Fast Devanagari numerals recognition using improved foreground sub-sampling technique

Abstract

In this manuscript, a fast improved technique is proposed for handwritten Devanagari numerals recognition. Today's scenario demands along with the high recognition rate require a technique which is fast and computationally efficient. Existing technique Foreground sub-sampling (FS) is based on the horizontal and vertical projection computation at each granularity level to find the division points. Here we proposed a technique through which the computation of horizontal and vertical projection at each granularity level becomes fast and efficient by using vertical and horizontal integral image. If sample image is 90 by 90 sized by FS technique at granularity level 3, 62100 addition (+) operations are required to find 85 division points while in our proposed technique only 18000 addition (+) operations are enough to compute same features. So the computation time of foreground subsampling technique is reduced 3 times. Our database (handwritten Devanagari numerals) is given by ISI (Indian Statistical Institute), Kolkata. The size of training and testing data are 18783, 3763 respectively and 98.97 % recognition accuracy is achieved.