基于改进Mask R-CNN的混凝土坝裂缝像素级检测方法  

Pixel-level crack detection method of concrete dam based on improved Mask R-CNN

作　　者：牛慧余 包腾飞[1,2,3] 李扬涛 黄思文 NIU Huiyu;BAO Tengfei;LI Yangtao;HUANG Siwen(College of Water Conservancy and Hydropower Engineering,Hohai University,Nanjing 210098,China;State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering,Hohai University,Nanjing 210098,China;College of Hydraulic&Environmental Engineering,China Three Gorges University,Yichang 443002,China)

机构地区：[1]河海大学水利水电学院,江苏南京210098 [2]河海大学水文水资源与水利工程科学国家重点实验室,江苏南京210098 [3]三峡大学水利与环境学院,湖北宜昌443002

出　　处：《水利水电科技进展》2023年第1期87-92,98,共7页Advances in Science and Technology of Water Resources

基　　金：国家重点研发计划(2018YFC1508603)。

摘　　要：为解决传统人工巡检混凝土坝裂缝效率低下的问题,将人工智能和计算机视觉技术相结合,提出了复杂背景下混凝土坝裂缝像素级精细化自动识别分割方法。为克服复杂环境背景因素干扰,基于多种数字图像处理手段对混凝土坝裂缝图像进行预处理,有效去除环境噪声。在Mask R-CNN基础上,对模型主干网络进行改进以提升裂缝特征提取能力。采集500幅包含单裂缝、多裂缝、交叉裂缝、龟裂等多种裂缝形态的混凝土坝裂缝图像用于模型训练与验证,并采用定性分析和定量评估相结合的手段多维度评估模型泛化能力和鲁棒性。结果表明:改进Mask R-CNN对含有多种裂缝特征及噪声背景的裂缝图像识别效果良好,模型在测试集上的目标检测和分割掩码平均精确度分别达76.3和61.9,满足高精度裂缝精细分割要求;与Cascade-Mask R-CNN、Yolact++等基准模型相比,改进Mask R-CNN在目标检测、分割掩码精确度及模型推理速度多方面均有一定的优势。In order to solve the problem of low efficiency of traditional manual inspection, artificial intelligence and computer vision technology was combined to propose a pixel-level refined automatic identification and segmentation method for concrete dam cracks under complex backgrounds.To overcome the interference of complex environmental background factors, the concrete dam crack image is preprocessed based on various digital image processing methods to effectively remove environmental noise. Based on the Mask R-CNN instance segmentation model, the model backbone network is optimized to improve the crack feature extraction ability. In the experiment, 500 images of concrete dam cracks, including single crack, multiple cracks, cross cracks, map cracks and other crack forms, were collected for model training and verification. A combination of qualitative analysis and quantitative evaluation was used to evaluate the generalization ability and robustness of the model in multiple dimensions. The test results show that the improved Mask R-CNN model in this paper has a good effect on crack image recognition with various crack features and noise backgrounds. The average accuracy values of the model’s target detection and mask segmentation on the test set are 76.3 and 61.9, respectively, which can meet the requirements of high-precision fine segmentationfor cracks. Compared with benchmark models such as Cascade-Mask R-CNN and Yolact++, the improved Mask R-CNN method has certain advantages in object detection, mask segmentation accuracy and model inference speed.

关 键 词：混凝土坝 裂缝检测 实例分割 改进Mask R-CNN 人工智能 

分 类 号：TV698.1[水利工程—水利水电工程]