Abstract
Automatic recognition of handwritten numerals is difficult because of the huge variety of ways in which people write. Attempts in the literature employ complicated features and recognition engines in trying to cope with the variety of symbols. But this makes the process slow. In this work, a hybrid technique is proposed to achieve the objective of recognition of handwritten Devanagari numerals with less time consumption and without sacrificing recognition accuracy. A database of 11,000 samples is created while ensuring that the samples include a variety of handwritings which are written with different writing instruments and in different colors. The features employed are density features and spline-based edge direction histogram features and combination thereof. The database size is reduced by using clustering to identify similar samples and putting only one representative sample in lieu of the whole cluster as well as reducing the number of features using PCA. This two-fold reduction provides a smaller database. A hybrid technique utilizing artificial Neural Networks (A-NN), K-nearest neighbour (K-NN) and other learning methods is implemented to ensure higher recognition accuracy and speed. These ideas are put together to provide a fast and robust scheme for recognition of handwritten Devanagari numerals with high recognition accuracy, i.e., 99.40% at a reasonable speed.