Font Size: a A A

Handwritten Chinese Character Recognition

Posted on:1997-07-01Degree:DoctorType:Dissertation
Country:ChinaCandidate:L W JinFull Text:PDF
GTID:1118360182997876Subject:Communications and electronic systems
Abstract/Summary:PDF Full Text Request
Computer recognition of handwritten Chinese characters (HCCR) is a popularresearch filed of considerable academic and industrial interests. This thesis studiesthe problem of HCCR on the basis of statistical approach, structural approach andneural networks approach respectively. The main work includes: 1,Building of Handwritten Chinese character database. About 300 sets ofChinese character handwriting samples are collected, each contains 3849 categoriesof characters in the GB23012-80 national standard Chinese character set I. Acharacter database preprocessing and management system has been developed. 2 , Based on the stroke analysis of handwritten Chinese characters, twostructural approaches for HCCR are presented:a, Consistent Relaxation Matching (CRM) Approach. CRM deals with amatching problem from the point-view of optimization. While applyingCRM to HCCR, we give a new method to define Consistent Functionand Initial Matching Assignment, which play important part in CRM.b, A deformable elastic matching model. In this model, handwrittencharacter is regarded as a kind of deformable pattern, with elasticproperties. We assume that for the same category of character, differentstyles of handwriting share the same topologic structure, but may differin shape details. When matching two characters, the strokes of thecharacters can be deformed so as to search for an optimal matchingbetween them. 3,On the study of statistical approach for HCCR. The Bayes classifier isemployed in our handwritten characters recognition system. As feature extractionacts an important role in pattern recognition system, we propose three novel featureextraction methods for HCCR:a, Two structural feature extraction methods, Stroke Cross CountingFeatures(SCCF) and Peripheral Features (PF) are studied. According tothe characteristics of Chinese handwriting samples, we extend the SCCFand PF respectively to weighted elastic SCCF and weighted elastic PF.Experiments show that the modified feature extraction approaches arebetter than original approaches.b, A new statistical feature extraction method——Elastic MeshingDirectional Decomposition Feature (EMDDF) extraction is proposed.According to the stroke statistical properties of Chinese character, wefirst decompose a handwritten character pattern into four directionalsub-patterns. Then a set of elastic meshes are applied to each of the foursub-patterns respectively to extract the pixel distribution features.Experiments show that this new feature extraction approach is of greateffec and the results we obtained are very promising.4,Application of Neural Networks to HCCR. Neural networks provides manyuseful tools which can be used for HCCR. By using a latest neural networksintegrated simulator ——SNNS, several neural networks models are studied for therecognition of handwritten Chinese character:a, Application of feedforward neural network to HCCR. Incorporatingwith three new features extraction methods, three multilayer neuralnetworks are constructed for HCCR. The results are compared withBayes' classifier.b, A new neural networks model——RBF-DDA is applied to HCCR.Impressive experimental results have been obtained.c, From the basic ideas of multi-expert system, we propos a multi-stageneural network model (MNNM) for HCCR. In this model, several neuralnetworks and various feature extraction approaches are integrated intoan unique pattern recognition system. Experiments show that theMNNM can improve the performance of the recognition systemeffectively.
Keywords/Search Tags:Pattern Recognition, stroke analysis, relaxation matching, elastic matching, handwritten Chinese character recognition, feature extraction, Bayes classifier, neural networks, multi-expert system
PDF Full Text Request
Related items