基于级联R-FCN的文档图像表格检测方法  

Table Detection Method in Document Image Based on R-FCN

作　　者：宋颖 易尧华[1] 汤梓伟 卢利琼[1] SONG Ying;YI Yao-hua;TANG Zi-wei;LU Li-qiong(School of Printing and Packaging,Wuhan University,Wuhan 430072,China)

机构地区：[1]武汉大学印刷与包装系,武汉430072

出　　处：《数字印刷》2020年第2期50-57,共8页Digital Printing

基　　金：国家科技重大专项(No.2017ZX01030102)。

摘　　要：表格检测是文档分析中的非文本内容检测部分的重要任务,表格检测的高准确率是提高文本检测准确性的必要条件。本研究提出了一种基于深度学习的文档图像分析的表格检测方法。该方法采用级联R-FCN(基于区域的全卷积网络)框架,首先检测出文档图像的公式区域并移除;然后在无公式的文档图像中,检测提取表格与图区域,最后通过参数调节筛选出最终的文档图像表格区域。该方法在ICDAR 2017 Competitionon Page Object Detection数据集上IoU(交叉重合区域)为0.8时,AP值和F1值相应为0.851和0.898。实验结果表明,该方法与传统的基于形态学变换和水平垂直投影的方法相比,可以简单而高效地检测文档图像中的表格。Table detection is an important task for non-text area detection in document images. A good table detection performance is a prerequisite for improving text detection performance. In this study, a table detection method in document image analysis based on deep learning was proposed. The cascade R-FCN deep learning framework was used. Firstly, the formula region of document image was detected and removed. Then, the table and figure regions in document images were detected and acquired. Finally, the table regions from document images was filtered by parameter adjustment. The AP and F1 values were 0.851 and 0.898 where Io U was 0.8. The test result showed that compared to the traditional morphological transformation, horizontal and vertical projection-based method, the proposed method is able to detect tables in document images in a simple but effective way.

关 键 词：文档图像 表格检测 级联R-FCN 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]