APE Weekly Seminar Abstracts - Spring 2004

Many existing biometric systems collect information like gender, age, height, and eye color from the users during enrollment. This information is used only in the event of false user rejection, when a human operator intervenes to verify these characteristics. This process can be made more efficient if these traits are automatically extracted and incorporated in the decision making process. We propose the utilization of these "soft" biometric traits to complement the identity information provided by the traditional (primary) biometric identifiers like fingerprint and face. Although soft biometric characteristics lack the distinctiveness and permanence to identify an individual uniquely and reliably, they provide some evidence about the user identity that could be exploited to our advantage. We present a framework for integrating the soft biometric information with the output of the primary biometric system. Preliminary experiments show that the recognition performance of a biometric system can be improved significantly by making use of additional soft biometric user information like gender, ethnicity, and height.

Aircraft industry has been relied on Non-destructive Inspection (NDI) in detecting sub-surface defects to increase the service life of aircraft and prevent accident. Magneto-optic Imager (MOI) is a relatively new NDI instrument which is easier and faster in defect detection than conventional NDI instruments. MOI generates magneto-optic images that reflect sub-surface structure of test sample. However, MO image contains serpentine pattern noise which degrades the inspection capability. In addition to that, manual inspection method degrades the confidence of test result and makes the task labor-intensive. Three major contributions to MOI inspection of aircrafts are presented in this seminar. First, Motion-based Filtering (MBF) algorithm that effectively removes serpentine pattern noise is developed. Second, a real-time program with MBF algorithm is implemented on a general-purpose computer. Third, an automated rivet inspection algorithm with MB filtered images is developed. A demo of real-time rivet inspection will be provided at the end of the seminar.

Conventional clustering algorithms utilize a single homogeneous clustering criterion for the entire data space that may not conform to the diverse shapes of the underlying clusters. We offer a new clustering approach that uses multiple clustering objective functions simultaneously. Our proposed multiobjective clustering is a two-step process. It includes detection of clusters by a set of candidate objective functions as well as their integration into the target partition. A key ingredient of the proposed approach is a cluster goodness function that evaluates the relative utility of multiple clusters independently using re-sampling techniques. Multiobjective data clustering is obtained as a solution to a discrete optimization problem in the space of clusters. At meta-level, our algorithm incorporates conflict resolution techniques along with the natural data constraints. An empirical study on a number of artificial and real-world data sets (including an image segmentation data set) demonstrates that multiobjective data clustering leads to high quality and robust data partitions.

The performance of face recognition systems that use two-dimensional (2D) images is dependent on consistent conditions such as lighting, pose and facial expression. We are developing a face recognition system that uses three-dimensional (3D) information to make the system more robust to these variations. 2.5D is a simplified 3D (x, y, z) surface representation that contains at most one depth value (z direction) for every point in the (x, y) plane. A 3D face model for each subject is constructed by integrating 2.5D facial scans captured from different views. The matching between a newly captured 2.5D scan and a 3D model is achieved for the recognition purpose. The Iterative Closest Point (ICP) framework is applied in the matching scheme. A similarity metric derived from different facial scan attributes is developed for matching. Results of matching a database of 18 3D face models with 113 2.5D face scans show the promise of the proposed scheme.

The identification of key feature points is important in order to register multiple scans in 3D space. We use three criteria for selecting key feature points; saliency, robustness, and utility. A shape index measurement is used to produce coordinate system independent information that is robust to changes in pose and lighting. Heuristics are then engineered and presented to take advantage of the best feature points using the local shape index information. The results of using simple heuristics will be presented and the limitations of this type of approach will be discussed. We will conclude with some design guidelines that could help develop future feature point extraction systems.

Cross-task learning is very important for an artificial agent to explore in the real world. By cross-task, we mean that the same system must learn multiple tasks incrementally in the same mode, dealing with task specific contexts correctly. With this capability, the agent can learn different tasks and transfer learned knowledge to new tasks. In this talk, I will survey cross-task learning in both psychology and machine learning areas. This talk focuses on two aspects: 1) multitask learning (how to learn multiple tasks) 2) task switching based on attention (how to switch between different tasks). Finally, I will discuss how to formulate cross-task learning as a component of developmental learning.

