Configurable Logic Blocks (CLBs), high-speed DSP hardware multipliers, embedded memory, and other specialized programmable elements - all surrounded by programmable Input/Output Blocks (IOBs), all interconnected by a hierarchy of fast, versatile routing resources as illustrated in Fig. 12. The DSP hardware elements, which comprise the heart of most algorithms, support independent functions such as a 25 x 18-bit multiplier with 48-bit accumulator and cascaded I/O to neighboring elements. There are over 500 true hardware multipliers per typical FPGA and each can execute independently at clock rates over 500MHz adding the high-bandwidth, parallel execution of the multiple algorithms needed for complex modular biometric systems. FPGAs offer this unprecedented algorithm-specific capable hardware and pipelined parallel processing of the multiple biometric algorithms. 4.2

Hardware Overview

A simplified FPGA-based Modularity Biometric System is shown in Fig. 13. All of the algorithms, as well as the data compression, are parallel processing the system inputs to maximize throughput and computation capability.


uoii*r

I4rc*

...c. PROGRAMMABLE INTERCONNECT

I/O BLOCK

. II;ii

.q1-.

LL1 V ._j!1I_I

L

,11

;UI

-

S.ITMN

LOGIC BLOCKS

4

Fig. 12. Typical FPGA Architecture (l), DSP Hardware Element (multiplier) (r)

VPR ALGORITHM

CCD

DESMEARING ALGORITHM

SPEECH

MICROPHONE

MFCC/TESPAR ALGORITHM

EO SENSOR

K-L ALGORITHM

FACE

RF I/F

IR SENSOR FINGER

DATA COMPRESSION

FPGA

RF

Fig. 13. High-level FPGA Implementation and Interfaces

4.3

Algorithm mapping

There are numerous tools from commercial FPGA vendors that interface directly with Matlab and Simulink to aid in the implementation and synthesis of designs from algorithm concept and system modeling to hardware simulation. As an example of implementing voice recognition, the authentication refers to the process of accepting or rejecting the identify claim of a user on the basis of individual information present in a speech waveform. A microphone and a high quality audio encoder-decoder (CODEC) with a sample rate of 96 kHz would be used to acquire an audio clip and convert it into digital speech for input to the FPGA. An algorithm such at the Mel-frequency Cepstral Coefficients (MFCCs) can be incorporated into an FPGA or embedded processor by processing the audio clip as illustrated in Figure 14. AUDIO CLIP

MEL FILTER

FFT

DCT

LOG( )

MFCC

Fig 14. Feature extraction of the MFCC Algorithm

To further simplify the memory intensive requirement of the Cepstral processing, Time Encoded Signal Processing and Recognition (TESPAR) [18,19] can be used to provide efficient user verification while providing a better fit in an FPGA utilizing internal embedded memory, DSP elements, and Look-up Tables (LUTs). The TESPAR algorithm hardware implementation can be divided into a coding engine and biometric comparison engine as shown in Figure 15. COMPARISON ENGINE

CODING ENGINE

SPEECH

FILTER + ADC

DURATION SHAPE MAPPER

AUTHENTICATE S-MATRIX GENERATOR

PATTERN MATCHING

DECISION

MEMORY (STORED TEMPLATES)

Fig 15. TESPAR Functional Elements Hardware Implementation


REJECT

The other biometric algorithms for face, vein-map, and fingerprint recognition, while slightly more complex to implement, can also be implemented in the same FPGA in a similar fashion using embedded memory, multipliers, LUTs, etc. as detailed in [20] for the QMF algorithm and in [21] for the K-L transform. 4.4

Data Compression and Wireless Communications Interface

Data compression in FPGAs is widely used and numerous FPGA IP soft cores are available commercially for lossless compression. Additionally, custom compression techniques can be implemented using internal resources of the FPGA such as block RAM and LUTs. The output interface of the system is a wireless communication link and an FPGA is suited to handle digital data up/down-conversion to the external DAC/ADC and associated RF subsystem. 5. Aided Target Recognition Modularity Biometric System (MBS) is an approach to support AiTR of the cooperated and/or, not-able to, not willingly cooperated biometric in a standoff area persistent surveillance. Advanced active and passive EOIR and RF sensor suite is not considered here. Neither will we consider the ROC, PD vs. FAR, versus the standoff POT in this paper. Our goal is to catch two dozen FBI “most wanted (MW),” separately furthermore the woman MW class from the man MW class, given their archrivals sparse front face data basis, by means of various new instantaneous input called probing faces for different types of variability. The challenge of sparse probing snapshot is that man may have facial hair, while woman, make up, and both may wear sun glasses, hats, scurf in different projection poses. We consider a small data basis, say 20 FBI most wanted suspects, and further divide them into 8 woman cohort group and 12 man cohort group for different disguises shown in Fig.16. We present an advanced nearest neighbor classifier algorithm: mini-Max classifier. The min-Max optimizes that the dispersions minimization among the same classes while, at the same time, maximize the separation among different classes, based on the simple orthogonal feature space of MIT Petland eigen-faces. It is important to know the correlation of the eigen faces in these two groups. Table 2 lists the angles between these two groups. In addition, the eigen value is another useful to provide correlation of these two group. Fig.18 plots the eigen values of the man and woman groups, and the conceptual operation of the min-Max optimization is illustrated in Fig. 19. It is a sparse realization of the celebrated Cramer-Rao Fisher bound of the Maximum Likelihood classifier. Tested images will be classified into man or woman groups, and then be identified. Selected simulations are demonstrated in Fig.17.

