SVM classifier for recognition of handwritten devanagari numeral
Mahesh Jangid, Renu Dhir, Rajneesh Rani, Kartar Singh · 2011
In this manuscript we recognize the handwritten Devnagari numerals. In our implementation we have used density and background directional distribution features for the zones, in which we divided the numeral samples already. We used the normalized images of samples of varying sizes of 32*32, 40*40 and 48*48. We divided these normalized images into 4*4 (16), 5*5 (25) and 6*6 (36) zones respectively to compute the features for each zone. In all the cases each zone is of size 8*8 pixels. Each zone contains 9 features consisting of one density feature and 8 backgrounds directional distribution features. The zonal density feature is computed by dividing the number of foreground pixels in each zone by total number of pixels in the zone i.e. 64. The other 8 features are based on directional distribution values of background in eight directions. These directional values are computed for each foreground pixel by summing up the value corresponding to neighbouring background pixels given in the specific mask for each direction. For each direction these directional distribution features are summed up for all pixels in each zone. Thus numbers of features finally used for recognition are 144, 225 and 324 for samples of respective sizes in increasing order. For classification purpose we have used SVM classifier with RBF kernel. Our dataset of handwritten Devnagari numerals used is provided by Indian Statistical Institute (ISI), Kolkata. Training data size is 18783 and testing data size is 3763, totally 22546. The optimum 5-fold cross validation accuracies of training data obtained for varying sizes of samples in increasing order are 98.76%, 98.91% and 98.94% respectively. By observing the cross validation results it is conclusive that at the cost of increasing the features size there is only minute increases in the performance. So we recommend 144 sized feature vector to recognize testing samples. The testing accuracy by using 144 features for 32*32 normalized samples observed is 98.51% which is prominent and cost-efficient.