摘要
针对与日俱增的隧道养护需求,为了节约时间与人力成本,提出基于卷积神经网络的公路隧道衬砌病害检测方法.利用自主研制的隧道智能快速检测车采集24条隧道衬砌的图像,构建超过20 000张病害图像的高质量数据集.结合隧道衬砌病害的成因及特点,分别构建单阶段SSD模型和两阶段R-FCN模型在自制的数据集上训练,对检测结果进行对比分析,提出离线式隧道衬砌病害检测方案.试验结果表明,SSD模型的识别准确率为98%,总的平均精度均值(mAP)为72%,检测速度较快,适用于隧道的快速诊断. R-FCN模型的识别准确率为85%,总的m AP达到91%,检测精度较高,适用于隧道病害的后期处理.利用这2种检测模型均可以提升检测效率和精度.
A highway tunnel lining disease detection method based on convolutional neural network was proposed aiming at the ever-increasing demand for tunnel maintenance in order to save time and labor costs. The selfdeveloped intelligent rapid detection vehicle for tunnels was used to collect 24 tunnel lining images. A high-quality data set of more than 20 000 disease images was constructed. Then single-stage SSD(single shot multiBox detector)models and two-stage R-FCN(region-based fully convolutional networks) models were constructed on a self-made data set combining the causes and characteristics of tunnel lining diseases. The detection results were compared and analyzed, and an offline tunnel lining disease detection scheme was proposed. The experimental results showed that the identification accuracy rate of SSD model was 98%, the total mean average precision(mAP) was 72%, and the detection speed was fast. The SSD model is suitable for rapid diagnosis of tunnels. The identification accuracy rate of R-FCN model was 85%, the total mAP value reached 91%, and the detection accuracy was high. The R-FCN model is suitable for the post-treatment of tunnel diseases. Using these two detection models can improve detection efficiency and accuracy.
作者
任松
朱倩雯
涂歆玥
邓超
王小书
REN Song;ZHU Qian-wen;TU Xin-yue;DENG Chao;WANG Xiao-shu(State Key Laboratory of Coal Mine Disaster Dynamics and Control,Chongqing University,Chongqing 400044,China)
出处
《浙江大学学报（工学版）》
EI
CAS
CSCD
北大核心
2022年第1期92-99,共8页
Journal of Zhejiang University：Engineering Science
基金
国家自然科学基金资助项目(51774057,52074048)
重庆市研究生科研创新资助项目(CYS18020)。
关键词
深度学习
卷积神经网络
隧道衬砌
隧道病害检测
裂缝
渗漏水
deep learning
convolutional neural network
tunnel lining
tunnel disease detection
crack
leakage
