藏文古籍手写文档文本区域检测方法研究  

Study on Detection Method of the Text Region in Tibetan Ancient Handwritten Documents

作　　者：格桑多吉 白玛旺久 扎西多吉 Gesang-Duoji;Baima-Wangjiu;Zhaxi-Duoji(College of Information Science and Technology,Tibet University,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

出　　处：《高原科学研究》2023年第3期102-111,共10页Plateau Science Research

基　　金：国家自然科学基金项目(62066042,61961038);西藏大学科研培育基金项目(ZDQMJH21-13)。

摘　　要：藏文古籍文本数字化是保护和传承藏文古籍的重要途径,由于藏文古籍具有版面复杂性、多样性和字体大小不一致等特点,传统的经典检测网络难以有效定位小目标文本区域,导致检测效果不理想。文章选择适用于小目标的检测算法YOLOv8基线模型,采用改进后的轻量级骨干网络Faster-Net作为特征提取网络,引入CA注意力机制以更好地利用上下文特征信息,将浅层信息和深层信息有效地结合在一起。实验结果表明,该方法可实现对藏文古籍多字体手写体版面不同字体大小的文本区域定位,其在测试集上的平均精度、准确率和召回率都达到99%以上,适用于藏文古籍版面中字体大小不一致的文本区域目标检测。Digitization of Tibetan ancient documents is an important way to protect and inherit Tibetan ancient documents,however,due to the complexity,diversity,and inconsistency of font size of the text in Tibetan ancient documents,it is difficult to effectively locate small text areas using the traditional classical detection network,resulting in unsatisfactory detection of text.In this paper,selecting the detection algorithm named YOLOv8 baseline model tailored to small targets,using the improved lightweight backbone network Faster-Net as the feature extraction network,and introducing CA attention mechanism for better utilizing the context feature information,shallow information is combined effectively with deep information.Experimental results show that this method can locate text regions with different font sizes and multi-font handwriting layouts of ancient Tibetan documents,and the average accuracy,accuracy,and recall rate of our test group are more than 99%,indicating it is suitable for detection of text areas with inconsistent font sizes in the layout of ancient Tibetan documents.

关 键 词：藏文古籍 注意力机制 Faster-Net 手写多字体 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]