融合图卷积网络的表格图像结构识别模型  

Table image structure recognition model based on graph convolutional network

作　　者：孙俊 苟刚[1,2] SUN Jun;GOU Gang(State Key Laboratory of Public Big Data,Guizhou University,Guiyang 550025,China;School of Computer Science and Technology,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学公共大数据国家重点实验室,贵州贵阳550025 [2]贵州大学计算机科学与技术学院,贵州贵阳550025

出　　处：《计算机工程与设计》2024年第10期3066-3073,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(62162010);贵州省科技支撑计划基金项目(黔科合支撑[2022]一般267)。

摘　　要：针对表格图像很难精确从文档中提取出表格结构的问题,提出一种融合图卷积网络的双分支识别网络模型。以ResNet+FPN为主干网络,引入矩阵分解头代替注意力机制重整全局特征。设计一个双分支网络以获取表格单元格间空间位置和逻辑邻接信息。以GCN感知单元格间连接关系辅助输出位置信息和逻辑邻接关系。实验结果表明,在多个数据集上相比基线模型F1指标平均提升10.6%,F(beta=0.5)指标提升18.6%。在TableGraph-24K数据集上,相比最近的TGRNet模型在F1指标上提升3.1%,F(beta=0.5)指标平均提升2.9%。In view of the problem that it is difficult to accurately extract the table structure from the document,a two-branch reco-gnition network model integrating the graph convolutional network was proposed.ResNet+FPN was taken as the main stem network,and the matrix decomposition head was introduced instead of the attention mechanism to renormute the global features.A two-branch network was designed to obtain spatial location and logical adjacency information.Position information and logical adjacency were output by GCN sensing cell connection relationship.Experimental results show that compared with the baseline model,the F1 index is increased by 10.6%,and the F(beta=0.5)index is increased by 18.6%.In the TableGraph-24K data set,compared with the recent TGRNet model,the F1 index is improved by 3.1%,and the F(beta=0.5)index is improved by 2.9%on average.

关 键 词：图像处理 表格图像结构识别 图卷积网络 特征融合 注意力机制 矩阵分解 双分支网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]