Date of Award

Spring 5-2014

Author's Department

Electrical & Systems Engineering

Degree Name

Doctor of Philosophy (PhD)

Degree Type



Security needs at access control points presents itself in the form of human identification and/or material identification. The field of Biometrics deals with the problem of identifying individuals based on the signal measured from them. One approach to material identification involves matching their x-ray scattering profiles with a database of known materials.

Classical biometric traits such as fingerprints, facial images, speech, iris and retinal scans are plagued by potential circumvention they could be copied and later used by an impostor. To address this problem, other bodily traits such as the electrical signal acquired from the brain (electroencephalogram) or the heart (electrocardiogram) and the mechanical signals acquired from the heart (heart sound, laser Doppler vibrometry measures of the carotid pulse) have been investigated. These signals depend on the physiology of the body, and require the individual to be alive and present during acquisition, potentially overcoming circumvention.

We investigate the use of the electrocardiogram (ECG) and carotid laser Doppler vibrometry (LDV) signal, both individually and in unison, for biometric identity recognition. A parametric modeling approach to system design is employed, where the system parameters are estimated from training data. The estimated model is then validated using testing data. A typical identity recognition system can operate in either the authentication (verification) or identification mode. The performance of the biometric identity recognition systems is evaluated using receiver operating characteristic (ROC) or detection error tradeoff (DET) curves, in the authentication mode, and cumulative match characteristic (CMC) curves, in the identification mode.

The performance of the ECG- and LDV-based identity recognition systems is comparable, but is worse than those of classical biometric systems. Authentication performance below 1% equal error rate (EER) can be attained when the training and testing data are obtained from a single measurement session. When the training and testing data are obtained from different measurement sessions, allowing for a potential short-term or long-term change in the physiology, the authentication EER performance degrades to about 6 to 7%.

Leveraging both the electrical (ECG) and mechanical (LDV) aspects of the heart, we obtain a performance gain of over 50%, relative to each individual ECG-based or LDV-based identity recognition system, bringing us closer to the performance of classical biometrics, with the added advantage of anti-circumvention.

We consider the problem of designing combined x-ray attenuation and scatter systems and the algorithms to reconstruct images from the systems. As is the case within a computational imaging framework, we tackle the problem by taking a joint system and algorithm design approach. Accurate modeling of the attenuation of incident and scattered photons within a scatter imaging setup will ultimately lead to more accurate estimates of the scatter densities of an illuminated object. Such scattering densities can then be used in material classification.

In x-ray scatter imaging, tomographic measurements of the forward scatter distribution are used to infer scatter densities within a volume. A mask placed between the object and the detector array provides information about scatter angles. An efficient computational implementation of the forward and backward model facilitates iterative algorithms based upon a Poisson log-likelihood. The design of the scatter imaging system influences the algorithmic choices we make. In turn, the need for efficient algorithms guides the system design.

We begin by analyzing an x-ray scatter system fitted with a fanbeam source distribution and flat-panel energy-integrating detectors. Efficient algorithms for reconstructing object scatter densities from scatter measurements made on this system are developed. Building on the fanbeam source, energy-integrating at-panel detection model, we develop a pencil beam model and an energy-sensitive detection model. The scatter forward models and reconstruction algorithms are validated on simulated, Monte Carlo, and real data.

We describe a prototype x-ray attenuation scanner, co-registered with the scatter system, which was built to provide complementary attenuation information to the scatter reconstruction and present results of applying alternating minimization reconstruction algorithms on measurements from the scanner.


English (en)


Joseph A O'Sullivan

Committee Members

Arye Nehorai, David G Politte, John W Rohrbaugh, Kilian Q Weinberger


Permanent URL: