摘要
在核电企业数字化转型过程中,人工识别图纸误差较高,易造成企业损失,有必要利用自动化提取表格信息技术代替人工.表格结构识别是表格信息提取的关键技术,但核电施工图纸表格结构复杂且样本少,影响了识别效果.针对这一问题,提出了结合高效注意力机制的多尺度扩展模型EPNet,引入了渐进式尺度扩展模块,增强了有效特征通道权重,实现了少样本的有效特征信息获取.另外,利用局部特征中的文本区域和全局特征中单元关系的视觉信息来获得可靠的单元格边界,提高边缘拟合的精细度.实验结果表明,该模型在识别核电施工图纸中的表格单元格结构方面表现出色,与Mask R-CNN模型相比,精确度提高了1%,F1值提高了3%,具有较高的准确性和鲁棒性.
Abstract
Table structure recognition is a key technology for table information extraction.The structure of the table of nuclear power construction drawings is complex and the sample is small,which affects the recognition effect.To solve this problem,EPNet,a multi-scale extension module combined with efficient attention,is proposed,and a progressive scale expansion module is introduced to enhance the weight of effective feature channels and realize the acquisition of effective feature information with fewer samples.The visual information of the text area in the local feature and the cell relationship in the global feature is used to obtain reliable cell boundaries and improve the fineness of edge fitting.Experimental results show that the proposed model performs well in identifying the table cell structure in nuclear power drawings,and the accuracy is improved by 1%and the F1 value is increased by 3%compared with the previous algorithm,which has high accuracy and robustness.
维普
万方数据