Comparing filter and wrapper approaches for feature selection in handwritten character recognition

It is generally agreed that the selection of an appropriate set of features is a fundamental process in the development of any pattern recognition system. Its purpose is to identify the truly distinctive subset of features to reduce the size of the search space, without decreasing the classification performance. This problem is particularly relevant in the field of handwriting recognition, due to the enormous variability of character shape, which has led to the development of a large variety of feature sets that are becoming increasingly larger in terms of the number of attributes. While promising, the results achieved so far have several limitations, which include, among others, the computational complexity of selecting and evaluating feature subsets and the difficulty in evaluating the interactions among features. In a previous study, we tried to overcome some of the above limitations by adopting a feature-ranking-based technique: a large study was carried out considering different filter-based techniques for feature subset evaluation. The aim of this work is to extend the previous study by presenting a broad comparison between filter and wrapper techniques for feature selection in the field of handwritten character recognition. In the experiments, we analysed one of the most effective and widely used set of features in handwriting recognition, applied to standard real-word databases of handwritten characters. The experimental results confirmed that filter and wrapper approaches achieve similar performances, with the former selecting fewer features at a lower computational cost.