Optical Digit Recognition for Images of Handwritten Historical Documents

This paper presents a study on recognition of handwritten digits extracted from images of historical documents. The work is part of ProHist, a project aimed to make easier the diffusion of historical documents. Our objective herein is to recognize the dates written in the document for indexing purposes. In this paper we investigate the influence of the feature extraction and the classification method on the performance of the recognition system. We report on experiments considering principal component analysis and undersampled bitmaps as feature extraction methods and k-nearest-neighbors, RBF neural networks and support vector machines (SVMs) as classifiers. The conclusion shows that using undersampled bitmaps combined with SVMs yield the best classification result.