Abstract: To solve the problems of low efficiency and poor accuracy in manually evaluating the similarity of complex process tables for process reuse design,a graph neural network combination algorithm was proposed to effectively extract the structure and semantics of process tables to measure similarity.An improved Mask Region-based Convolution Neural Networks (Mask R-CNN) algorithm was proposed for table detection,which included highlighting table features with distance transformation,improving detection accuracy by Confluence algorithm and achieving precise position by corner positioning and adjustment of detection frame.Meanwhile,the Optical Character Recognition (OCR) technology was used to extract table text information.According to the extracted key unit information,the structural characteristic graph network and the semantic relation graph network of the process table were respectively modeled.Furthermore,a graph neural network combination algorithm was proposed to extract the structural features and node attributes of the graph network model,and convert them into low-dimensional real-valued vectors to drive the proposed comprehensive evaluation method of joint similarity to realize the semantic similarity of the measurement process table.The experimental analysis showed the effectiveness of the proposed method,and the feasibility of the method was verified by an example of process reuse.

Key words: table detection, process table modeling, graph neural network, similarity measure

摘要： 为解决人工评估复杂工艺表格的相似性用于工艺重用设计存在效率低、精度差等问题,提出一种图神经网络组合算法,以有效提取工艺表格的结构、语义等特征以度量相似性。首先提出改进Mask R-CNN算法用以进行表格检测,包括距离变换突出表格特征、Confluence算法提高检测精度、角点定位调整检测框以实现精准定位,同时利用光学字符识别(OCR)技术提取表格文本信息。然后,针对提取的关键单元信息,分别建模工艺表格的结构特性图网络与语义关系图网络。进一步,提出图神经网络组合算法提取图网络模型的结构特征与节点属性,并转化成低维实值向量,以支撑提出的一种联合相似度综合评估方法,实现度量工艺表格语义相似性。最后,经实验分析表明了所提方法的有效性,并以工艺重用实例验证了方法的可行性。

关键词: 表格检测, 工艺表格建模, 图神经网络, 相似性度量