Feature Selection for Recognizing Handwritten Arabic Letters

Abstract

There are many feature extraction methods for handwritten letters. These methods provide large sets of features that include redundant and irrelevant features. Feature selection is needed to select a subset of features that gives good recognition accuracy and has low computational overhead. We use feature selection techniques to evaluate a large set of features extracted from handwritten Arabic letters. We extract 96 features from the letter’s secondary components, main body, skeleton, and boundary. These features are evaluated and best subsets of varying sizes are selected using five feature selection techniques. These techniques vary in complexity from selecting best individual features, through sequential forward selection, to evolutionary optimization algorithm. The best subsets of selected features include secondary components features, letter form, low-order elliptic Fourier descriptors, moments, size features, and features extracted from the boundary. We use three popular classifiers to evaluate the subsets selected by the five selection techniques and to find the recognition accuracy as a function of the feature subset size. The evolutionary algorithm has the highest time complexity but it selects feature subsets that give the highest recognition accuracies. In most cases, feature subset sizes of about 20 features achieve best recognition accuracy.