A Method Of Handwriting Identification Based On Wavelet Packet Analysis And Neural Networks

The computer writer identification based on handwriting is one of the research focuses in the field of Pattern Recognition and Image Processing. Because wavelet analysis has excellent characteristic in Timefrequency Multiresolution, it has been rapidly developed in the field of Information Processing in recent ten years. Applying wavelet analysis to computer handwriting identification and exploring some handwriting characteristic distributed in different frequency domains have become a new research idea of computer handwriting identification.The writing process is regarded as the distributing process of handwriting energy by author. According to this point of view, a set of characteristic analysis methods for computer handwriting texture are presented. Firstly, author presents a normalization method which is entirely different to tradititnal methods. The method can not only retain the spatial information of handwriting samples, but also efficiently simplify the preprocessing of them. Secondly, a 2-D Wavelet Packet best basis characteristic extraction method which is essentially different to the familiar basis matching method is described. The method directly executes wavelet packet decomposition of handwriting texture using wavelet packet basis db6 at scaling 3 in 2D space, then reconstructs the decomposition coefficient of 15 wavelet packet best basis which are took by Shannon Entropy Cost Function. In order to obtain the energy characteristic of texture sub-image, a nonlinear handwriting energy measure method is presented in this paper. It has been proved that the method has the adaptive capability to match the texture. After a series of above-mentioned processing, a Chinese character image can be compressed into an energy measure matrix which includes 15 elements. A BP Neural Networks is designed to learn and classify the result coming fom the combination and standardization of every energy measure matrix.The experiment proves that the recognition rate of the system can reach to 95% or more when the number of the experimental sample is limited.This system is implemented by mixed programing based on C++ and Matlab.