Research On Xixia Character Detection Based On KR-CTPN

Xixia script is a kind of script used by the Tangut people in ancient China.It has a history of thousands of years.The Xixia literature composed of Xixia script has important historical significance and research value for the study of the history and culture of China’s ethnic minorities.At present,there are a large number and various types of documents in Xixia script,but most of them are not preserved completely.Due to objective reasons,the interpretation and arrangement methods for Xixia script and documents lag behind the traditional ones.However,with the development of artificial intelligence—especially the convolutional neural network and the development of deep learning,it provides a more efficient way for Xixia text detection and document retrieval,and also provides new technical support for creating a digital knowledge base of Xixia literature.At present,there are problems such as high font similarity,messy order,blurred fonts,etc.in the existing ancient Xixia script documents.Obviously,traditional text detection methods cannot detect Xixia texts accurately and quickly,so this paper uses the self-made Xixia script dataset as the basis,and the expanded data set was manually labeled,and the deep learning algorithm was used to improve the CTPN network model for Xixia text detection research.A large number of data experiments show the effectiveness of the detection method proposed in this paper.The main research work of this paper is as follows:1.Expansion and establishment of Xixia language dataset.Firstly,the basic Xixia dataset was formed by scanning the ancient books of Xixia and other methods.It is proposed to use GANs and CycleGAN to expand the dataset of Xixia characters,and use VGG-16 and ResNet to classify the generated text images.By detecting GANs and CycleGAN in For the expansion effect on the Xixia text dataset,it was decided to use the improved CycleGAN data generation model specific to the Xixia text dataset as the main model for data set expansion.In the end,six rounds of experiments were conducted to obtain a Xixia text dataset containing 9,300 images.2.CTPN model may not perform well for low-resolution images,and ResNet network is prone to bottleneck when processing high-resolution images due to the small number of channels in the convolutional layer,which will also lead to a series of problems such as reduced network performance.In this paper,a Resnet-based KResNeSt Network is designed to replace VGG-16 as the backbone of CTPN(Connection Text Proposal Network).First,the performance of ResNet has been improved with the introduction of the Kaming initialization and BatchNormalization layers,giving it greater generalization capabilities.Secondly,we add an Attention layer in front of the Bi-LSTM layer to improve the accuracy of the model with weights.Finally,after many experiments,we found that compared with VGG-16,KResNeSt has a deeper network structure,and KResNeSt can contain hundreds or even thousands of layers,which improves the expression ability of the model and thus improves the accuracy of the model.KResNeSt also has a faster convergence rate.The initialization of kaiming makes the weight initialization of the model more reasonable,thus speeding up the convergence of the model.By replacing the VGG-16 network in the CTPN model and using the KResNeSt network,its performance has been improved to some extent.3.Design a Xixia character detection system based on KR-CTPN.The system uses the BootStrap+java framework to realize a text detection function for the input image and visualize it.