Jae Hung Yoo

An ellipse detection method from the polar and pole definition of conics

Abstract A new multistage Hough transform approach to the problem of ellipse detection in images is presented. The proposed approach is based on the polar and pole definition of conics and its advantage lies in avoiding the propagation of parameter estimation error over the successive stages of the Hough transform. The proposed method is capable of detecting partially visible ellipses, overlapping ellipses, and groups of concentric ellipses. A set of experimental results is presented to demonstrate these capabilities of the suggested approach.

Structure-driven induction of decision tree classifiers through neural learning

The decision tree classifiers represent a nonparametric classification methodology that is equally popular in pattern recognition and machine learning. Such classifiers are also popular in neural networks under the label of neural trees. This paper presents a new approach for designing these classifiers. Instead of following the common top-down approach to generate a decision tree, a structure-driven approach for induction of decision trees, SDIDT, is proposed. In this approach, a tree structure of fixed size with empty internal nodes, i.e. nodes without any splitting function, and labeled terminal nodes is first assumed. Using a collection of training vectors of known classification, a neural learning scheme combining backpropagation and soft competitive learning is then used to simultaneously determine the splits for each decision tree node. The advantage of the SDIDT approach is that it generates compact trees that have multifeature splits at each internal node which are determined on global rather than local basis; consequently it produces decision trees yielding better classification and interpretation of the underlying relationships in the data. Several well-known examples of data sets of different complexities and characteristics are used to demonstrate the strengths of the SDIDT method.

Design of multicategory multifeature split decision trees using perceptron learning

Abstract The decision tree methodology is an important nonparametric technique for building classifiers from a set of training examples. Most of the existing top-down decision tree design methods make use of single feature splits at successive stages of the tree design. While computationally attractive, single feature splits generally lead to large trees and inferior performance. This paper presents a new top-down decision tree design method that generates compact trees of superior performance by using multifeature splits in place of single feature splits at successive stages of the tree development. The multifeature splits in the proposed method are obtained by combining the concept of information measure of a partition with perceptron learning. Several decision tree induction results for a broad range of classification problems are presented to demonstrate the strengths of the proposed decision tree design methods.

A general approach for token correspondence

Abstract This paper presents a generalization of the correspondence approach of I. K. Sethi and R. Jain ( IEEE Trans. Pattern Anal. Mech. Intell. 9 , 56–73, 1987), by extending the path coherence criterion to high dimensional vector spaces, where tokens are represented as points. Two algorithms for obtaining correspondence of tokens are described. Experimental results for line and region tokens are presented to demonstrate the general purpose nature of the proposed correspondence approach.

Symbolic mapping of neurons in feedforward networks

It is common to view multiple-layer feedforward neural networks as black boxes since the knowledge embedded in the connection weights of these networks is generally considered incomprehensible. This paper proposes a solution to this deficiency of neural networks by suggesting a mapping procedure for converting the weights of a neuron into a symbolic representation and demonstrating its use towards understanding the internal representation and the input-output mapping learned by a feedforward neural network. Several examples are presented to illustrate the proposed symbolic mapping of neurons.

Extraction of diagnostic rules using neural networks

A neural learning methodology is presented that is capable of providing rules from learned weights. An application of this methodology to systematic lupus erythematosus is demonstrated. It is shown that the proposed approach can disregard irrelevant features in the data and can generate different criteria combinations indicating the presence of systemic lupus erythematosus in a patient.<<ETX>>

SIR: simultaneous induction of rules using neural networks

One major drawback of the decision-tree-based inductive knowledge acquisition methodology is its inability to form high-level features from raw attributes. While neural learning has no such problem, its difficulty is in the opaqueness of the acquired knowledge. The authors address both these issues and present a neural learning methodology that yields production rules formed on the basis of high-level features that are also learned during the learning phase. Furthermore, the competitive component of the learning in the proposed methodology automatically determines the number of rules for a given learning situation. Two examples are presented to illustrate the methodology.<<ETX>>

Mobile robot localization with multiple stationary cameras

The problem of determining the position and orientation of a mobile robot has been addressed by several researchers using sensors of different modalities, including video cameras. Invariably, all the vision-based approaches for robot localization consider that the camera is mounted on the robot and that the robot working environment is assumed to contain prominent landmarks at known locations. In this paper we propose a robot localization scheme where the robot itself serves as the landmark for cameras that are positioned in the environment to cover the entire work area of the robot. Although the proposed approach is applicable for the robots of any regular shape, we develop the solution to the localization problem by assuming a cylindrical shape for the robot. A compete mathematical analysis of the localization problem is given by extending the three-dimensional structure-from-rotational motion approach to the present task. We also examine the implementation issue of the proposed approach and present experimental results to show its effectiveness.

Correspondence using property coherence

This paper presents a generalization of the correspondence approach of Sethi and Jain by extending the path coherence criterion to high dimensional vector space. This allows the same correspondence procedure to be used for a variety of tokens including points, lines, planes and regions. To demonstrate the generalized approach, we apply it to track lines and present experimental results.

Design of radial basis function networks using decision trees

A radial basis function (RBF) neural network is considered as one of the universal input-output functional mapping learning systems. Important issues in designing an efficient RBF neural network are the number of neurons and the shape and location of neurons to define local receptive fields in feature space. This paper presents a solution to these problems using decision tree partitioning. A conversion algorithm from decision tree to RBF neural network is described. Two examples are presented to illustrate the proposed approach.