Fig. 16. 20 faces divided into 8 woman cohort group and 12 man cohort group.

0)

4F

Fig. 17. Selected simulations show nearest neighbor classifier algorithm: mini-Max classifier.


M

Fig.18. Two eigen value plots for man and woman groups

Fig. 19. Conceptual operation of min-Max optimization.

Man Eigen Face

Table 2. Angles of the eigen faces between these two groups Woman Eigen Face 4 5

1

2

3

1

309.0473

293.2391

270.305

298.701

262.3371

6

7

8

271.0343

270.0523

270.5659

2

309.2589

290.0942

290.9434

285.1245

3

307.5122

294.5003

272.0702

296.9166

282.3966

270.245

281.6407

272.1887

289.3038

255.9143

301.0757

4

284.2863

260.2259

264.4086

277.6278

5

306.9482

306.7149

249.4829

280.2226

295.0594

279.9569

314.6803

296.0775

262.5122

299.7273

321.4564

239.4008

6

275.9339

267.4515

262.4012

268.4499

272.812

265.3167

284.7632

312.9154

7

284.8234

277.1925

8

298.0571

253.1648

252.3627

307.626

245.4826

244.4746

272.9838

284.9064

291.9032

286.6426

303.734

252.7563

314.541

236.0639

286.1546

9

304.5627

289.8365

307.6271

273.3001

239.0983

242.2918

230.9922

295.5474

10

280.5897

285.0167

278.1775

207.207

250.0299

270.9221

256.1562

273.6633

11

305.084

249.5918

293.6633

227.1094

255.0956

322.8796

358.8371

255.2204

12

276.4322

270.3637

276.5285

300.9718

274.782

263.2025

288.4101

99.5895

(1) We assume after a proper image taking, segmentation, projection extrapolation, and zero-mean normalization, we obtain woman and man archival front face image data bases of an identical size: A={Ai} and B={Bj}, etc. (2) We further denote the eight woman cohort covariance matrix which in principle shall be averaged over the time of image acquisition. [Aij] = t based on the set {Ai};namely Matlab/PCA produces the contracted square 8x8 matrix: [Aij]fi=λifi ; λ8>λ7>…>λ1 generates the eight eigenfaces f={fi } i=8,…1, inversely listed from the largest eigenvalue, or the degree of freedom, to the smallest on , Matlab conversion, so obtained from the 8 woman group. Similarly, we obtain 12 eigenfaces g={gj} from 12 man B={Bj}; etc.


(3) Since for a real and symmetric covariance matrix the eigen-function is real and orthogonal to one another. Thus, we do not need to make them orthogonal feature vectors, for they come in naturally. All we need to do is to select the best feature vector to represent ones class and separate from other. Thus, we construct direction cosine table among 8 woman eigenfaces versus 12 man eigenfaces. (4) We choose the orthogonal pairs of the table to represent the group class separation. It often turns out to be f8 and g12 (5) The nearest neighbor classifier is intuitively has two steps. (i) Step one: cohort group ID and (ii) Step two: membership ID. By inspection, a dot plotted in the eight dimensional hyperspace represents a person, which requires take the direction cosine inner product with respect to all eigen-basis/faces. Then, the cloud of those dots represent the data basis. Any probing image enters into the classifier must require an identical pre-processing, segmentation etc. By entering the classifier, it must be taking the inner product direction cosine computation, so that it becomes a likewise a across in the cloud, from which the membership ID follows by the nearest one distance. (6) In order to handle different disguises in different cohort group, we have separated the archival data basis likewise. Thus, we must stitch together two hyperspaces together at the common origin. The is done by the maximization of inter-class separation. We can now generalize this mini-max classifier beyond Petland covariance matrix eigen-faces representation.

Theorem of Min-Max Classifier The tradeoff cost function consists of minimizing the dispersion of each intra-class feature candidate f versus images Wa; feature candidate F versus images Mb; as well as the inverse LMS of inter-class feature separation as follows: Then, the energy function minimization can automatically achieve the min-Max compromise:

, where use is made of the binomial expansion to bound the relative tolerance level of look-alike error a=1/σa with respect to the variance of “a”, for a specific “a” of the class, and likewise, the level b for a specific “b” of another class. We can simplify the computation of standard deviation with a-priori combinatorial factor weighting factor when single best features f and F, in the min-Max sense, are chosen for all features candidates, say top 5 for each of 2 cohort groups in our experience. These seemingly complex formula help us achieve intuitively clear min-Max separation to stitch two separation hyperspace together at a common origin as revealed by the estimated bound of the right-hand summation, once all the hopeful candidates fk give up their delegation right to only one feature f representing them:

We then apply arbitrary probing image after proper pre-processing including segmentation, perspective correction, etc., we can test the ability separate one class from the other in the most wanted archival sample, before further zero in identified the membership within the subclass of gender. Our final goal is to develop a common platform to conduct the standoff biometric performance virtually in terms of the combined Receiver Operation Characteristics (ROC), which will be plotted of probability of detection (PD) versus the False Alarm Rate (FAR) for various standoff distances in terms of the pixel on target (POT).


