Writer Adaptive Model For Chinese Handwritten Characater Recognition

The classical handwritten characters recognition process, which includes thewriting of characters, character recognition, and results verification, usuallyinterrupt user’s writing influence, and thus reduce the input efficiency. Inaddition, when it came to chose recognition candidate, the classical input modelsare unable to use the context information, which may lead to the lowerrecognition rate. While various users have quit different writing behaviours, theuser independent handwriting recognition, models do not fully use users’ writinghabits, which also affect on the increasing of recognition rate for specific users.To address above problems of classical handwriting input methods, thisdissertation designs a chapter based on-line handwriting recognition method.This method presents handwritten characters and the recognized text charactersto the user at the same time, which provides users the ability of editing thehandwritten characters and verifying recognition results on the same interface.The method is able to use the chapter context to improve the recognition rate. Itcan also satisfy the requirement of fluent writing. By using the contextualinformation of chapter, the system can create and update user specific languagemodels timely. With the language model, the method could improve therecognition performance and increase the precision rate in some extent.The focus of this dissertation includes three parts. Firstly, the similaritycomputing method is introduced. According to the users’ handwriting documents,the method counts handwritten characters that are similar with each other. TheWDTW algorithm is designed to compute the similarity of handwrittencharacters. It is used in the similarity analysis of users’ handwritten documents.Secondly, according to the context information, the charater-based n-gramlanguage model is used to adjust the ranking of candidat es. Finally, the useradaptive model is constructed. By incremental learning algorithm, the user’shandwritten features are learned continuously and the user specific models areadapted. During the process of updating user specific models, the incrementallearning of Linear Discriminant Analysis (LDA) and the incremental learning ofModified Quadratic Discriminant Function (MQDF) are combined together.According to above method, the chapter based handwriting recognitionsystem is modified to improve the recognition rate. By using WDTW algorithm,the users’ interactions are reduced effectively. According to the similaritiesamong handwritten characters, when the users modify the recognition result ofone handwritten character, the system may adjust the recognition results of thosehandwritten characters that are similar with this character automatically. It improves the handwriting recognition accuracy of the real application indirectly.Adding the character-based Bigram language model can also improve the systemperformance by adjusting the first recognition candidate. The effectiviness ofuser adaptive model and character based bigram language model are finallyverified via the experiments conducted in this dissertation.