基于深度学习的桥梁设计图纸中钢筋数量信息归集方法

Methods of collecting rebar quantities information in bridge design drawings based on deep learning

桥梁设计图纸中含有重要的钢筋数量信息,施工过程中需要利用钢筋数量信息进行钢筋的分配工作,同时这些信息可以在新项目设计时提供设计经验,然而这些钢筋数量信息通常以表格形式保存在纸质图纸中,对查阅和参照造成了不便。基于此,提出了一种从桥梁设计图纸中归集钢筋数量的方法,首先在YOLOv7的骨干网络中,融入卷积注意力模块(CBAM)注意力机制获取更多细节特征,在头部结构中引入多分支卷积RFB模块,提升模型对小目标的特征表达能力,基于改进后的YOLOv7对桥梁设计图纸的表格和图签栏进行目标检测;其次利用PP-StructureV2的表格识别功能对图纸中的表格和图签栏进行表格识别,经过汇总验证,最终生成包含所有钢筋数量信息的Excel文档,实现桥梁设计图纸中钢筋数量信息的归集。利用安徽省交通规划设计研究总院股份有限公司提供的桥梁设计图纸进行实验研究,实验结果表明,改进的YOLOv7算法的F1 Score可达到98.35%,和原始YOLOv7算法相比,提升了0.86%,可满足从桥梁设计图纸中检测表格以及图签栏的要求。

There exists rebar quantity information in bridge design drawings,which is used in the construction process for rebar allocation work,and at the same time,this information can provide design experience in the design of new projects,however,this rebar quantity information is usually saved in paper drawings in the form of tables,which is inconvenient for access and reference.Based on this,a method is proposed to summarize the rebar quantity from bridge design drawings.Firstly,in the backbone network of YOLOv7,the convolutional attention module(CBAM)attention mechanism is incorporated to obtain more detailed features,and the multi-branch convolutional RFB module is introduced in the head structure to improve the feature expression ability of the model for small targets,object detection of normal tables and title block of bridge design drawings based on improved YOLOv7;secondly,the table recognition function provided by PP-StructureV2 is used to identify the norma tables and title block in the drawings;after summarization and validation,an Excel table containing all the rebar quantity information is finally generated to realize the summarization of rebar quantity information in bridge design drawings.Experimental studies were conducted using the bridge design drawings provided by Anhui Transport Consuling&Design Institute Company Limited.The experimental results show that the F1 Score of the improved YOLOv7 algorithm provided in this paper can reach 98.35%,which is 0.86%higher compared with the original YOLOv7 algorithm,and it can meet the requirements of detecting the common table as well as the title block table from the bridge design drawings.