Feature Selection and Model Design through GA Applied to Handwritten Digit Recognition from Historical Document Images

This paper presents a genetic algorithm-based approach that integrates a radial basis function kernel support vector machine applied to pattern recognition. The proposed approach performs feature selection on handwritten digits from historical document images and model design on the adopted support vector machine in order to obtain the best possible recognition performance with the minimum possible feature selection subset. The results generated by this research display that the proposed approach is efficient in finding the parameter set of the support vector machine and the feature selection subset that optimize the pattern distinction into their respective classes. Additionally, this method produces solutions that achieve rate of correct digit classification of 95.10% on average and obtain a pattern representation with approximately 50% reduction in dimensionality regarding the original representation. This paper contributes to the preservation and diffusion of historical documents. Our goal is to recognize handwritten dates in documents for indexing purposes.