References: [1]

“History, Summary of Different Biometric Traits, Comparison Chart”, National Center for State Courts (NCSC).

[2]

“Information Fusion in Biometrics”, Arun Ross, Anil Jain, Department of CS & Eng., Michigan State University.

[3]

Harold Szu, Charles Hsu, Joseph Garcia, Brian Telfer, “Fingerprint data acquisition, de-smearing, wavelet feature extraction, and identification,” p. 96-118, Proc. SPIE Vol. 2491, Wavelet Applications II, SPIE Conference, Orlando 1995.

[4]

Tom Hopper, Fred Preston, “Compression to Gray-scale Fingerprint Images”, pp 309-317, DCC 1992 Data Compression, IEEE Computer Society Press, Los Alamitos, CA 1992.

[5]

Harold Szu, Brian Telfer, Joseph Garcia, “Wavelet transform and neural networks for compression and recognition,” pp695-708, Neural Network, Vol. 9, Issue 4, June 1996.

[6]

Special Journal Issues in Wavelets: IEEE Trans. Information Theory Vol, 38, March 1992; Optical Engineering Sept. 1992; IEEE Trans. on Signal Processing Dec. 1993; Optical Engineering July 1994.

[7]

Sellahewa, H.; Jassim, S., "Face recognition in the presence of expression and/or illumination variation Automatic Identification Advanced Technologies," Fourth IEEE Workshop on Volume, Issue , 17-18 Oct. 2005 Page(s): 144 – 148, 2005.

[8]

M. Kirby and L. Sirovich," Application of the Karhunen-Loeve procedure for the characterization of human faces, "IEEE Trans. PAMI., Vol. 12, 103-108, 1990.

[9]

Phillips, P.J., et al, “The FERET Evaluation” in H. Wechsler, et al (eds), Face Recognition: From Theory to Applications, Berlin, Springer-Verlag, 1998.

[10] Sang-Kyun Im, Hyung-Man Park, Young-Woo Kim, Sang-Chan Han, Soo-Won Kim, Chul-Hee Kang and Chang-Kyung Chung, “An Biometric identification system by extracting hand vein patterns”, Journal of the Korean Physical Society, 38(3): 268-272, March 2001. [11] J. M. Cross and C. L. Smith, “Thermographic imaging of the subcutaneous vascular network of the back of the hand for biometric identification”, Proceedings of 29th International Carnahan Conference on Security Technology, Institute of Electrical and Electronics Engineers, 20–35, 1995. [12] Toshiyuki Tanaka, Naohiko Kubo, ”Biometric authentication by hand vein patterns” SICE, Annual Conference in Sapporo, 249-253, Aug. 2004. [13] C.L Lin, K.C. Fan, “Biometric verification using thermal images of palm-dorsa vein patterns”, IEEE Trans Circuits Sys Video Tech, 14(2): 199-213, Feb. 2004. [14] N. Miura, A. Nagasaka, and T. Miyatake, “Extraction of Finger-Vein Patterns Using Maximum Curvature Points in Image Profiles,” pp. 347-350, Proceedings of the 9th IAPR Conf. on Machine Vision Applications, MVA2005, Tsukuba Science City, Japan, 2005. [15] Miller, Dan. “Voice Biometrics Market Potential Summary: Applications Review and Assessment,” July 2007, (http://opusresearch.net/wordpress/pdfreports/vbio_mktptnl_leadup.pdf.) [16] Bishop, C. M. (1995) Neural networks for pattern recognition. Oxford, UK: Oxford University Press. [17] Ripley, B.D. (1996) Pattern recognition and neural networks. Cambridge, MA: Cambridge University Press. [18] “TESPAR, A powerful New Voice Biometric For Forensic Applications” First International Conference On Forensic Human Identification, London 23-26 October 1999 [19] M. H. George, and R. A. King, “Time Encoded Signal Processing And Recognition For Reduced Data High Performance Speaker Verification Architectures”, 1st Int. Conference Audio And Video-based Biometric Person Authentication (AVBA 97), pp. 377-384, March 1997 [20] R. D. Turney, C. Dick, and A. Reza, “Multirate Filters and Wavelets: From Theory to Implementation”, Xilinx Inc. [21] M. Fleury, R. P. Self, and A. C. Downton, “Multi-spectral Satellite Image Processing on a Platform FPGA Engine”, University of Essex Electronics Systems Engineering Dept.
