Research On The Recognition Of Tangut Character Based On Deep Learning

As a kind of writing recording the history of the Western Xia Dynasty,Tangut characters have very unique cultural connotations.The recognition of Tangut characters is of great significance to the whole Xixia study.Up to now,the problem of low recognition rate of Tangut characters has limited the development of Xixia studies.Compared with widely used characters,such as Chinese characters,there are more difficulties in the recognition of handwritten Tangut characters in ancient literature.it mainly includes:the structure of Tangut characters is more complex than Chinese characters,more character strokes,higher similarity between Tangut characters;The training sample set of handwritten Tangut characters is limited to the source of samples,and there is a problem that the number of samples of some character categories is scarce.Based on this,this paper combines the characteristics of deep convolution neural network to extract more surface information of complex Tangut characters,the dynamic routing algorithm of CapsNet network makes attention to the position relationship between strokes of Tangut characters,and builds an A-CapsNet depth recognition network for the recognition of Tangut characters with a small number of samples.The main contents of this paper are as follows:1.According to the characteristics of Xixia text sample image,such as different size,uneven proportion of text region and much background noise information,an improved MSER algorithm is proposed to extract the text region of Xixia text sample image,filter out the redundant background and noise information,and normalize it.2.The CapsNet network based on the classical deep convolution neural network classification model and dynamic routing algorithm is used to train and recognize the preprocessed Tangut characters.The experimental results show that the network structure is suitable for Tangut characters recognition.3.On the basis of CapsNet network,the A-CapsNet network structure is constructed to complete the recognition of Tangut characters.On the basis of the same data set,experiments are carried out on the recognition effect from the aspects of the number of dynamic routing iterations,training rounds and reconstruction effect.The final experimental results show that the A-CapsNet network has higher recognition accuracy and better applicability to Tangut characters.