基于深度学习的政务表格单元格结构检测

Deep learning based government affairs table cell structure detection

当前政务领域中涵盖大量复杂异构表格,其结构检测困难,识别准确率较低并且单元格边缘拟合效果较差.针对该问题,在掩膜区域卷积神经网络(Mask R-CNN)的基础上,以政务表格单元格结构为对象,提出一种改进的政务表格单元格结构检测方法.首先,在Mask R-CNN算法的主干网络及特征金字塔中加入通道注意力机制,增强有效特征通道权重;然后,对分割产生的掩膜结果使用基于规则和形态学方法进行优化以提升单元格分割边缘拟合度.实验结果表明:改进后的表格单元格结构检测模型在此数据集G-Tab及公开表格数据集ICDAR2013上的精确率和召回率都有明显提升,能够验证改进模型的有效性.

The current government affairs field contains a large number of complex heterogeneous tables,its structure detection is difficult,the recognition accuracy is low,and the table cell edge fitting effect is poor.To solve this problem,based on the mask region convolutional neural network(Mask R-CNN),and taking the cell structure of government affairs table as the object,an improved method for recognizing the structure of government affairs table cell is proposed.First,a channel attention mechanism is added to the backbone network and feature pyramid of the Mask R-CNN algorithm to enhance the effective feature channel weights.Then,the rule-based and morphological processing method is used to optimize the mask results generated by the segmentation to improve the edge fit of the cell segmentation.Experimental results show that the precision and recall of the improved table cell structure recognition model on the G-Tab dataset and the public table dataset ICDAR2013 are significantly improved.These can verify the effectiveness of the improved model.