Research On Fast Handwritten Character Recognition In Natural Scenes

From traditional handwriting input methods to optical character recognition,such as checks,manuscripts,forms,and documents,to image recognition systems,handwritten Chinese character recognition(HCCR)is of great use.Especially in recent years,with the rapid development of touchscreen devices such as smart phones and Pads,HCCR gained more attention.However,there are still two problems in HCCR——robustness and speed,which may block its widespread use.The first problem,robustness reflects in two aspects.One is that the activation function of the current HCCR models is mostly softmax,which leads to a fact that the model will identify a sample which isn’t in training set lexicon as a class in it,with a high degree of confidence.This may seriously affect accuracy and security of recognition.The other is that the distribution of HCCR in natural scenes is quite different from experimental HCCR dataset.Although a high accuracy has been achieved in the training set,there is a great room of accuracy in natural scenes.With regards to the second problem,the recognition speed cannot meet the industrial requirements now.Because the convolutional neural network(CNN)is usually adopted in HCCR while it has a large number of parameters due to the existence of the full connection layer.How to solve these two problems has become a new challenge,restricting the development of technology in this field.In this paper,we’ll introduce three mothods to deal with the problems mentioned above.Firstly,the distance based rejection rule is used to improve the poor performances while using softmax.We gained an over 35% promotion in accuracy in open set recogonition.Secondly,we designed a new cross-model prototype learning to reduce the difference between experimental data and data in natural scenes.Based on experimental data,the accuracy can be significantly improved by only adding a small number of natural scene HCCR samples.Thirdly,the candidate reduction algorithm is used to improve speed of HCCR.Under reasonable conditions,the algorithm can accelerate by 4.73× or even higher on 3,755 categories dataset and14.67× or even higher on 21,003 categories dataset.It even can promote the accuracy while achieved with a 2-times speedup.Extensive data are used in every part of experiment.The largest industrial level dataset is up to 21,003 categories,with over 12 million samples.Moreover,we designed comprehensive control variable experiments to clarify the improvement of proposed algorithms,and designed three demonstration programs for intuitive display.The results show that compared with previous studies,we has significantly improved in several aspects and reached the industrial leading level.