APE Weekly Seminar Abstracts - Spring 2005

Fingerprint based identification has become more and more accepted by people. It is partly based on the uniqueness of fingerprint. We build a model based on N.H.P.P to fit the pattern of the minutiae for fingerprint minutiae pattern and try to catch some characteristics of the minutiae distribution to get more accurate estimation of the probability.

In order to counteract the inherent intra-class variability in the signature biometric, the template set for a signature verification algorithm must be selected appropriately. Instead of increasing a user’s threshold to compensate for high variability, this algorithm determines the number of templates for a user by grouping similar individual signatures and assigns each an adaptive threshold based on the sample distribution.

Recent advances in biometric authentication have been accomplished by using multiple classifier methods. One successful and popular example in this regard is the hybrid fingerprint recognition system using minutiae and texture developed in the PRIP Laboratory. More examples of recognition strategies combining local and global information can be found in almost every biometric trait described in the literature. The talk will be focused on the work done by the Spanish ATVS/Biometrics Research Laboratory regarding multiple classifier approaches for signature and multimodal authentication. On the one hand, the core system used by ATVS for the 1st Intl. Signature Verification Competition will be described, and a recent multilevel extension will be introduced. On the other hand, the novel strategies to multimodal biometrics proposed by the ATVS group will be described, including adapted user-dependent score fusion, and quality-based score fusion.

In the increasingly e-commerce oriented society, verifying a user’s identity is critical for carrying out financial and other transactions with trust. Most current authentication systems are password based making them susceptible to problems such as forgetting the password and passwords being stolen. One way to overcome these problems is to employ biometrics (e.g., fingerprints, face images, signatures, etc.) for authentication. This talk will provide an overview of our research in methods to authenticate a person’s identity based on their biometrics. In particular, application of correlation filters to verify the identity based on face images, fingerprint images and iris images will be discussed. Correlation filters are designed in frequency domain and offer several advantages such as shift-invariance, closed-form designs and graceful degradation. Although the focus of this talk is on verification, we will also show results of applying these methods to the task of face identification.

With the "Enhanced Border Security and Visa Entry Reform Act of 2002", Congress called for the installation "not later than Oct. 26, 2004" of "equipment and software to allow biometric comparison and authentication of all United States visas and other travel and entry documents issued to aliens". On January 28, 2004, the Departments of State and Homeland Security asked Congress for a two-year extension of this deadline, saying such a process of document and system redesign "normally takes years." Congress granted only a one-year extension in the deployment mandate. Since 1992, there have been at least a dozen attempts internationally to add biometrics to the border crossing process. None of these attempts has been successful over a large population. In this talk, we will review several case studies: the Schipol "Travel Pass" (1992), the US INSPASS (1994), the Australian SmartGate (2003) and the US-VISIT program (2004). We will discuss the technology advances made in each and the remaining engineering challenges in applying these programs to a general population of travellers. We will discuss why DHS and DOS testified that the US-VISIT program will need more time for development than Congress was ultimately willing to grant. Finally, we will discuss the more basic progress being made in standards and technology for large-scale biometric applications.

Face recognition has been extensively studied during the past decades but still not a well-solved problem. Current state of the art technology shows high performance on a strictly controlled recognition scenario. However the performance suffers from the variations in the face data such as pose, illumination, expression, etc. In this talk, the utilization of 3D models to generate a gallery data with various pose and illumination is introduced. The 3D model is represented as a VRML object and generates a set of 2D projection images with various pose and illumination. The generated gallery images are used for the 2D face recognition system to provide a better performance than that of a single frontal image gallery. The effect of the quality and reality of the 3D models on the face recognition performance will be discussed.

Recent advances in technology have created many large, high-dimensional data sets in pattern recognition, machine learning and data mining. It is important to develop tools that can be used to summarize, analyze and transform the high-dimensional data to a more usable form for different applications. In this talk, I shall talk about some recent advances in clustering and dimensionality reduction. We shall also describe how side-information,i.e., information other than class labels and feature vectors, can be utilized for more useful clustering and dimensionality reduction results.

Autonomous vehicles can dramatically increase throughput on existing highways, reduce commute time and hence less pollution will be generated; can free former drivers for more productive tasks. Autonomous vehicles will also be safer since they will eliminate accidents due to fatigue or human error. The talk will describe the AutoNav program, a joint effort to develop an autonomous vehicle for off-road navigation, and techniques used in this program; also the talk will describe the ongoing autonomous navigation research in EI Lab.

