探地雷达图像隧道衬砌病害智能识别与形态分割方法

YOLACT based targets categorization and shape identification from ground penetrating radar images

探地雷达具有高效率、非破坏性、穿透性强和分辨率高等优点,使其成为隧道衬砌病害检测的主流方法.针对目前大多数深度学习网络只能检测出探地雷达图像中隧道衬砌病害的类型和位置,难以准确识别病害的轮廓的问题,本文提出了一种基于探地雷达图像的隧道衬砌智能病害分类与轮廓识别方法,首先构建了基于YOLACT(You Only Look At CoefficienTs)的病害轮廓及类型识别网络模型,然后结合裂缝、脱空、空洞等六类病害的介电常数模型采用时域有限差分法正演生成仿真探地雷达数据集进行模型训练,最后通过无水隧道衬砌病害和含水隧道衬砌病害的仿真数据以及隧道现场数据进行该方法的测试验证,通过模型分析和识别探地雷达B-Scan图像中的病害分类与轮廓.研究试验表明,表明本文方法在仿真数据中对裂缝、空洞、脱空、含水裂缝、含水空洞和含水脱空这六类典型隧道衬砌病害的类别和轮廓的识别结果较为准确,识别平均准确率达92.28%,同时该方法在现场试验真实探地雷达数据中也能够有效识别衬砌病害的轮廓,在实际工程中具有广泛应用前景.

Ground Penetrating Radar (GPR) has the advantages of high efficiency, non-destructive, strong penetrability and high resolution, making it the mainstream method of tunnel lining disease detection. To solve the problem that most of the current depth learning networks can only detect the type and location of tunnel lining diseases in the GPR image, and it is difficult to accurately identify the contour of the disease, this paper proposes an intelligent disease classification and contour recognition method for tunnel lining based on the GPR image. First, a disease contour and type recognition network model based on YOLACT (You Only Look At CoefficienTs) is constructed, and then combined with the dielectric constant models of six kinds of diseases, such as cracks, voids and cavities, the finite difference time domain method is used to forward generate the simulated ground penetrating radar data set for model training. Finally, the method is tested and verified by the simulation data of waterless tunnel lining disease and water bearing tunnel lining disease as well as the tunnel site data. The disease classification and contour in the B-Scan image of GPR are identified through model analysis. The research results show that the method in this paper is more accurate in identifying the categories and contours of six typical tunnel lining diseases, including cracks, cavities, voids, water containing cracks, water containing cavities and water containing voids, in the simulation data, with an average recognition accuracy of 92.28%. At the same time, the method can also effectively identify the contours of lining diseases in the field test real ground penetrating radar data, which has broad application prospects in practical projects.