摘要
为了克服传统数字图像处理方法进行桥梁裂缝识别时面临的效率低、效果不佳等问题,提出了集成深度学习YOLOv5和U-Net3+算法的一体化桥梁裂缝智能检测方法.通过调整算法宽度和深度参数,优化边界框损失函数,构建基于YOLOv5目标检测算法的裂缝识别定位模型,实现桥梁裂缝快速识别与定位;引入结合深度监督策略及预测输出模块的U-Net3+图像分割算法,训练并构建桥梁裂缝高效分割模型,实现像素级裂缝智能化提取;建立结合连通域去噪、边缘检测、形态学处理的八方向裂缝宽度测量法,基于U-Net3+裂缝分割结果实现裂缝形态及宽度高精度测量;利用LabelImg图像标注软件制作包含4414张图像的裂缝识别定位模型训练数据集;利用LabelImg图像标注软件及CFD数据集制作包含908张图像的裂缝分割模型训练数据集;利用无人机航拍的485张5280×2970 pixels桥梁索塔裂缝图像,来制作裂缝智能检测模型的测试对象.将所提出的裂缝检测方法应用于上述裂缝测试对象,其裂缝识别定位准确率91.55%、召回率95.15%、F1分数93.32%,裂缝分割准确率93.02%、召回率92.22%、F1分数92.22%.结果表明,基于YOLOv5与U-Net3+的桥梁裂缝智能检测方法,可实现桥梁裂缝高效率、高精度、智能化检测,具有较强的研究价值和广泛的应用前景.
To overcome the problems of low efficiency and poor effect when using traditional digital image pro⁃cessing methods to detect bridge cracks,this paper proposes an integrated bridge crack detection method that integrates deep learning YOLOv5 and U-Net3+algorithms.By adjusting the width and depth parameters and optimizing the bounding box loss function,a crack identification and location model based on the YOLOv5 target detection algo⁃rithm is constructed to realize the rapid identification and location of bridge cracks.The U-Net3+image segmenta⁃tion algorithm combined with a deep supervision strategy and the output prediction module is introduced to train and build an efficient segmentation model of bridge cracks,and to realize pixels-level intelligent extraction of cracks.An eight-direction crack width measurement method combined with connected domain denoising,edge detection,and morphology processing is developed.And the morphology and width of cracks are measured with high precision based on U-Net3+segmentation results.LabelImg image annotation tool is used to make a dataset containing 4414 images to train the crack identification and location model.LabelImg image annotation tool and CFD dataset are used to make a dataset containing 908 images to train the crack segmentation model.UAV is used to capture 485 images of size 5280×2970 pixels,which are taken from the bridge tower.The crack images of the bridge tower are used as the test object of the intelligent crack detection model.The proposed crack intelligence detection method is applied to the above test objects,the overall precision,recall,and F1 score of crack identification and location are 91.55%,95.15%,and 93.32%,respectively,and the overall precision,recall,and F1 score of crack segmentation are 93.02%,92.22%,and 92.22%,respectively.The results show that the intelligence detection method of bridge cracks based on YOLOv5 and U-Net3+algorithms can achieve high efficiency,high precision,and intelligence de⁃tection of bridge cracks,which has much research value and broad application prospects.
作者
余加勇
刘宝麟
尹东
高文宇
谢义林
Yu Jiayong;Liu Baolin;Yin Dong;Gao Wenyu;Xie Yilin(Key Laboratory of Wind Engineering and Bridge Engineering in Hunan Province,Hunan University,Changsha 410082,China;College of Civil Engineering,Hunan University,Changsha 410082,China;Jiangsu Hydraulic Research Institute,Nanjing 210017,China)
出处
《湖南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第5期65-73,共9页
Journal of Hunan University:Natural Sciences
基金
湖南省自然科学基金资助项目(2021JJ30102)
湖南省自然资源厅科技计划项目(No.2023-20)
湖南省水利科技项目(XSKJ2021000-46)。