Telecommunication systems enable new and enhanced forms of remote collaboration through sophisticated environments that use reality augmentation to enhance the presence factor. To make these systems more effective, it is necessary to properly convey facial expressions of the users. In this work, we have established a system to capture and synthesize two side views of the face and a supplementary view during training. By modeling the three views using an active appearance model (AAM) and finding a regularized solution to the mapping between the two side views and the supplementary view, we can estimate the parameters for the missing view at runtime. In using the designed approach, we have achieved a completely automatic face-to-face communication system with the added advantage of providing a highly compressed image stream.

In this study, we propose a new method to achieve sparseness via a competitive learning principle for the linear kernel regression and classification task. We form the duality of the LASSO criteria, and transfer an $\ell _ 1$ norm minimization to an $ \ell _ \infty $ norm maximization problem. We introduce a novel solution derived from gradient descending, which links the sparse representation and the competitive learning scheme. This framework is applicable to a variety of problems, such as regression, classification, feature selection, and data clustering.

Prototype-based classification can be understood in a broad sense as any classification rule that explicitly deals with exemplars or representatives from a given training set. Particular instances of this range from the old and well-known nearest neighbor rules to recent approaches as Support Vector Machines. Usually prototype-based methods require dramatic reductions in the complexity as a way of obtaining efficient solutions to difficult problems and avoiding overfitting. This talk will review some relevant previous work and present some progress on heuristic methods for prototype selection methods.

Computer vision research has (slowly) progressed, over the past 25 years, from recognizing the motion of rigid, planar objects to 3-dimensional, articulate non-rigid objects. Today, one of the major objectives is the tracking, recognition and understanding the activities of humans and vehicles moving in everyday surroundings. The effort to develop computer systems able to detect humans and to recognize their activities is part of a larger effort to develop personal assistants. The understanding of human activity is a diverse and complex subject that includes tracking and modeling human activity, and representing video events at the semantic level. Its scope ranges from understanding the actions of an isolated person to understanding the actions and interactions of a crowd. In general, human interactions are diverse and difficult to interpret. Occlusion, shadows lighting conditions make tracking and recognition difficult. Professor Aggarwal will present his research on modeling and recognition of human actions and interactions. The work includes the study of interactions at the gross level as well as at the detailed level. The two levels present different problems in terms of observation and analysis. The issues relating to segmentation and tracking of individual body parts will be explored. At the gross level we model persons as blobs, and at the detailed level we conceptualize human actions in terms of an operational triplet ‘agent-motion-target’ similar to ‘verb argument structure’ in linguistics. We use dictionary-based definitions of human interactions as domain knowledge and construct the classification rules for the human interactions. The applications of research to surveillance and computer animation will be discussed, as well as the issue of what the computer ‘sees’ vs. what a human ‘sees’.

Biometrics-based user authentication has several advantages over traditional password-based systems for standalone authentication applications, such as secure cellular phone access. This is also true for new authentication architectures known as crypto-biometric systems, where cryptography and biometrics are merged to achieve high security and user convenience at the same time. We explore the realization of a previously proposed cryptographic construct, called fuzzy vault, with the fingerprint minutiae data. This construct aims to secure critical data (e.g., secret encryption key, SSN) with the fingerprint data in a way that only the authorized user can access the secret by providing the valid fingerprint. The results show that 128-bit AES keys can be secured with fingerprint minutiae data using the proposed system.

In this talk, I will introduce a new formulation for Bayesian image segmentation which facilitates the use of spatial priors. In particular, I will show how wavelet-based priors can be used for image segmentation. This formulation can be applied for supervised, unsupervised, or semi-supervised segmentation, and with any probabilistic observation model (intensity, multispectral, texture, etc.).

The main obstacle to using wavelet-based priors for segmentation (or other priors for continuous-valued images, such as Gauss Markov models) is that they're aimed at representing real values, rather than the discrete labels needed for segmentation. I'll show how this difficulty can be sidestepped by introducing real-valued hidden fields, to which the labels are probabilistically related. I will show how a (generalized) expectation-maximization algorithm can be derived to perform "maximum a posteriori" segmentation under this model. Experiments on synthetic and real data testify for the adequacy of the approach.