Researchand Designof Layout Analysis Systemfor Tibetan Historical Document Images

With the development of the country and society,and driven by big data,cloud computing and other technologies,the digital protection of cultural resources has been paid more and more attention,and the related research has become a hot spot at home and abroad.In recent years,researchers are more and more interested in the digitization of ancient books.Layout analysis,as an important basic step in the process of digitization of historical documents,is also a significant part of OCR system.It divides the document image into different parts according to certain characteristics,and judges that each part is text,title,image,graph or table.Different processing methods will be used in the subsequent processing of each area obtained from layout analysis.For example,the text area will be recognized as internal code text by character recognizer,and the table area will be processed by special table recognizer.Aiming at the layout analysis of Tibetan ancient books,taking the images of Uchen Tripitaka of Beijing edition and Lijiang edition as examples,divides the layout analysis of Tibetan ancient books into four stages: layout segmentation,text line recognition,layout description,and layout restoration,on this basis,design and implement a layout analysis system for Tibetan ancient book images.The main research work is as follows.(1)Based on the semantic segmentation network Deep Lab,training Tibetan ancient book images layout segmentation model.Build a Tibetan ancient book images layout segmentation data set,named THDID-LS.This data set is manually marked by the image annotation software labelme.The Tibetan ancient book images layout elements are divided into five categories: background,text,left title,right title and image.Different colors are used to distinguish the categories and achieve the marking.Training and testing the Tibetan ancient book images layout segmentation model on the data set THDID-LS,the accuracy can reach 90.92%.(2)Based on the text recognition network CRNN,the end-to-end recognizer of text lines for Tibetan ancient book is trained.Firstly,a text line data set of ancient Tibetan books is generated,named TTLDS-G.the data set includes five popular Uchen fonts.Background samples are taken from real ancient Tibetan books,and a total of 1.15 million simulation samples with normal,slanting,blurring,smudging and damage effects are generated.The real data set is labeled and named TTLDS-R,which is used to improve the generalization of the model in the second training.Training and testing on TTLDS-G and TTLDS-R data sets,the comprehensive recognition rate is 85.41%.(3)Propose a method for the layout description and restoration of Tibetan ancient book.According to the location information of text,left title,right title and graph after the segmentation of Tibetan ancient book image,combined with the recognition results of Tibetan ancient book text line,the corresponding data structure is defined,and the layout description of Tibetan ancient book image is carried out,so as to generate the retrievable page which is consistent with the original Tibetan ancient book layout style,higher layout quality and text code representation.(4)Design and implement a Tibetan ancient book images layout analysis system.The system integrates layout segmentation,text lines recognition,layout description and restoration of Tibetan ancient book images,and provides two analysis methods,one click analysis and step-by-step analysis.It is a web platform for displaying the effect of layout analysis for Tibetan ancient book images.