为解决水工隧洞尺寸较小、断面形式多变、洞内环境复杂、无信号覆盖给检测图像快速采集带来的难题，研制出一种无信号覆盖下图像实时传输以及适用于不同断面、尺寸下隧洞全断面采集的水工隧洞检测系统，采集重叠率和精度满足三维重建及对自主研发的衬砌展开影像生成器的要求；针对隧洞表观病害传统算法速度慢、精度低等问题，引入DenseNet主干网络和ECA注意力机制对YOLOv5检测算法进行改进，优化特征提取能力；针对单张图像识别难以反映病害整体分布、全景展开图尺寸过大网络难以捕获有效特征等问题，以全景展开图底层图像为种子点，利用方向搜索结合目标检测算法提出一种“段-环-点”的隧洞常见表观病害快速识别算法。现场应用表明： 1）相比YOLOv7-tiny、YOLOv8、YOLOv9等模型，改进的YOLOv5模型在mAP和F1上均有明显提高，可有效提升复杂背景下病害识别与边界定位精度，显著减少误检和漏检现象； 2）所提技术适用于圆涵、箱涵等不同类型、尺寸的水工隧洞表观图像快速获取与全景识别，能够用于隧洞表观病害的展示和快速定位，可为病害整治措施的及时制定提供依据。
Challenges posed by
hydraulic tunnels, such as confined space, varying cross-sections, complex environments, and lack of signal, hinder rapid
image acquisition. To address these challenges, a rapid image acquisition
system that can capture cross-sectional images of the
tunnel while traveling in a short period is presented. This system can capture
images under different cross-sectional shapes and sizes
and meets the acquisition overlap rate and accuracy requirements of three-dimensional reconstruction and image generator for lining expansion.
To enhance the speed and accuracy of the traditional algorithms, the YOLOv5
detection algorithm is improved using a DenseNet network and an efficient
channel attention module, thereby enhancing the effectiveness of feature
extraction. In addition, a "segment-ring-point" method is proposed for the efficient identification of
common surface defects on tunnel linings using panoramic images, with the
bottom image of the panorama serving as the seed point. This approach addresses
the limitations of single images that fail to reflect the overall distribution
of the damage and large images that do not capture the effective features of
the network. Field applications demonstrate the following: (1) The improved
YOLOv5 model significantly improves the mean average precision and F1 performances compared with YOLOv7-tiny, YOLOv8, and YOLOv9 models, effectively improving the accuracy
of defect identification and boundary positioning in complex environments and
substantially reducing the occurrence of false positives and missed detections.
(2) This approach is suitable for rapid acquisition and panoramic image
stitching in hydraulic tunnels with diverse cross-sectional
shapes and sizes. It can also be used for identifying and locating surface
diseases promptly in tunnels to provide defect treatment measures.
