Offline Recognition of Handwritten Numeral Characters with Polynomial Neural Networks Using Topological Features

Group-Method of Data Handling (GMDH) has been recognized as a powerful tool in machine learning. It has the potential to build predictive neural network models of polynomial functions using only a reduced set of features which minimizes the prediction error. This paper explores the offline recognition of isolated handwritten numeral characters described with non-Gaussian topological features using GMDH-based polynomial networks. In order to study the effectiveness of the proposed approach, we apply it on a publicly available dataset of isolated handwritten numerals and compare the results with five other state-of-the-art classifiers: multi layer Perceptron, support-vector machine, radial-basis function, naive Bayes and rule-based classifiers. In addition to improving the classification accuracy and the per-class performance measures, using GMDH-based polynomial neural networks has led to significant feature dimensionality reduction.