The evaluation of feature extraction criteria applied to neural network classifiers

Feature extraction is a crucial part of classification procedures. In this paper we present an approach to utilize feature extraction criteria to predict the potential efficiency of a neural network classifier. Statistical and geometrical criteria are introduced for analysis. The complete system of our research consists of a class of generalized Hough-transformations for feature extraction and a subsequent neural network. The neural network performs the classification based on respective features. For an example we concentrated on a pattern recognition problem-the classification of handwritten numerals. As a result of our work we assign two feature extraction criteria to the employed network for a significant estimation of its efficiency.