Research On Text Detection Of Tibetan Ancient Books Based On Deep Learning

The rapid development of deep learning has attracted the attention of many researchers,and as the most direct way of communication and information acquisition,words are indispensable in our daily life.Due to the slow development of information technology in Tibetan areas,less research on detection and recognition of Tibetan language,the overall structure of Tibetan language is complex,which is quite different from that of Chinese and English,so the mature technology of detection and recognition in Chinese and English cannot be fully used.It is necessary to improve the detection and recognition technology in Chinese and English according to the characteristics of Tibetan language,and design a more suitable detection and recognition system for Tibetan language.This dissertation focuses on the detection of the text of Tibetan ancient books based on deep learning,and flexibly uses neural network to realize the positioning of the text line of Tibetan ancient books.The algorithm of text detection in natural scene is that the Tibetan ancient books and documents have similar text characteristics to those in natural scene,with many external interferences and complex background.This dissertation mainly uses two different models to test the Tibetan ancient books.The specific work is as follows:(1)This dissertation analyzes the data set of Tibetan ancient books applied to the college,increases the number of data sets by using GANdata expansion,and establishes a data set of Tibetan ancient books,which contains 3396 pictures in total.The pictures are labeled and analyzed for the training of detection methods of Tibetan ancient books.(2)In this dissertation,the text detection method of Tibetan ancient books based on CTPN model is adopted.CTPN is a detection algorithm based on the combination of CNN and RNN.It uses VGG16 to extract image features,BiLSTM to learn character sequence features,and finally uses NMS to refine the predicted text box,and uses text line construction method to merge the predicted text box into a whole text box.Experimental results show that the CTPN model of Tibetan ancient literature text detection has achieved good performance,the accuracy of the algorithm is 0.89.(3)In this dissertation,EAST model is used to detect the text of Tibetan ancient books.EAST is a detection algorithm based on the combination of FCN and NMS.FCN is used to generate text box prediction directly,LNMS is used to delete redundant text boxes,unnecessary intermediate steps are deleted,and end-to-end training and optimization are carried out.However,the accuracy of text detection in Tibetan ancient books is not high,which needs further study.