On-Line Handwritten Chinese Character Recognition Approach Based On Sentence Level

Though online handwritten Chinese character input had been fully developed, for the various of Chinese character types, the diversity of shapes for the same character, shape similarities of different characters and the Ligatures etc., the recognition algorithms that play key role in recognition systems are still far from satisfied. And a lot of researches focus on seeking better solutions by extracting more distinct features.Most of the existing handwriting recognition products limit the input area of Chinese character to a small block to reach an acceptable performance, and can only recognize one character at each time. It not only keep the users out from fluent input process since they have to select the right character from a list of candidates at each time, but also limit the improvement of recognition precision since the context information is not applied by this method. With the popularity of full touched screen devices, writing more characters or even a sentence at one time is very convenient. It motivates the research on "Sentence level" or even "Document level" Chinese character recognition methods.In this paper, we design and implement a sentence-level online handwriting recognition system based on our recognition algorithm. In our system, users can input under an unconstrained condition. When the user completes a sentence or a paragraph, the whole text will be recognized and displayed on the screen. The handwritten document is also saved at the same time. The system consists of two parts: recognition and post-processing. The accuracy of single word recognition is an important part in the system, especially for feature extraction. In this paper we modify the continuous NCFE method by preprocessing the vector endpoints. We also applied various assignments for different situations to enhance the robust of the recognition system. The AP clustering algorithm is applied in the step of coarse classification. Being compared with other clustering algorithms, AP clustering algorithm has advantage on both the recognition accuracy and the clustering efficiency. In the post-processing stage, we integrate the character recognition results with n-gram language models by weighted interpolation method.To evaluate the performance of our algorithms, experiments are conducted on the HIT-OR3C and CASIA data sets. The results showed that, being compared with other methods, our algorithms get higher recognition accuracy with the convenience of sentence-level handwriting system.