The New Jersey Institute of Technology's
Electronic Theses & Dissertations Project

Title:	Automatic analysis of electronic drawings using neural network
Author:	Shi, Yi
View Online:	njit-etd1998-102 (xix, 100 pages ~ 3.7 MB pdf)
Department:	Department of Electrical and Computer Engineering
Degree:	Doctor of Philosophy
Program:	Electrical Engineering
Document Type:	Dissertation
Advisory Committee:	Shi, Yun Q. (Committee chair) Hou, Edwin (Committee member) Hung, Daochuan (Committee member) Manikopoulos, Constantine N. (Committee member) Su, Wei (Committee member)
Date:	1998-08
Keywords:	Pattern perception. Neural networks (Computer science). Digital techniques.
Availability:	Unrestricted
Abstract:	Neural network technique has been found to be a powerful tool in pattern recognition. It captures associations or discovers regularities with a set of patterns, where the types, number of variables or diversity of the data are very great, the relationships between variables are vaguely understood, or the relationships are difficult to describe adequately with conventional approaches. In this dissertation, which is related to the research and the system design aiming at recognizing the digital gate symbols and characters in electronic drawings, we have proposed: (1) A modified Kohonen neural network with a shift-invariant capability in pattern recognition; (2) An effective approach to optimization of the structure of the back-propagation neural network; (3) Candidate searching and pre-processing techniques to facilitate the automatic analysis of the electronic drawings. An analysis and the system performance reveal that when the shift of an image pattern is not large, and the rotation is only by nx90°, (n = 1, 2, and 3), the modified Kohonen neural network is superior to the conventional Kohonen neural network in terms of shift-invariant and limited rotation-invariant capabilities. As a result, the dimensionality of the Kohonen layer can be reduced significantly compared with the conventional ones for the same performance. Moreover, the size of the subsequent neural network, say, back-propagation feed-forward neural network, can be decreased dramatically. There are no known rules for specifying the number of nodes in the hidden layers of a feed-forward neural network. Increasing the size of the hidden layer usually improves the recognition accuracy, while decreasing the size generally improves generalization capability. We determine the optimal size by simulation to attain a balance between the accuracy and generalization. This optimized back-propagation neural network outperforms the conventional ones designed by experience in general. In order to further reduce the computation complexity and save the calculation time spent in neural networks, pre-processing techniques have been developed to remove long circuit lines in the electronic drawings. This made the candidate searching more effective.