Unconstrained Online Handwritten Chinese Text Recognition

Due to the popularity of smart phones, touch screen has become an indispensable part of daily life. Increasing the size of the touch screen is benefit for the handwritten text input method. Currently, 12.5% of the users employ handwriting input method, and this proportion is rising year by year, which provides strong support to this research field. Online handwritten Chinese text recognition include online text line handwriting recognition and online overlaid handwriting recognition. The existing research focused on the long text lines, but the handwritten text input method prefer to short text, and there is a lack of relevant research on overlaid handwriting recognition as well.In view of this real scenarios, this paper expects to study unconstrained online handwritten Chinese text recognition technology and apply it to handwriting input method on mobile devices. In order to achieve this goal, this paper mainly contains the following work:(1) In this paper, the traditional text line recognition technology is applied to short text line recognition, but also apply these techniques to overlaid handwriting recognition. Confidence transformation convert the classifier outputs to posterior probabilities. Experimental results show that the outlier probability of the sigmoidal confidence is suitable for text line, but not for the overlaid handwriting, while the D-S evidence confidence can achieve the best effect under both circumstances. We proposed improved Path Evaluation Functions Modified Weighting with Character pattern Width and Modified Weighting with primitive Segments Number, both of which perform better than the original algorithm.(2) The output of our character classifier is large(about 10000) and complex, and thus we proposes short text rules to improve the recognition rate. By this rule, accurate rate of the text line recognition module increases from 93.78% to 94.23% and the first candidate sentence rate of the text line recognition module increases from 94.60% to 95.17%. Although the recall rate of segmentation algorithm is not ideal, but we still won the correct rate of 85.98%, the character accuracy of 85.10%, the first candidate sentence rate of 83.93%.(3) Character recognition rate is low under overlaid handwriting mode, so we propose an overlaid handwrting judgment algorithm based on linear density, to improve the recognition rate. Finally we get character relative error rate of 15.25%, the relative error rate ratio of 1.69 and character recognition rate increases from 90.95% to 92.33%. After using this algorithm to improve character recognition rate, we won the correct rate of 93.75%, the character accuracy of 93.13%, the first candidate sentence rate of 90.46% and character recognition rate of 92.33%.(4) Fade algorithm is proposed in this paper to friendly interact with users. We designed a text recognition engine and related infrastructure, and applied them to the handwriting input method-- SCUT gPen Chinese handwriting input method and SCUT gPen Traditional Chinese handwriting input method. The engine provides handwriting input services for more than ten thousands users.