藏文手写乌梅体古籍文献文本识别方法研究  

Research on Text Recognition Method for Tibetan Handwritten Umei Ancient Book

作　　者：扎西多吉 格桑多吉 杨宇帆 白玛旺久 Zhaxiduoji;Gesangduoji;YANG Yu-fan;Baimawangjiu(College of Information Science and Technology,Tibet Universty,Lhasa Tibet 850000,China)

机构地区：[1]西藏大学信息科学技术学院,西藏拉萨850000

出　　处：《计算机仿真》2024年第10期232-236,共5页Computer Simulation

基　　金：国家自然科学基金项目(62066042)。

摘　　要：藏文文献是藏民族传统文化的重要载体,数量在国内仅次于汉语,其中手写文献占有一定比例。文献数字化能够有效缓解藏文文献资源“藏”与“用”的矛盾,对藏文化挖掘、保护、传播、利用有重要作用和意义。目前藏文手写乌梅体文本识别研究甚少,导致文献数字化主要靠人工录入。在分析藏文手写乌梅体古籍文献特点的基础上,其首先标注8466张藏文手写乌梅体古籍文献文本行图片,同时合成多粒度文本行数据共623257张,然后针对藏文手写乌梅体古籍文献识别难点,提出一种基于VITSTR的新识别方法。经试验,方法的准确率达到81.44%,对藏文手写乌梅体文本中存在的长尾分布、相似字符多等问题有较好的表现,并且具有较好的泛化能力。Tibetan literature is an important carrier of Tibetan traditional culture.Its quantity is second only to Chinese in China,and handwrtten literature occupies a certain proportion.Document digitization can effectively alleviate the contradiction between"Tibetan"and"use"of Tibetan literature resources,and has an important role and significance in the excavation,protection,dissemination and utilization of Tibetan culture.At present,there is little research on the recognition of Tibetan handwritten Umei text,leading to the fact that the digitization of literature mainly depends on manual input.This paper first analyzes the characteristics and recognition difficulties of Tibetan handwritten Umei ancient books,and then labels 8466 text line images of Tibetan handwritten Umei ancient books.At the same time,a total of 623257 multi granularity text line data are synthesized.Finally,aiming at the recognition difficulties of Tibetan handwriten Umei ancient books,some improvements are proposed to the VITSTR method and model training.The recognition accuracy of the improved VITSTR method is 81.6%,which has a good performance on the long tail distribution and many similar characters in Tibetan handwritten mume text,and has a good generalization ability.

关 键 词：手写乌梅体 文本识别 模型训练 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]