基于行列信息门的表格结构识别网络研究

Research on Table Structure Recognition Network Based on Row and Column Information Gates

为解决检测表格行与列法容错率大的问题,本文提出了行信息门、列信息门两种信息传输模块,采用特征切片与平铺相结合的方式,对行或列信息门进行预测,保证简化行与列预测结果的精确性。运用行列两种信息门,完成对语义分割模型的构建,实现对表格行与列的精确化分割。此外,基于ICDAR格式数据集,完成对表格行与列掩模的有效构建,并评估该模型性能。结果表明,与基于特征金字塔网络的分割模型相比,本文提出的语义分割网络模型具有较高平均查准率、查全率和F1。其中,平均查准率超过0.55%,查全率超出2.78%,F1超出1.48%。希望本研究可以为相关人员提供有益借鉴和参考。

To address the issue of high fault tolerance in detecting table rows and columns,this paper proposes two information transmission modules:row information gate and column information gate.By combining feature slicing and tiling,the row or column information gates are predicted to ensure the accuracy of simplified row and column prediction results.By utilizing both row and column information gates,the construction of a semantic segmentation model is completed to achieve precise segmentation of table rows and columns.Additionally,based on the ICDAR format dataset,the effective construction of table row and column masks is completed,and the performance of the model is evaluated.The results show that compared with the segmentation model based on the feature pyramid network,the semantic segmentation network model proposed in this paper has higher average precision,recall,and F1 value.Specifically,the average precision exceeds 0.55%,the recall rate exceeds 2.78%,and the F1 value exceeds 1.48%.It is hoped that this research can provide effective reference and inspiration for related personnel.