Research On Real-time Recognition Of Online Handwritten Chinese Text

Online handwritten text recognition has been a hot topic for many research and engineering studies.Unlike the English text handwriting recognition task,Chinese text has more character classes,hyphenated characters and similar characters are more difficult to distinguish,and documents are generally written with longer text lengths,which brings great difficulties to both recognition accuracy and speed.With the gradual advancement of various deep learning methods,image-based text line recognition methods have made some progress in recent years,but these methods often require images of larger size to retain more detailed features.At the same time,there are a large number of long text lines in handwritten data,which also leads to longer overall data image size width.In online Chinese handwritten text recognition in real time,the larger image size takes longer time to recognize,which is very likely to cause the recognition system to lag and affect the user experience,and it is very difficult to achieve smooth online handwritten text recognition in real time with this method.This paper optimizes and improves the shortcomings of the current online Chinese handwritten text real-time recognition in terms of accuracy and speed,and the main contributions and contents are as follows:1.In this paper,two text recognition models are designed: one is the general text recognition model CRNN-E128,which is mainly designed for convolutional neural networks and used for handwritten text recognition;the other is the graph convolution-based text recognition model CNN-GCN,which uses graph convolution for graph modeling between time-step nodes on image-based handwritten text recognition.Meanwhile,this paper designs four different online Chinese handwritten text recognition models based on graph convolution to compare and analyze the effect of graph convolution modeling in different cases.Experiments on CASIA-OLHWDB and ICDAR-2013 competition datasets show that the CRNN-E128 model proposed in this paper achieves high accuracy so far(CR values of 97.35 and 94.49,and AR values of 96.43 and 93.45,respectively);the CNN-GNN model discards recurrent neural networks to avoid the time consuming effect of a large number of recurrent The CNN-GNN model discards the recurrent neural network and avoids the time consuming operation of recurrent neural network,and the model executes faster with better accuracy.2.In this paper,I propose a local feature map dynamic update method for on-line handwritten Chinese text recognition in real time.A local convolution and local update operation is performed according to the newly written strokes to obtain a new feature map at the current moment.This operation avoids a large number of repeated calculations,increases the recognition speed,and makes the recognition more smooth,thus realizing a smooth and lagfree real-time recognition.In addition,the real-time recognition method proposed in this paper can solve the problem of backward pen insertion and significantly increase the real-time recognition speed without any loss of accuracy.It is noteworthy that the real-time recognition method proposed in this paper maintains the original recognition accuracy and solves the backward pen insertion problem while significantly improving the real-time recognition speed.Experiments on CASIA-OLHWDB and ICDAR-2013 competition datasets show that the realtime recognition speed of this paper is exponentially improved compared with the traditional text real-time recognition method.The real-time recognition speedup is 4.08 times on the generic model and 7.31 times on the graph convolution-based model,allowing smooth realtime recognition of large-size text images.