Automatic recognition of handwritten numerals via orthogonal moments using statistical and neural network classifiers
Mandyam D. Srinath, Robert R. Bailey · 1993
This research is concerned with the automatic machine classification of handwritten Arabic numerals. In particular it examines in detail a variety of approaches based on the application of two-dimensional orthogonal polynomials for feature extraction and the use of parametric and non-parametric statistical and neural network classifiers. This work utilizes a data base of over 16000 binary digitized images collected from a large number of individuals covering a wide range of style, neatness, character size, and pen width. Polynomials, including Legendre, Zernike, and pseudo-Zernike, are used to generate features invariant to location, size, and (optionally) rotation. An efficient method for computing the moments via geometric moments is presented. A side effect of this method also yields scale invariance. A new approach to location invariance using a minimum bounding circle is presented, and a detailed analysis of the rotational properties of the moments is given. Pre- and post-processing steps are used to improve the classification accuracy. Classifiers include the Bayes quadratic, k-nearest neighbor, three varieties of Parzen, and multilayer perceptron neural network. Since numerous combinations of features and classifiers are evaluated, a simple ranking of how well they perform would not adequately discern real differences in their relative performance due to statistical scatter in the observed results. To help draw statistically justifiable conclusions about their performance, the method of analysis of variance is employed in this research to compare the mean classification error of both feature types and classifiers. In the past researchers into pattern recognition have, unfortunately, ignored such methods. For rotational invariant character recognition, the highest percentage of correctly classified characters was 91.7%, and for non-rotational invariant recognition it was 97.6%. This compares with a previous effort, using the same data and test conditions, of 94.8%. An analysis of all the results indicate that, overall, the Parzen classifiers performed best, although in particular cases the k-nearest neighbor and multilayer perceptron sometimes were better. The quadratic classifier nearly always gave the most errors. The pseudo-Zernike moments, using either centroid or minimum bounding circle location invariant method, were the best features. The techniques developed here should also be applicable to other areas of shape recognition.