Study On Off-line Chinese Handwriting-based Writer Identification

Writer identification is a science and technology to judge the identification of the writer by analyzing and comparing the writing style and feature of different people. With the development of the technology of biometrics identification, writer identification as a necessary part of biometrics identification has a widespread application in many fields, such as finance, insurance, public security, justice and archeology. The kind of the method has quick distinguish, high efficiency, free of inspector's subjective factor etc.The paper discusses the method of text independent off-line Chinese handwriting writer identification, especially focuses on feature extracting of the handwriting text image. Under the high dimension condition, wavelets can't make the best of the unique geometrical property of data itself, and it's not the optimal or sparsest representation. To solve the problem, the paper applies the multi-scale geometric analysis theory to extract the feature of handwriting text image. The main topics in this thesis are as follows:Firstly, the theory of the anti-aliasing Contourlet transform is discussed. The anti-aliasing Contourlet transform not only contains the advantage of achieving a multi-scale and multi-dimensional description of Contourlet transform, but also eliminates the disadvantage of frequency aliasing of Contourlet transform. Iit's better able to describe the texture information of text image. Therefore, the off-line handwriting-based writer identification with anti-aliasing Contourlet transform is presented. A large number of simulation experiments show that the coefficients of anti-aliasing Contourlet transform is accorded with the GGD model at different decomposition level. The paper adopts the method of the combination of GGD model and KL distance to get the image retrieval precision rate. In comparison with a single wavelet transform,the complex wavelet transform and Contourlet transform, the method increases the accuracy about 22.3%, 7.5%, 2.3%, respectively.Secondly, the theory of PDTDFB is discussed. The PDTDFB transform combines the no-separable directional filter bank, not only contains the multidirection property which the wavelets can't express, but also captures the contour information effectively. It provides a multi-scale, multi-direction and scalable decomposition for text image. Theerfore, the off-line handwriting-based writer identification with PDTDFB transform is presented. The experiments show that the coefficients of PDTDFB transform is accorded with the GGD model at different decomposition level. The experiment uses the same method as anti-aliasing Contourlet transform to get the precision rate. In comparison to the scalar wavelet transform, the complex wavelet transform and Contourlet transform, the method increases the accuracy about 22.3%, 7.5%, 2.3%, respectively.