A Study on the Effect of Varying Training set Sizes on Recognition Performance with Handwritten Bangla Numerals

A study showing how the recognition performance of an MLP based classifier varies with variation in the training set size is presented in this paper. The training set for the work is formed with samples of handwritten Bangla numerals. For recognition of handwritten Bangla numerals, we have used directional features extracted from the contour of each numeral. To extract these features, the minimum bounding box containing the image of each numeral is first segmented into few blocks and then the direction code histogram is computed with each of these blocks. Peak values of each such histogram are considered as the feature values of the corresponding blocks. Considering all the blocks a 100 element feature set is formed for representation of each image pattern and a database of 12000 numerals are used for the same.