Machine recognition of hand printing

Abstract

The problem of pattern recognition as applied to line drawings has been studied and a generalized method developed which is based on a hierarchical sequence of classification stages. Recognition proceeds in two steps: First the individual pen strokes are abstracted into straight line segments; then the normalized pattern of lines is identified with one member of the possible pattern set. The procedures used to abstract the line segments are reasonably invariant to size, position, line width, noise, and certain small topographic variations. This line recognition stage produces normalized, quantized measurements of each segment's length and angle, and its bearing direction and distance from the pattern's centroid. These data are presented in a matrix form which is convenient to manipulate both for the recognition programs and for certain topographic preprocessing procedures. Recognition is accomplished by comparing the unknown matrix with a catalog of arrays which are formed by accumulating statistics on known samples. The classification process makes use of two measures of the unknown against the arrays: The first is the correlation coefficient which yields a provisional identification. The second is an insignificance index which, by providing a measure of the reliability of the correlation coefficient, serves to improve the identification. The method was tested on hand printed letters. The identification was correct for 93.5% of the unknown samples. Using the insignificance index resulted in a reduction of the error rate over that of the unmodified correlation-based identification by a factor of 2.5. The operation of the entire system was simulated on a general purpose digital computer.