Files in this item
|(no description provided)|
|Title:||Management of standard graphic symbols in a computer-aided design and drafting environment using neural network approaches|
|Doctoral Committee Chair(s):||Rendell, Larry A.|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||Computer-Aided Design and Drafting (CADD) systems have become prevalent for producing building design drawings. An ultimate goal of CADD systems is to automate analyses and communication of high-level design information extracted from CADD drawings, a difficult task because of the lack of CADD standards. Using standard graphic symbols attached with symbolic information can help, but locating symbols in large libraries is difficult. AUGURS is a new interactive tool designed to assist CADD users in utilizing standard symbols.
The task of recognizing symbols sketched by CADD users differs from traditional pattern recognition problems in several ways. Standard libraries have over 1000 symbols, grouped into seven disciplines. The large symbol set makes training data difficult to obtain. Since AUGURS is embedded in the CADD system, it must be efficient and compact. Also, it needs to handle irregular distortion in symbols sketched by users. These difficulties are lessened by the special output format that requires AUGURS to perform only "admissible" recognition, classifying the input to a small set of plausible symbols.
The symbol recognition program in AUGURS is a neural network similar to the Neocognitron, but is more compact and efficient and having better recognition performance. The main thrust of the AUGURS approach is a novel network structure encoded with general knowledge balancing the discriminant power and the noise tolerance of the network. To handle large symbol sets, another thrust of the AUGURS approach is to construct a network by first building an integrated network from the internal structures of smaller networks trained on sub-tasks, and then pruning unnecessary components from this integrated network.
This research contains an extensive empirical study of numerous related work varying conditions and parameters. The results demonstrate the superiority of the AUGURS approach over many alternatives, including Zipcode Nets, an unconstrained network, networks using such invariant features as Zernike moments, pseudo-Zernike moments, normalized moments, and Fourier-Mellin descriptors, the Integrated Neural Network, and the connectionist gluing approach. A practicality analysis shows that AUGURS can handle around 100 symbols, about the size of a discipline library. To enable AUGURS to handle even more symbols, future work is planned to augment it with domain-specific knowledge and other improvements.
|Rights Information:||Copyright 1994 Yang, Der-Shung|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9512605|