基于分割方法的繁体中文报纸文本检测  

Segmentation-based detector for traditional Chinese newspaper

作　　者：姜宇[1,2] 潘家铮 陈何淮 符凌智 齐红[1,2] JIANG Yu;PAN Jia-zheng;CHEN He-huai;FU Ling-zhi;QI Hong(College of Computer Science and Technology,Jilin University,Changchun 130012,China;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012,China)

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]吉林大学符号计算与知识工程教育部重点实验室,长春130012

出　　处：《吉林大学学报（工学版）》2023年第4期1146-1154,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(62072211,U20A20285)。

摘　　要：现阶段文本检测的研究主要面向自然场景数据集进行,针对繁体中文图像内嵌文本场景难以有效检测的问题,本文提出了一个基于分割方法的繁体中文报纸图像文本检测模型。该模型使用Resnet50和FPN作为特征提取网络,采用分割实例缩放加扩展算法的方法生成用于预测文本框的二值图,并提出了周围填补、循环检测加区域覆盖的方法增强检测效果。该模型针对自建繁体中文报纸数据集的实验结果的三项指标均在0.9左右,且相比于目前文本检测效果较好的DBNet的实验结果均提升了5%~7%,针对繁体中文报纸图像文本检测任务具有一定的优越性。Most of the research on text detection has been conducted on natural scene datasets,few on such specific scene.And the existing models are not good enough for text detection on traditional Chinese newspaper.In order to solve this problem,a segmentation-based text detector for traditional Chinese newspaper is proposed in this paper.The model uses Resnet50 and FPN as feature extraction network,employing a segmentation instance scaling and extension algorithm to generate the binary map for predicting text boxes.And the methods of surrounding filling and loop detection plus region coverage are proposed to enhance the detection effect.In addition,a traditional Chinese newspaper dataset is built to satisfy the research needs.The experimental results of this model on traditional Chinese newspaper dataset are around 0.9 and are improved by 5%to 7%compared with DBNet,which indicates that the model is effective and accurate for text detection on traditional Chinese newspaper.

关 键 词：计算机应用 特定场景 繁体中文报纸 图像内嵌文本 文本检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]