复杂环境公路隧道裂缝快速识别与分割算法研究

随着我国公路隧道由建设为主朝建养并重转化,在运营里程快速增长与既有隧道劣化加剧的双重作用下,移动检测及结构安全快速诊断已成为目前公路隧道运营维养领域的研究热点之一。我国已研发了多种类型的隧道检测车,为裂缝、渗漏水等表观病害的快速检测提供了手段,然而公路隧道衬砌图像背景复杂、干扰因素多、裂缝占比小的特点,给检测数据的快速分析带来巨大挑战,已成为制约技术推广的主要瓶颈。基于深度学习算法,本文提出了一种将目标识别与语义分割进行组合的裂缝快速提取方法,首先采用Faster R-CNN网络对原始衬砌图像进行目标识别,判定所采集图片是否存在裂缝并智能框选出裂缝区域;随后对框选出的裂缝区域自动裁切,由此过滤不含裂缝的图片并去除含裂缝图片中的干扰背景,再利用U-Net语义分割网络对裂缝进行像素级分割。通过实际工程验证发现,单幅图像裂缝整体分割时间小于0.15 s,在常见各类干扰因素下,目标识别F1值可达到92%,语义分割像素准确度可达到98%以上。与阈值分割和同类智能分割算法相比,本方法显著提高了识别速度与精度,为从隧道检测车海量数据中进行快速准确的裂缝提取提供了良好手段。

Research on Fast Identification and Segmentation Algorithms for Cracks of Highway Tunnels in Complex Environment

With the rapid construction of highway tunnels in China turns to the stage of both construction and maintenance, it will face the double pressure of the rapid growth of operating mileage and the deterioration of existing tunnels in the future, and the mobile detection and rapid diagnosis of structural safety have become research hotspots in the field of highway tunnel operation and maintenance. China has developed a lot of detection vehicles being used to tunnel apparent defects, which can provide an efficient method for rapid detection of tunnel surface cracks, water leakage, etc. However, highway tunnel lining images are characterized by complex backgrounds, multiple interference factors and low cracks occupancy, and accordingly, it's difficult to do rapid identification and analysis of mass tunnel detection data, which has become the main bottleneck of technology promotion. Based on deep learning algorithm, this paper proposes a method that combines the target recognition with semantic segmentation algorithm. Firstly, Faster R-CNN deep neural network is used for target recognition on the original lining images to determine whether there are cracks and intelligently mark the crack regions within a rectangle. Then, the crack area selected by the frame is automatically cut, thereby filtering the pictures without cracks and removing the interference background in the pictures containing cracks, and then using the U-Net semantic segmentation network to segment the cracks at the pixel level. In practical engineering, it takes less than 0.15s to recognize the cracks in one image. Besides, this method can effectively identify various cracks in complex environments, with the F1 score of target recognition reaching 92% and the segmentation accuracy rate 98%. Compared with the previous feature recognition and global segmentation intelligent algorithms, this method significantly improves the speed and accuracy of recognition, and provides a feasible method for rapid and accurate recognition of mass defect detection data.