基于Mask R-CNN的满文文档版面分析  

Manchu Document Layout Analysis Based on Mask R-CNN

作　　者：陈璇 贺建军 李厚杰 武林秀 CHEN Xuan;HE Jian-jun;LI Hou-jie;WU Lin-xiu(School of Information and Communication Engineering, Dalian Minzu University, Dalian Liaoning 116605,China)

机构地区：[1]大连民族大学信息与通信工程学院

出　　处：《大连民族大学学报》2019年第3期240-245,共6页Journal of Dalian Minzu University

基　　金：国家科技支撑计划项目(2012BAJ18B06);国家民委科研项目(12DLZ011);辽宁省自然科学基金项目(20180550625);辽宁省教育厅科学研究一般项目(L2014540);中央高校基本科研业务费专项资金资助项目(DC110313,DC120101073)

摘　　要：提出一种基于Mask R-CNN深度学习框架的满文文档版面分析方法,将满文文档版面分析问题转化为基于深度学习的图像实例分割问题。使用ResNet101网络和FPN网络构成的卷积神经网络自动提取满文文档图像特征,特征图经过RPN网络和RoI Align层生成新的特征图。新特征图经过全连接层完成感兴趣区域的类别和边框预测,并经过全卷积神经网络对感兴趣区域的像素进行分类得到mask预测,最终实现满文文档图像的实例分割。通过《新满汉大辞典》的文档图像构建了满文文档图像数据集,算法在该满文文档图像数据集上进行了实验。实验结果表明,本算法在满文文档版面分析中取得了较好的检测和分割效果。Manchu document layout analysis method based on Mask R-CNN deep learning framework was proposed in this paper, which turned the Manchu document layout analysis problem into an image instance segmentation problem based on deep learning. The convolutional neural networks which were composed of ResNet101 network and FPN network extracted Manchu document image features automatically. The feature map was sent to the RPN network and RoI Align layer to generate a new feature map. Then, the fully connected layer was used to complete the classification and the bounding box regression. At the same time, the full convolutional neural network was used to complete the pixel classification of the regions of interest to obtain the mask prediction. Finally, the instance segmentation of the Manchu document image was realized. In this paper, the Manchu document image dataset was constructed by the document images of New Manchu Dictionary , and the algorithm was performed on this dataset. Experimental results show that the algorithm can achieve effective detection and segmentation in Manchu document layout analysis.

关 键 词：满文文档 版面分析 实例分割 MASK R-CNN 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]