Study On Algrithoms Of Offline Handwritten Chinese Identification And Recognition

With the development of pattern recognition technology, the research of handwritingidentification and character recognition is arousing more and more people’s attention.Chinese characters have their own characteristics with multiple character types andcomplex fonts, and handwriting characters also have too much different writing styles. Inorder to meet the requirements of application, the thorough research of handwritingidentification and character recognition of handwritten Chinese characters has academicsignificance and extensive application value. The main research contents and academiccontribution are presented in the following several aspects.Firstly, considering the problems such as different backgrounds, noise and differentsize of handwriting images samples collected at present, a system for preprocessinghandwriting images is designed in this paper. First of all, considering the backgrounds suchas the grid lines which affect the handwritings, a threshold segmentation method is adoptedto removethe backgrounds. Then, to reflect the writing style of handwriting characters, agrayscale and binarization processing is applied to the image samples. Next, comparing thedenoising effectiveness to the image samples of different methods in experiment, anadaptive median filtering method is selected for image denoising. At last, aiming at theproblem of different sizes of characters or texture image samples, a line, wordsegmentation and size normalization method is designed. The whole preprocessing systemguarantees the following feature extraction effect.Secondly, aiming at the handwriting identification problem of off-line handwrittenChinese characters based on text-dependent samples, we proposethe anisotropic Gaussianfilters to extract featuresof the image samples. And the importance of scale and angleparameters to the features is analyzed through experiment. In addition, according to theproblem that the selection of the filters’ parameters takes too much time, a parameteroptimization method that combines the artificial colony algorithm and LDA algorithm isput forward. Experiment results show this parameter optimization method can greatlyimprove the time efficiency of identification.Considering that the recognition rate of the present algorithms of text-independentoff-line handwritten is low, a feature fusion method for handwriting feature extraction is proposed. This method combines the advantages that the ability local binary pattern methodcan represent the local features and the abilitymulti-channel decomposition methodcanrepresents the global features. In addition the features can be extracted in spatial and timedomain simultaneously. Meantime, the method obtains a considerate recognition ratein thehandwriting identification of text-independent.Finally, we research on the handwritten Chinese characters recognition. Due to theproblem that traditional algorithms are easy to loss feature information in the process offeature extraction, which leads to the decrease of identification rate of the classifier.Aconvolution neural network which extracts features directly from the basic pixels ofChinese characters is adopted for the first time. This method can extract feature andclassify simultaneously, reduce the intermediate links and effectively resolve the problemof information loss. The experimental results show that this method can achieve a goodrecognition rate on handwritten Chinese characters.