基于声呐成像的水下桩墩表观病害深度学习与智能检测  

Deep learning and intelligent detection of apparent defects on underwater foundations of bridges with sonar imaging

作　　者：骆剑彬[1] 姜绍飞[1] 沈圣[1] 陈建腾 Luo Jianbin;Jiang Shaofei;Sheri Sheng;Chen Jianteng(Fuzhou University,Fuzhou 350116,China)

机构地区：[1]福州大学,福建福州350116

出　　处：《土木工程学报》2021年第7期90-100,共11页China Civil Engineering Journal

摘　　要：声呐成像检测水下桩墩表观病害时,其图像与光学图像的病害特征存在较大差异,病害的位置和类型需要人工识别且易出错。为解决这个问题,提出基于声呐成像的水下桩墩表观病害深度学习与智能检测方法。首先对水下实桥桩墩以及试验模型进行声呐扫描获取大量图像,并分析声呐图像中的病害特征;然后对Faster R-CNN框架下VGG16网络模型进行改进,采用水平、垂直等线性变换实现原始声呐图像的数据增强,对深度学习模型进行近似联合优化训练,用一定概率保证率的矩形识别框实现水下桩墩多类病害的分类定位;最后选取150幅未参与训练的声呐图像进行识别,验证所提出方法的有效性,并通过混淆矩阵、精确率、召回率、准确率以及F1值等评价指标对识别方法性能进行研究。研究结果发现,桩墩孔洞、剥落和位移等病害以及无病害类型的识别结果的总体准确率为88.3%,F1值分别为90.1%、84.9%、78.7%和94.6%,平均F1值为87%。这说明该方法在水下桩墩表观病害识别、定位以及自动化处理方面是可行、有效的,为桥梁水下桩墩表观病害的图像处理、智能化检测与桥梁安全评估提供技术支撑。When sonar imaging is used to detect the apparent defects of underwater piles and piers,the defect characteristics from the sonar images are remarkably different from those of the optical images,so that manual identifications prone to errors are required to detect the location and type of the defects.In order to solve this problem,this paper presents a novel deep learning and intelligent detection of apparent defects on underwater foundations of bridges based on sonar imaging.Firstly,the MS1000 sonar was used for imaging the underwater foundations and experimental models,and a large number of sonar images were obtained,so that the defects characteristics of the underwater foundations were analyzed based on these sonar images.Secondly,based the VGG16 network model,a modified faster region-based convolutional neural network(Faster R-CNN)was built.Data augmentation was implemented by rotation at every 90°,vertical and horizontal flipping operations.An approximate joint optimization training method was adopted.Rectangular bounding boxes for detection were obtained to locate the defects on underwater foundations with the category labels and classification probabilities.Finally,150 images not used to form the training sample set were selected for detection,validating the effectiveness of the proposed method.Furthermore,the evaluation indexes,including confusion matrix,precision ratio,recall ratio,accuracy and F1 score,were employed to study the performance of the detection method.The results show that the F1 scores of the holes defects,spalling,displacement,and no defects are 90.1%,84.9%,78.7%,and 94.6%,respectively.The overall accuracy of detection is 88.3%and the average F1 score is 87%.This indicates that the proposed method is feasible and effective in the identification and positioning of underwater apparent defects by the automatic processing of sonar images,and provides technical support for the intelligent detection of underwater apparent defects and safety assessment of bridge.

关 键 词：水下桩墩表观病害 Faster R-CNN 声呐成像 深度学习 

分 类 号：N945.14[自然科学总论—系统科学]