The authors present a discussion of the different aspects and requirements of the CAD conversion problem and describe the general architecture and algorithms of one commercially available CAD conversion system, the GTX 5000. Scanning, vectorization, text recognition, symbol recognition, context processing, and cleanup editing subsystems of the GTX 5000 are described. Several possible alternative approaches and algorithms are also compared. EPRI (Electric Power Research Institute)-funded enhancements are discussed, including neural networks for character and symbol recognition, touching and broken character processing, and text/symbol associativity.>

The GTX 5000 drawing processor is a commercial device that scans engineering drawings and converts them to computer-aided design format. As part of this process, it automatically classifies the hand-lettered (or occasionally machine-printed) characters that appear on the drawing. The techniques used in this process are described. Although the basic approach (decision trees using numerical features) is standard, the nonparametric recursive technique described for building the decision trees and the use of multiple decision trees for confidence determination are useful refinements. The author also discusses many of the requirements and practical engineering consideration needed to make this commercial product. Finally, results are given on the accuracy of the character-recognition system, based on a set of approximately 9000 hand-lettered characters from engineering drawings.>

The usual approach to numerical training in syntactic pattern recognition involves estimation of the production probabilities of stochastic context-free grammars. Here we describe a different approach, namely that of finding an LMSE discriminant hyperplane between sets of class samples in a space of ``structural indices'' determined by a context-free grammar.