Sketch matching algorithms are commonly used for indexing and retrieval of documents based on printed or hand-drawn sketches. One could use a hand-held computer to do sketch- based queries to a database containing handdrawn and printed sketches. The basic problems in matching sketches include the representation of sketches, which is invariant to different drawing styles and defining a distance measure between two sketches based on that representation. I will present a matching algorithm using a line-based representation of sketches.

Image segmentation is often defined as a partition of pixels or image blocks into homogeneous groups. These groups are characterized by a prototypical vector in feature space, e.g., the space of Gabor filter responses, by a prototypical histograms of features or by pairwise dissimilarities between image blocks. For all three data formats cost functions have been proposed to measure distortion and, thereby, to encode the quality of a partition.

Robust algorithms for image processing are designed according to the following three steps: First, structure in images has to be defined as a statistical model. Second, an efficient optimization procedure to find good structures has to be determined. I advocate stochastic optimization methods like simulated annealing or deterministic variants of it which maximize the entropy while maintaining the approximation accuracy of the structure measure. Other optimization algorithms like interior point methods or continuation methods are equally suitable. Third, a validation procedure has to test the noise sensitivity of the discovered image structures. This three step strategy is demonstrated in the context of image analysis based on color and texture features.

The information of dental work has been used in forensic dentistry for a long time. We are also trying to utilize this information for person identification. The difference is that the forensic dentists can get the information by observing the actual teeth and jaws, while for our system, the only input is the dental radiographs. So in this seminar, I will talk about detecting the presence of the dental work, and extracting the contours of dental work from X-rays. Problems of improving the detector and applying the contours of dental work for matching will be discussed.

There is an increasing interest in a robot under uncertain environments. Bayesian learning frame is usually used to learn the environment model. In this talk, I first survey techniques such as Gaussian distribution assumption (Kalman Filter) and Markov chain Monte Carlo (MCMC) algorithm in the context of robot navigation. Also I am proposing 1)an Incremental Designed State HMM (IDSHMM) to combine 2-D metric information and topological model together and 2) a Quasi-Bayesian learning algorithm to learn the parameters of IDSHMM incrementally. Finally, I will discuss how to integrate this interactively generated environment model for cross-task learning.

Biometrics-based personal auhtentication systems have several advantages compared to traditional token or knowledge-based systems, but they are also vulnerable to attacks that can decrease their security. In this talk, these attacks will be summarized using a fingerprint matcher as a test bed. A new attack system that bypasses the feature extractor and uses synthetic feature sets in order to gain access to a fingerprint-secured system will be presented. Based on a hill-climbing procedure using the matching scores as inputs, the system generates these sets in a feasible amount of time. Experimental results conducted on a large database will be presented.

Information Bottleneck (IB) is a general approach to extraction of relevant information from data. IB method creates new set of variables {T} that compresses input variables {X} while keeping as much information as possible about relevant target variables {Y}. Numerous applications of IB are known in text categorization, genetic data analysis, Unsupervised image clustering. We provide a review of IB framework and illustrate it with the problems of feature extraction and clustering in several domains.

Personal identification from 3D face images has been investigated since the 1980s. Recently, the amount of interest in the 3D modality has increased and systems for 3D face recognition are appearing in the research literature and on the market. This talk summarizes recent research on face recognition using dense range data. Using the largest publically available database of 3D face images and matchers based on principal components and on alignment, we highlight impressive performance as well as sensitivity to facial expression. In addition, this talk introduces the use of 3D range maps of the hand as matchable biometrics. Fingers extracted from the range data and described using curvature-based labelings are matched using a simple scorer. Preliminary results validate the concept but also highlight challenges that must be addressed. This research is supported by DARPA, ONR and the National Science Foundation.

A sparse representation is defined as one in which most of the components are zero and only a few non-zero. Sparse representation is desirable because 1) it simplifies the model by setting the redundant or irrelevant basis to zero; 2)In kernel-based method the generalization ability improves with the degree of sparseness. In this talk, we will discuss the sparseness problem from a Bayesian Learning framework. From this point of view, the sparseness is achieved from MAP (maximum a posteriori) estimation with a Laplace priori. We also discuss the solutions of this problem and propose to solve this problem in its dual space.