The Recognition System Of Uchen Tibetan Historical Documents Based On Deep Learning

There are a large number of Tibetan ancient books,involving a wide range of contents,which are precious cultural heritage in China.However,due to the age,long preservation time,and improper preservation methods,the paper of ancient books is seriously degraded and cannot be read repeatedly by researchers.In order to better study the ancient books,it is urgent to organize the text of the ancient Tibetan books and store the text information as a Unicode file.This paper makes a detailed study on layout analysis,text line segmentation and text line recognition of Tibetan ancient books in Uchen by using related technologies of document analysis and recognition.The main contents are as follows:(1)An image layout analysis set containing 212 Tibetan historical documents in Uchen was constructed by manual annotation.In addition,by analyzing the word construction rules of Tibetan combined with the semantic information of ancient books and the baseline fluctuation information of text lines,a text line synthesis scheme of Uchen Tibetan historical documents was proposed,and a data set containing 3.2 million Tibetan text line images was generated.(2)A layout analysis model of Uchen Tibetan ancient books based on multi-task branch is proposed to solve the problem of layout segmentation.On the other hand,the model can complete different levels of layout segmentation tasks by adding different task branches,thereby completing more complex layout segmentation.The final precision of binarization and layout segmentation was 92.53% and 91.69%,respectively.(3)A text line segmentation method based on local baseline and connected component centroid is proposed.Firstly,the text line area is detected,and the local baseline position is obtained;secondly,the position relationship between the connected component and the local baseline as well as a neural network are used to complete touching connected component detection,and the watershed algorithm is used to segment touching connected components.Finally,the broken strokes are assigned to the text lines in which they belong according to the characteristics of Tibetan character structures.This method can effectively reduce the influence of text line distortion and skew on text line segmentation,and has high robustness.This method effectively solves the problem of text line segmentation of Uchen Tibetan historical documents,the precision was 99.48% in 322 document image line segmentation experiments.(4)An end-to-end recognition model of Uchen Tibetan historical documents is proposed,which can realize the whole line recognition of text line image.The method avoids the difficulty of character segmentation,so the difficulty of character recognition is greatly reduced,it achieves 92.79% accuracy in 640 thousand synthetic text line document images.According to the proposed algorithm,the design and implementation of the Uchen Tibetan ancient book recognition system has been completed.The test results show that the system meets the requirements for image recognition of Tibetan historical documents.