Size of Training Set Vis-à-vis Recognition Accuracy of Handwritten Character Recognition System

Support Vector Machines (SVMs) have successfully been used for character recognition. In the present study, we have shown how the recognition accuracy of a SVM classifier varies with variation in the training set size. The training set for this work is taken from samples of offline handwritten Gurmukhi characters. For recognition of a handwritten Gurmukhi character, we have used curvature features extracted from the skeletonized image of each Gurmukhi character. Features of a character have been computed based on statistical measures of distribution of points on the bitmap image of character. To extract these features, the image of each Gurmukhi character is first segmented into few zones and then the curvature shape is computed within each of these zones. Considering all the zones, a feature set is formed for representation of each image pattern and a database of 3500 isolated handwritten Gurmukhi characters has been used for the same. The results of investigation presented in this paper show that the size of training set has a significant effect on the accuracy of offline handwritten Gurmukhi script recognition system. Index Terms—Feature extraction; curve fitting; handwritten character recognition; SVM.