面向复杂环境下混凝土桥梁裂缝病害的视频检测方法研究

The automatic concrete bridge surface crack detection is an important approach for fast evaluating the concrete structure’s durability. However, the existing vision-based bridge crack detecting methods mostly perform badly in the complicated service surroundings, and that is because of the weak crack image feature extraction, the imprecise crack region detection and the troublesome background false alarms. To address these issues, the laboratory model experiments and the on-site field crack detections are conducted in the following researches: the sparse imaging property of concrete bridge crack regions is analyzed, and the multi-layer sparse feature extraction model for concrete bridge images is presented, which can improve the surrounding anti-noise capability. On the basis of image feature representation, we advocate a twin adversarial crack classified detecting network, which would fully suppress the diversity for different crack images, and boost the discrimination between crack images and background disturbances. In order to remove the background false alarm effectively, the spatio-temporal probability statistic model will be established. Via taking full advantage of the space-time and content relevant information of video images, the robustness of crack detector can be further improved. The above research achievements will provide the novel bridge crack detecting technique and related theoretical supports, and contribute to the fast and precise operation for the bridge infrastructure engineering.

混凝土桥梁表面裂缝病害自动检测是快速评估其结构耐久性的重要途径。针对复杂服役环境下基于视觉分析的桥梁裂缝检测方法普遍存在的图像特征提取性能差、病害区域检测不精准以及背景虚警剔除难等问题，本课题采用室内试验和现场检测分析相结合的方法开展以下研究：分析混凝土桥梁裂缝区域的稀疏成像特性，建立桥梁结构区域图像的多层稀疏特征提取模型，提高系统对环境噪声的抗干扰能力。在完成图像特征提取的基础上，考虑抑制不同裂缝样本之间的多样性以及强化裂缝与桥梁背景杂波之间的差异性，建立混凝土桥梁裂缝区域孪生对抗分类检测模型。为有效剔除裂缝检测背景虚警，提出采用视频图像检测的方式，考虑融合利用视频数据的时空内容相关信息建立裂缝时空链概率统计模型，进一步提高桥梁裂缝检测模型的鲁棒性。本课题研究成果可为混凝土桥梁表面裂缝缺陷的自动无损检测提供新的技术与理论支持，对促进桥梁基础设施的快速科学检修具有重要意义。

利用数字图像分析技术检测与识别桥梁结构表面病害已成为桥梁结构数字化巡检的一大亮点。然而在复杂多变的桥梁服役环境中，相比于其他表面病害（比如锈蚀、掉块等），裂缝病害目标的信噪比较低且更易受各种背景噪声（如光照、划痕、麻面、黏贴物等）的干扰。基于此，本项目提出首先利用丰富的图像特征表示方法和强大的模式分类技术，建立基于单帧图像分析的裂缝病害局部检测与识别模型：在图像区域特征表示方面，利用方块图像裂缝区域像素特有的结构稀疏性，建立基于多层超限自编码器的裂缝图像特征提取方法；在样本图像分类与识别方面，提出以裂缝和背景样本特征构建训练样本对，并基于超限分类学习框架，建立基于孪生对抗学习的裂缝分类识别网络，强化裂缝检测模型的识别能力。此外，考虑桥梁巡检图像移动拍摄特点，提出利用视频图像序列之间的时空相关信息，在桥梁裂缝病害单帧图像处理的基础上，建立了基于时空链图像相关分析的背景虚警剔除方法，进一步提高了裂缝病害检测模型在复杂环境中的适用性。本项目的研究成果为探索复杂环境下桥梁结构快速巡检以及相关结构表面裂缝病害检测与识别等应用领域提供了新的理论和技术支持。