基于高分辨特征的隧道病害检测算法研究

Research on Tunnel Disease Detection Algorithm Based on High-resolution Features

现阶段,自动化隧道病害检测的准确率往往受限于隧道内部的成像条件与检测车的采集设备,波动较大。本文提出一种基于HR(High Resolution)-Net的隧道病害检测算法HR-Net(1se),通过使用高分辨率特征可在成像条件较差的隧道洞壁图像中,更准确地提取隧道洞壁病害,且具有较高的鲁棒性。为提升算法运行效率,本算法使用HR-Net的第一阶段的特征提取网络作为主干网络,并使用该特征进行病害检测。通过在隧道病害图像数据集上验证,本文算法的mIoU指标达到81.74%,高于其他对比算法,并且耗时低于原始的HR-Net约25%。

At the present stage,the accuracy of automated tunnel disease detection is often limited by the imaging conditions within the tunnel and the acquisition equipment of the inspection vehicle,which fluctuates greatly.This paper proposes a tunnel disease detection algorithm HR-Net(1se)based on High Resolution(HR)-Net,which can extract tunnel wall diseases more accurately and exhibit higher robustness in tunnel wall images with poor imaging conditions by using high-resolution features.To improve the efficiency of the algorithm operation,this algorithm uses the feature extraction network of the first stage of HR-Net as the backbone network,and uses the features for disease detection.Through a verification on the tunnel disease image dataset,the mIoU index of algorithm in this paper achieves 81.74%,which is higher than other comparative algorithms,and takes about 25%less time than the original HR-Net.