Chinese Handwriting Recognition Algorithm Research

Recently, with the rapid development of biometric identification technology, computer handwriting identification is becoming an important and necessary part, just as speech recognition, fingerprint identification, iris recognition and face recognition. Handwriting identification is a technique that aims to decide the identity of writers by comparing the same character written by different people and analyzing the writing style. It has widespread application in domain of finance, insurance, public security and judicial department criminal survey and court trial and it has many advantages such as quick distinguishing, high efficiency, free of inspectors' subjective factor and so on. Therefore, to avoid human factor and to make the handwriting identification automatic and intellectual is becoming important research aim in this field.The methods of off-line handwriting identification are discussed, especially the techniques of handwriting image pre-processing and feature extraction are focused on. The main purpose of this research is realizing main algorithm and technology which computer-aided handwriting is involved and offering technical support for computer handwriting identification system. The application background and history of development of writer identification techniques are introduced at first. The nature of the writer identification problem and difficulties are analyzed, and then the overall strategy and the plan of building computer writer identification system, which includes current theoretical algorithms about image processing and pattern recognition, are proposed.Texture analysis is applied in the manipulation, analysis and recognition of images widely. The thesis summarizes many methods that were excogitated in these several decades, in which Gabor transform, also named Short-Time Fourier Transform or Windowed Fourier Transform, is an important method of joint time-frequency analysis in modern signal processing. It shows the joint temporal-frequency property of signal analysis and overcomes the shortcoming of traditional Fourier Transform which fails to present any temporal discrimination ability in frequency domain. It has favorable time-frequency localization, direction characteristic and multi-resolution wavelet characteristic. Under Heisenberg's uncertainty principle, it has been proved that it has optimal joint temporal-frequency resolution. Based on the studies on the physiology of mammal perception system, it has been shown that 2-D Gabor Elementary Function can fit well the receptive fields of the majority of simple cells in mammal visual cortex.For free-format handwritten Chinese characters, a complete and systematic algorithm on preprocessing of handwriting image is proposed in this paper. Connected characters removing is carried out by combining statistical characteristic of different people's characters width with the gauss distribution. According to the location of valley point on projecting images, we select a threshold value as best division point. Then size normalizing and characters mosaic are all realized easily. At the same time line slope is corrected. The proposed methods make the whole preprocessing subsystem more stable and perfect. It lays a dependable foundation for subsequent handwriting recognition. At stage of feature extraction based on multi-channel Gabor filters, a new method combining human visual system and statistical property of character stroke width is proposed. It has good performance of robustness and is applied in the feature extraction of uniform handwriting images and compared with the results when employing traditional experimental method. In the experiment, 49 writers' random handwritings are tested and its correct identification rate was over 98%, which can achieve excellent performance of recognition. In addition, Support vector machine classifier based on statistical learning method is employed in classification stage. We compare three kinds of kernel function including polynomial, radial basis function (RBF) and Sigmoid. And RBF have higher recognition rate when using the same number of training and testing samples.