APE Weekly Seminar Abstracts - Spring 2003

In this presentation I will survey modeling the selective visual attention. One possible reason that natural visual systems outperform current computer vision systems lies in the fact that natural visual processing systems have an optimal selective attention mechanism. Attention selection can not only reduce the amount of data that is for further analysis, it can also suppress the distracting information, which makes quick focusing on relevant information possible. This presentation also pays attention to the model that explains how the visual system develops. We assume that attention is learned from the visual experience, so that it is possible to put visual system development and visual attention development into a single theoretical framework.

Magneto-Optic Imager is a powerful instrument for the purpose of Nondestructive Inspection of aging aircrafts with its faster speed and easiness of inspections in comparison with conventional NDI instruments. However the MOI images have serpentine patterns as background noises, which makes it a little hard to interpret the images. I will present a new approach, Additive Frame Subtraction, to remove the background noises and boost up interest objects by utilizing the characteristics of MOI images. This idea can also be used to any motion segmentation algorithms to extract motions in sequence of images.

The use of airbags into automobiles has significantly improved the safety of the occupants. Unfortunately, when airbags are deployed in the case of a crash, they can also cause fatal injuries if the occupant is a child smaller (in weight) than a typical 6 year old. In response to this, The National Highway Transportation and Safety Administration (NHTSA) has mandated that, starting in the 2006 model year, all automobiles be equipped with an automatic suppression system. These systems are supposed to suppress the airbag if a child or an infant is occupying the front passenger seat. We are investigating the use of machine vision to classify front-seat passenger occupants into four classes: (i) adult, (ii) empty, (iii) RFIS (Real Facing Infant Seat), and child. The design and integration of such a vision system into automobiles is very difficult due to (i) occupant variability (e.g., different types of infant seats and children clothing), and (ii) extreme lighting variability (e.g., bright sunny days, and night time operation). Our approach is based on recognition-driven segmentation. Preliminary results show that by integrating the segmentation and classification stages of processing, we are able to more reliably recognize the various occupant classes.

This is a presentation of Miller's work on ocular dominance and how correlation models are used to explain the plasticity associated with mechanisms of neural development. Ocular dominance is the eye-preference found in the visual cortical cells. Ocular dominance columns are the stripes or patches of cortex that are dominated across the cortical depth by a single eye. This is a study trying to understand these ocular dominance patches in terms of the neural activity and trying to model it using simple mechanisms as the correlation in activity among input neurons; connectivity between the input neurons and receiving neurons, interactions in the cortical region, etc. A simple model proposed by Miller serves to unify a large group of neural phenomena such as development of patches; determination of their width; refinement of individual cortical receptive geniculate cells; the effects of the breadth of correlation or of opposite-eye anti-correlations on the degree of segregation that develops; and the effects on development of abnormal experience such as monocular deprivation including elements of a critical period.

Biometric authentication systems can store and use multiple templates to eliminate effects of variability in a person's biometric data. This variability can originate from improper interaction with the sensor, variations in environmental factors or temporary changes in the biometric trait. But the number of templates that can be used is limited due to storage and computational complexity considerations. In this work, I will describe two techniques for selecting prototype fingerprint templates from a given set of fingerprint impressions. The experimental results show that the proposed template selection mechanism leads to better performance compared to random selection. The methods can be generalized to other biometric modalities such as face, voice print and signature.

Incremental hierarchical discriminant regression faces several challenging issues: (a) a large input space with a small output space and (b) nonstationary statistics of data sequences. In the first case (a), there maybe few distinct labels in the output space while the input data distribute in a high dimensional space. In the second case (b), a tree has to be grown when only a limited data sequence has been observed. In this paper, we present the Locally Balanced Incremental Hierarchical Discriminant Regression (IHDR) algorithm. It is applicable for various content-based retrieval and data engineering applications. A novel node self-organization and spawning strategy is proposed to generate a more discriminant subspace by forming multiple clusters for one class. We applied the new algorithm to three types of data: face image data, marketing data and visual image data in navigation. The experimental results showed that the proposed Locally Balanced IHDR tree produced a significantly better recognition rate than the existing IHDR and worked well for different data sets.

With the ever increasing reservoir of textual information available, knowledge discovery from textual material becomes increasingly important. Information extraction (IE) is concerned with extracting salient information from documents and piecing it together in a coherent framework. In this talk, I will give an introduction to IE. In particular, I will describe a trainable information extraction system.

Clustering is one of the most fundamental problems in machine learning and pattern recognition. For most partitional clustering algorithms, a partition is always generated, irrespective of the true structure of the underlying data. We propose to measure the validity of a partition by interpreting a clustering algorithm as a statistical estimator and examine the variability of this estimator by bootstrapping. Because of the generality of bootstrapping, the proposed procedure can compare validity of partitions produced by (i) a clustering algorithm with different parameters, (ii) different clustering algorithms, and (iii) different distance measures used in clustering. Experimental results on both synthetic and real data in different scenarios are presented to illustrate the applicability of the proposed measure of cluster validity.

Face and iris identification have been employed in various biometric applications. Besides improving verification performance, the fusion of these two biometrics has several other advantages. We use two different strategies for fusing iris and face classifiers. The first strategy is to compute either an unweighted or weighted sum and to compare the result to a threshold. The second strategy is to treat the matching distances of face and iris classifiers as a two-dimensional feature vector and to use a classifier such as Fisher's discriminant analysis and a neural network with radial basis function (RBFNN) to classify the vector as being genuine or an impostor. We compare the results of the combined classifier with the results of the individual face and iris classifiers.

We present work that has as starting point a very concrete application: segmentation of the left ventricle in cardiac ultrasound images. The generally accepted strategy is to (a) build a model of the heart wall from ground truth, (b) fit the model to data (coarse fit) and (c) do a local refinement of the fit. In this talk we tackle the problem of building a model, and discuss in some detail the problem of curve registration.

The main purpose of the dental biometrics is to identify deceased individuals for whom other means of identification (e.g., fingerprint, face, etc) are not available. We try to identify people using their post-mortem (PM) and ante-mortem (AM) radiographs. In other words, given a PM radiograph, we search the database to locate a matching AM radiograph. The variant quality of the radiographs requires us to perform preprocessing procedures such as image enhancement and restoration. The matching is based on the contour of the teeth currently.

The FBI Wavelet Scalar Quantization (WSQ) compression standard was developed by the US Federal Bureau of Investigation (FBI). The main advantage of WSQ-based fingerprint image compression has been its superiority in preserving the fingerprint minutiae features even at very high compression rates which standard JPEG compression techniques were unable to preserve. With the advent of JPEG 2000 image compression technique based on Wavelet transforms moving away from DCT-based methods, we have been motivated to investigate if the same advantage still persists. In this paper, we describe a set of experiments we carried out to compare the performance of WSQ with JPEG 2000. The performance analysis is based on three public databases of fingerprint images acquired using different imaging sensors. Our analysis shows that JPEG 2000 provides better compression with less impact on the overall system accuracy performance.

What "constraints" are exactly wired into the human developmental program? What "constraints" are minimally necessary for a developmental robot? These are open questions. In this presentation, a neurologically inspired mechanism of developing experience-based priming - predicting the future contexts including sensation and action based on the previous experience - is proposed as a powerful "constraint" for developmental robots.We further report how SAIL developed a sense of novelty based on this mechanism and learned its environment on-line. Especially, this novelty detection model is applied on a well-known "drawbridge" experiment which sheds some light on a controversial issue of "Object Permanence" in psychology.