Directional Continuous Wavelet Transform Applied to Handwritten Numerals Recognition Using Neural Networks

The recognition of handwritten numerals has many important applications, such as automatic lecture of zip codes in post offices, and automatic lecture of numbers in checknotes. In this paper we present a preprocessing method for handwritten numerals recognition, based on a directional two dimensional continuous wavelet transform. The wavelet chosen is the Mexican hat. It is given a principal orientation by stretching one of its axes, and adding a rotation angle. The resulting transform has 4 parameters: scale, angle (orientation), and position (x, y) in the image. By fixing some of its parameters we obtain wavelet descriptors that form a feature vector for each digit image. We use these for the recognition of the handwritten numerals in the Concordia University data base. We input the preprocessed samples into a multilayer feed forward neural network, trained with backpropagation. Our results are promising.