Wavelet Coefficients and Gradient Direction for Offline Recognition of Isolated Malayalam Characters

Abstract

This work attempt to use the wavelet transform coefficients combined with image gradient direction as feature vector for the recognition of isolated handwritten Malayalam characters. It has been established that the number of zero crossings in wavelet transform distinctly characterizes an image. This property has been exploited in this work for the recognition of handwritten characters. The images of 71 characters in Malayalam are considered for the recognition purpose. The segmented image of the symbols are thinned and smoothed for further processing. The feature vector proposed in this work is the combination of number of zero crossings in two level Daubechies (Db4) wavelet transform and gradient direction of the image mapped to twelve regions with each region having 30 degree span. A two level Db4 wavelet transform is applied on each processed symbol and the number of zero crossings in each of 20 sub images are counted and recorded. Gradient direction is combined with this to form the feature vector. Multilayer Perceptron classifier is used for classification. We have obtained an accuracy of 98.8%.