Improvement And Analysis Of Online Handwritten Chinese Character Recognition Based On Deep Convolutional Neural Networks

With the growing maturity of information technology,the trend of electronization in various fields is becoming more and more obvious.In frequent human-computer interaction,efficient,fluent and natural interaction will bring a good experience to the user.As an important part of human-computer interaction,handwritten input is widely used in the field of specific crowd and vehicle navigation because of its low learning cost,natural input mode,less ambiguity,and separation of sight from input equipment.So the requirement for recognition accuracy is higher than before.In recent years,computing cost and storage cost have been declining,computing speed and storage capacity are increasing.Deep learning technology has made breakthrough progress in the fields of image,audio and video,natural language processing and so on.It is also gradually applied to the traditional pattern recognition field of online handwritten Chinese character recognition.In this paper,a handwriting optimization algorithm is proposed based on the traditional preprocessing method.The improved four deep convolutional neural networks are used to recognize the online handwritten Chinese characters.The specific contents are as follows:1.The online handwritten Chinese character data is converted into binary images with direction information.A convolutional neural network is selected as the classifier.The aspect ratio adaptive normalization method based on twelve different mapping functions is used respectively,and the optimal normalization method is selected.Four feature extraction methods are used for the normalized data and the best one is selected by classifier.2.Aiming at the noise interference problems such as error handwriting and redundant handwriting which are easy to appear in the process of data acquisition,a handwriting optimization algorithm based on direction and proportion to remove adjacent handwriting is proposed.Experiments are carried out by combining the optimal normalization method,feature extraction method and selected optimal network structure.By comparing the recognition rate of four direction,eight directions,sixteen direction and different retention ratio of handwriting,the parameters of handwriting optimization are determined.3.Use the best normalization method,feature extraction method and handwriting optimization algorithm to process online handwritten Chinese character dataset.Four kinds of classical network VGGNet,Goog LeNet,ResNet and DenseNet are selected as the design basis.The network structure which meets the characteristics of experimental data is designed and trained by MXNet framework.Finally,various handwriting processing methods and networks are compared and analyzed.After the comparison and analysis of the experimental results,the eighteen channel features based on the central moment adaptive normalization method and the eight direction handwriting optimization algorithm are selected,and the accuracy of the test dataset is 97.79% by using the ResNet network.