Application Of Style Adversarial Network And Siamese Network To Handwritten Chinese Character Recognition

Handwritten Chinese character recognition has a wide range of applications in doc-ument retrieval,postal sorting,bill transcription,etc.Traditional character recognition mostly uses artificially designed features,and on this basis,it uses machine learning models to classify.The traditional recognition method has made good progress after a long period of research,but the recognition effect still cannot meet the actual needs.In recent years,with the development of GPU,TPU and other parallel computing hardware and the deepening of deep learning theory research,the study of handwritten Chinese character recognition based on deep learning has attracted the attention of academia,and its performance significantly exceeds traditional methods.However,the existing deep learning models also have shortcomings.First,the model is prone to overfitting the writing style of the training sample,which will cause its classification performance on the new writing style sample to be greatly reduced.Second,most of the existing models are only suitable for closed-set characters whose fixed classification categories are determined during training.The new character categories outside the closed set cannot be classified,which greatly limits the application scenarios of the modelIn order to solve the above problems,this paper applies the style adversarial net-work and siamese network in deep learning to the problem of handwritten Chinese char-acter recognition.The specific research work is as follows:1.Aiming at the problem of the impact of writing style changes on recognition performance,a style adversarial network(SAN)is proposed.Through the adversary training of character recognition and writer recognition,the model can better extract character features that are not related to writing style,thus improve the generalization performance of the model.The net-work includes three parts:feature extractor,character classifier and writer clas-sifier.The gradient reversal layer(GRL)is added between the feature extractor and the writer classifier,so that when backpropagating,the parameter of the fea-ture extractor is updated in the negative direction of the writer classifier gradient,thereby suppressing the overfitting tendency of the writing style.The recognition rate of the model is 97.27%in ICDAR2013,which is 0.39%higher than that of the model without adding style adversarial model2.Aiming at the problem that the existing network is only suitable for closed sets with predetermined categories,a siamese neural network for open set character recognition is proposed.The network learns the similarity be-tween handwritten character and character discrimination template dur-ing the training stage.In the test stage,the multi-classification task of handwritten characters is converted into multiple two-classification tasks between handwritten characters and character discrimination templates,thereby achieving character recognition on the open set.The siamese net-work has two sub-networks with different structures and do not share weights to adapt to different inputs.Proved by a series of experiments conducted on CASIA-HWDB1.0-1.2 and ICDAR2013,the accuracy of the siamese network proposed in this paper on the 3755 classification of the first-level closed set and the 3008 clas-sification of the second-level closed set is 0.41%and 1.9%higher than the state of art method.3.In order to improve the siamese network suitable for open set charac-ter recognition,a siamese neural network-soft attention alignment(SNN-SAA)based on soft attention alignment is proposed.The soft attention alignment mechanism can learn the correspondence between similar fea-tures of handwritten character and character discrimination template,so as to better measure the similarity of the two.The soft attention alignment layer calculates the attention matrix of handwritten characters and character dis-crimination templates,and adjusts the weights of the two features,so that the model learns the similarity between the same strokes.A series of experiments conducted on CASIA-HWDB1.0-1.2 and ICDAR2013 proved that the accuracy of SNN-SAA with soft attention alignment in the 3755 classification of the closed set of first-level character is 4.92%higher than the model without the addition,the ac-curacy rate on the 3008 classification of the closed set of second-level characters is increased by 7.55%.