基于图卷积网络的任意形状藏文文本检测方法研究  

Study on Arbitrary-Shape Tibetan Text Detection Method Based on Graph Convolutional Network

作　　者：珠杰 许泽洲 Zhujie;XU Zezhou(School of Information Science and Technology,Tibet University,Lhasa 850000,China;Provincial and Ministerial Collaborative Innovation Center for Tibetan Informatization,Lhasa 850000,China)

机构地区：[1]西藏大学信息科学技术学院,西藏拉萨850000 [2]西藏信息化省部共建协同创新中心,西藏拉萨850000

出　　处：《高原科学研究》2023年第3期94-101,共8页Plateau Science Research

基　　金：国家自然科学基金项目(62066042);教育部人文社会科学研究项目(21YJCZH059);西藏大学培育计划项目(ZDCZJH21-10);西藏大学珠峰学科建设计划项目(zf22002001)。

摘　　要：在文本检测领域,多形状文本检测一直是一个棘手的问题,对于藏文文本检测来说,这更是一个亟待解决的问题。文章构建了基于CC-Based思想的任意形状藏文文本检测模型,该模型首先使用卷积神经网络来获得文本构件预测,然后使用图卷积网络来进行关系推理。模型在藏文自然场景数据集与现代藏文书籍数据集上都取得了优良的实验结果。Detection of multi-shape text has been a challenging issue in the field of text detection,and this is also an urgent problem that needs to be solved for Tibetan text detection.In this paper,an arbitrary Tibetan text detection model based on the CC-Based is constructed,which first uses convolutional neural networks to obtain text component prediction,and then uses graph convolutional networks to perform relational inference.Our experimental results show that the model performed excellently on the text dataset of the Tibetan natural scene and the modern Tibetan books.

关 键 词：文本检测 任意形状 藏文 图卷积网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]