基于深度学习的盾构隧道渗漏水病害图像识别

随着城市地铁隧道急剧增加的养护需求,地铁盾构隧道结构病害尤其是渗漏水病害亟需快速精准的识别诊断。利用计算机视觉对盾构隧道进行健康检测是近年来国内外的一种新趋势,但目前渗漏水病害图像的识别效果尚不能满足实际工程的需要。在分析盾构隧道衬砌表面图像特点的基础上,将渗漏水图像分为6种类别,采用深度学习的方法,提出一种新颖的基于全卷积网络的盾构隧道渗漏水病害图像识别算法,并从图像识别结果、错检率和运行时间三个方面与大律法、区域生长法、分水岭法等传统图像识别方法进行对比分析。研究表明:基于全卷积网络的盾构隧道渗漏水病害的图像识别能够有效地避免管片拼缝、螺栓孔、管线、支架等干扰物的影响,特别是在克服管线遮挡方面具有优越的鲁棒性;与传统图像识别算法相比,提出方法在错检率和运行时间上具有较大优势,能够更好地满足工程需要。

Image recognition for water leakage in shield tunnel based on deep learning

With the sharp increasing in the requirements for tunnel maintenance of urban metro，the structural defects of the metro shield tunnels，especially the water leakage，need be inspected fast and accurately. Using the computer vision to inspect the structural defects of shield tunnel is a thriving trend home and abroad during recent years. However，the traditional methods of image recognition on water leakage cannot meet the need in engineering practice. In this paper，a novel method employing the fully convolution network(FCN) based on the deep learning(DL) is proposed to improve the image recognition on water leakage in shield tunnel. The water leakage images are divided into six categories according to the lining surface of shield tunnel and some interference on image recognition. The recognition results，the error rates and the running time from FCN are compared with those from the traditional image recognition methods of Otsu algorithm(OA)，region growing algorithm(RGA) and watershed algorithm(WA). The results show that DL-based image recognition on water leakage effectively avoided the interference from the segment joints，bolt holes，cables，brackets etc.，and has excellent robustness in overcoming the defects shelter from cables.