On-line Handwritten Chinese Characters Analysis And Recognition Based On Deep Learning

Handwritten Chinese characters can generally be divided into online and offline handwriting according to the collection method,and divided into handwriting style analysis and character recognition according to the goal of related tasks.With the widespread use of touch screens in life,entertainment,education,etc.,the demand for research on Handwritten Chinese Character Analysis and Recognition(HCCAR)has become more and more important in pattern recognition.Handwriting style analysis,including writer identification and writer verification,is frequently used for authentication.The writer identification is mainly aimed to identify the writer of a single character or sentence;and the writer verification is to judge whether a specific word or text is written by a specific person.Handwritten Chinese Character Recognition(HCCR)is to enable the machine to understand the handwritten information,including single word and text.There are various styles of handwritten text line,such as horizontal,vertical,overlapping,rotation,oblique and multi-lines.However,most of the traditional handwriting recognition methods only concentrate on horizontal text or vertical text recognition,respectively.Features of HCCAR based on traditional methods are difficult to design and inefficient in achieving high performance,hence traditional methods is no longer popular.Focusing on the HCCAR,this paper mainly studies the application of deep learning in Online Handwriting Writer Identification(OHWI)and Unconstrained Online Handwriting Chinese Text Recognition(UOHCTR)based on deep learning.Specifically,the content and innovation of this thesis consist of the following two major components:1.For deep learnning based OHWI,this thesis studied building a new OHWI system with data enhancement techniques,Path-signature features and different deep learning neural networks.The proposed Drop Stroke(DStr)has made a great contribution to the generalization and robustness of the model and the improvement of learning ability as efficient data enhancement methods for online handwriting.We use Path-signature as part of the input features.The experimental results show that the proposed method has achieved better performence on benchmark dataset CASIA-OLHWDB1.0 database.2.We designed a Multi layer Distilling Gated Recurrent Unit(MLD-GRU)model.The model includes compression layers and recursive neural network layers to speed up training while ensuring recognition performance.We proposed a data enhancement method to synthesize handwritten data set which contains horizontal,vertical,overlapping,screw-rotation,right-down,and multi-line texts.Due to the insufficience of online handwritten text lines and unconstrained online handwriting data set,we synthesized data by the proposed data augmentation methods.The proposed framework achives good performance on UOHCTR,and get an accuracy of 91.36% on benchmark dataset ICDAR2013 online text data without language models,which demonstrate superior performance over state-of-the-art techniques.