Research On Handwritten Chinese Character Recognition

This paper does research on the off-line handwritten Chinese character recognition, which is important to the Chinese character information processing automation and new generation of intelligent computer input.The sample collection is necessary for the research on the handwritten Chinese character recognition. The character collection of NUST603HW, including the samples of nineteen capital Chinese currency characters, make the design of the handwritten Chinese character recognition system for special fields possible. The character image from the input device must be preprocessed to filter noises and enhance useful information. The nonlinear normalization, as the important preprocessing component, is used to resolve the handwritten Chinese character distortion on the character image lattice. All the nonlinear normalization methods resolve the above distortion based on density equalization, and differs on the stroke density description. This paper presents a nonlinear normalization method, which not only adjusts the relative character stroke position to make the stroke distribution uniform, but also adjusts the stroke thickness to make it consistent so as to narrow the character difference within the same class.The Chinese character is made up of strokes or sub-strokes. This paper presents the sub-stroke extraction method to resolve the unsteadiness because of the sub-stroke interaction. The feature matrix may be formed based on the character sub-strokes, including the sub-stroke length, position and direction information and so on. A handwritten Chinese character classifying algorithm is designed based on the character feature matrix with the excellent classifying effect.Fuzzy mathematics is the power utility to resolve the conflict between the computer precise calculation and the fuzzy brain thought. This paper adopts the fuzzy mathematics thought to present the fuzzy direction feature. The fuzzy division of the character lattice image can overcome the stroke position influence to the feature extraction. The fuzzy direction feature combines the fuzzy image division with the fuzzy direction property feature, which describes the stroke direction property of single edge point.There are several distance measurements in the distance classifier, and different distance measurement makes the classifier output different result. In some case the recognition rejection is necessary for the classifier reliability, so this paper designs a kind of recognition rejection policy for the distance classifier.Data fusion is a new data processing technique, including three levels of pixel.feature and decision. This paper presents a data fusion method on the feature level based on Fisher criterion, which analyzes the classifying capacity of different features and corresponding discriminative vectors. The new feature derived from the multiple features fusion benefits from the advantage of single feature to the pattern classifying, which is superior to each fused feature on terms of the classifying performance.The multiple classifiers combination fuses the decision level data. The output information of single classifier has three forms of abstract, rank and measurement Single classifier supplies both the unknown pattern classifying information on the measurement level and the wrong classifying distribution information of the training samples on the abstract level, which are used to design the fuzzy multiple classifiers combination method. The combination classifier judges the unknown patterns based on the fuzzy integration of the above two classes of information. The classifier's output information to the training and test samples on the measurement level is used to design another multiple classifiers combination method, in which the decision information of the combined classifier to the training samples is denoted with the prior knowledge matrix and the decision information of single classifier to the samples to be recognized is denoted with the similarity measurement. The similarity measurement is denoted with...