- Biomedical Optics & Medical Imaging
- Defense & Security
- Electronic Imaging & Signal Processing
- Illumination & Displays
- Lasers & Sources
- Micro/Nano Lithography
- Optical Design & Engineering
- Optoelectronics & Communications
- Remote Sensing
- Sensing & Measurement
- Solar & Alternative Energy
- Sign up for Newsroom E-Alerts
- Information for:
Defense & Security
Correlation pattern recognition for biometrics
Correlation filter technology developed in the 1990s for automatic target recognition is being used to recognize people based on their biometric images.
17 May 2006, SPIE Newsroom. DOI: 10.1117/2.1200604.0143
Think of withdrawing cash from an automatic teller machine (ATM). What if a thief steals your ATM card and comes to know your personal identification number (PIN)? For many secure buildings, access is restricted to bearers of an appropriate magnetic swipe card or a radio frequency identification (RFID) tag. But what if that RFID tag or swipe card falls into the wrong hands? Similarly, when we login to computers, we rely on passwords that can be forgotten or stolen.
In our increasingly security-conscious world, it is important to improve our ability to control access to physical (e.g., buildings) and virtual spaces based on who the person is, rather than on what she knows (e.g., passwords) or what he possesses (e.g., RFID tags). Biometric recognition offers this ability.
Biometrics refers to the physiological (e.g., face, fingerprint, iris image, retinal patterns, DNA, etc.) or behavioral (e.g., gait, dynamic signature, keyboard dynamics, etc.) characteristics that distinguish one person from another. Thus, a given person has a unique biometric pattern since biometric signatures cannot be lost, forgotten, or easily stolen. It is of course possible to compromise biometric verification systems (e.g., using fake fingerprints made out of gelatinous materials) and to counter these efforts, liveness detection methods are being actively developed.
Biometrics can be used to authenticate a person's claimed identity (called verification or the 1:1 matching problem) or to ascertain whether that person is in a database or not (called identification or the 1:N matching problem).
Verification usually involves cooperative subjects in applications such as access control whereas identification applications (e.g., recognizing subjects in surveillance videos) can involve subjects that may or may not be cooperative. An example application of biometric identification is found in some United Arab Emirates (UAE) facilities where access is controlled using the iris patterns of workers. Their patterns are matched to the stored iris patterns of expellees who are then prevented from gaining illegal access. Biometric recognition is used to refer to both verification and identification. Usually, techniques that yield improved performance in verification tend to improve identification performance and vice versa.
Many biometric modalities produce images, e.g., face, fingerprints, and iris. The goal is to match the live biometric images to those collected during enrollment. The major challenge is that biometric patterns often exhibit significant inter-class variability (e.g., the appearance of face images is affected by pose, illumination and expression) and inadequate differences between various subjects.
Well-designed biometric verification systems therefore attempt to decrease both the false accept rate (FAR) i.e., the rate at which impostors are incorrectly accepted as authentic, and the false reject rate (FRR) i.e., the rate at which authentic users are incorrectly rejected as impostors.
Most biometric recognition approaches use features computed from the images. The choice of features is critical to the efficiency of recognition systems and many types of features (e.g., Gabor wavelet-based features for iris images, minutiae for fingerprints, etc.) have been investigated. The resulting feature vector selected as the test biometric is compared to the stored feature vector (also called a template) to obtain a similarity value between them. If the similarity is above a chosen threshold, a match is declared and a no match is declared otherwise. By changing the threshold, FAR and FRR values can be adjusted.
While several image domain-based biometric recognition methods have been proposed, frequency-domain methods—such as Fourier transform (FT)—had until recently received little attention. In this context, FT is used as an information-preserving operation. There are several advantages to approaching biometric recognition in the spatial frequency domain (i.e., using the 2D FT of images) and this class of recognition methods is increasingly referred to as correlation pattern recognition (CPR).1
CPR owes its origins to the pioneering work of VanderLugt,2 who showed how correlations between two images can be computed using coherent optical systems employing holographically-prepared complex masks in the frequency plane.
Another important development in this field3 was the ability to create correlation filters that can represent multiple distortions (e.g., in-plane rotations) of a target of interest. Since these initial contributions, both correlation filter theory and design have made significant advances.1 While the initial focus was on automatic target recognition applications, we have recently started to successfully apply correlation filters to biometrics. Another major difference from the early work is that we now do correlations in digital hardware due to the high speed with which 2D fast Fourier transforms (FFT) can be implemented.
Figure 1 schematically describes the use of CPR for face verification. During enrollment, a few face images of the authentic user are acquired. The 2D FTs of these images are then used to generate a 2D filter. This filter array is stored (perhaps on a smart card) to identify that authentic user. When this authentic user presents his face image during verification, the 2D FFT of his live face image is multiplied pixel-wise by the stored filter array and the resulting product array is input to a 2D inverse FFT to produce a correlation output array. As illustrated in Figure 1, the correlation output exhibits a sharp peak for the authentic user and no such discernible correlation peak for an impostor.
Figure 1. Schematic of the use of correlation pattern recognition for biometrics.
CPR offers other advantages as shown in Figure 2. In this example, the filter was designed using three training images (representing the face of the subject with illumination-shadow in the left half, in the right half and no shadow) of one subject retrieved from the well-known Carnegie Mellon University PIE (pose, illumination and expression) face database.4 The filter was tested with a face image (of the authentic subject) where part of his face was blocked and the image was off-center.
Figure 2. Left: Centered and full test image (top) and resulting correlation output (bottom). Right: Shifted and partially occluded test image (top) and resulting correlation output (bottom).
The shifting of the image also shifted the correlation peak to a new position (Figure 2, bottom right), a result of the shift-invariance of the filter, but without affecting peak sharpness and thus the verification performance. In contrast, many image-based algorithms fail completely when the live face image has one of the eyes closed or obscured, since they use the eye centers as a starting point to normalize the image.5 However, no single region is overly critical for correlation-filter performance, thus conveying to CPR methods the ability to offer graceful degradation.
Face recognition is gaining increased importance. In recognition of this fact, the National Institute of Standards and Technology (NIST) has recently created the Face Recognition Grand Challenge (FRGC)6 to promote the development of face recognition research. We have applied CPR methods to FRGC data with excellent results.7,8
While this short summary focused on face recognition, CPR is proving equally applicable for other image biometrics.
Vijayakumar Bhagavatula and Marios Savvides
Electrical and Computer Engineering, Carnegie Mellon University
Prof. Vijayakumar Bhagavatula is a Professor in the ECE Department at CMU. He served as the Acting Department Head during 2004–2005. His research interests include pattern recognition and coding and signal processing for data storage. He has co-authored some 400 publications and is a Fellow of both the OSA and the SPIE. In addition, he has served on the technical program committees of SPIE conferences Biometrics for Human Identification and Optical Pattern Recognition and has published several papers in SPIE conference proceedings.
Dr. Marios Savvides obtained his PhD in Electrical & Computer Engineering, Carnegie Mellon University in 2004.He is currently a Research Assistant Professor there with a joint appointment in Carnegie Mellon's CyLab. His research interest is in Biometric Recognition of Face, Iris and Fingerprint and Palmprint modalities.He is a member of IEEE and SPIE and serves on the Program Committee of several Biometric conferences. sHe is also listed in 2006 Edition of Marquis' Who's Who in America. He has authored and co- authored over 60 conference and journal articles in the Biometrics area and has filed two patent applications in this field.