一种基于关键信息监督的隧道雷达数据衬线识别算法

An Algorithm for Liner Identification of Tunnel Radar Data Based on Critical Information Supervision

【目的】探索探地雷达数据解析和基建病理检测新的处理方法。探地雷达作为桥梁和隧道缺陷检测中常用无损技术手段,一直面临数据解析困难问题,提高解析结果的准确性,对交通基础设施的缺陷检测具有重大应用价值。【方法】将衬砌结构作为识别对象,分解为关键点和曲线来表示。关键点检测基于双热图方法,借助“软标注”来加快模型收敛。曲线拟合模块通过神经网络回归拟合,加入对抗扰动机制,抗图像噪声干扰。【结果】结果表明,该算法识别的衬砌线较真实偏移量为2.23个像素点,较CenterNet网络提升1.24个像素点,较CornerNet网络提升0.71个像素点。【结论】解析识别效果提升显著,具有较高应用价值。

[Objective]This paper mainly explores new processing methods for ground-penetrating radar data parsing and infrastructure pathology detection.The current ground-penetrating radar,as a common non-destructive technical tool in bridge and tunnel defect detection,is facing the problem of difficult data parsing.Improving the accuracy of the parsing results has significant application value to the defect detection of transportation infrastructure.[Methods]The lining structure is represented as an identification object,decomposed into key points and curves.Key point detection is based on a bipartite heat map approach,with the help of"soft annotation"to speed up model convergence.The curve fitting module is implemented by neural network regression,incorporating a counteracting perturbation mechanism to resist image noise interference.[Results]The results show that the algorithm identifies a liner offset of 2.23-pixel points from the true one,a 1.24-pixel point improvement over the CenterNet network and a 0.71-pixel point improvement over the CornerNet network.[Conclusions]The proposed method has significantly improved resolution recognition results and has obvious application value.