首页|基于改进Faster R-CNN与U-Net算法的桥梁病害识别与量化方法

基于改进Faster R-CNN与U-Net算法的桥梁病害识别与量化方法

扫码查看
原文链接
  • 万方数据
  • 维普
为实现桥梁病害检测的自动化,对基于图像处理技术的混凝土桥梁表观病害的智能识别和尺寸确定方法展开研究.提出基于改进Faster R-CNN算法的病害识别方法,利用K均值聚类和遗传算法对区域候选网络锚框进行优化设计;以裂缝预测区域为基础,提出ResNet34结合U-Net的裂缝形态提取方法,并结合裂缝形态学研究了裂缝像素宽度和长度的确定方法.结果表明:锚框优化设计可改进Faster R-CNN算法的表观病害识别效果,5类常见病害的预测准确率、召回率、平均精确率分别由68.40％、69.87％、74.64％提升到85.40％、83.59％、83.72％;利用病害预测框,结合改进U-Net算法的裂缝像素尺寸计算,可实现裂缝病害尺寸的自动测量;基于改进Faster R-CNN和改进U-Net的方法可实现混凝土桥梁常见病害的智能识别和尺寸量化,从而提高桥梁病害检测效率并促进桥梁技术状况评定的智能化.
Bridge defects detection and quantifying method based on modified Faster R-CNN and U-Net
In order to realize the automated detection of bridge defects,the intelligent identification and dimen-sion measurement of surface diseases of concrete bridges was studied based on the digital image processing technology.The disease identification method based on improved faster region convolutional neural networks(Faster R-CNN)algorithm was improved,the design of region proposal network(RPN)anchors was opti-mized using K-means clustering and genetic algorithm.Based on the predicting region of strip crack disease,the crack morphology was extracted by combining the ResNet34 algorithm and U-Net segmentation method.Finally,the pixel width and length of crack were calculated by crack morphology.The results show that the optimized design of RPN anchors can improve the recognition effect of Faster R-CNN algorithm for surface dis-eases.The prediction accuracy,average recall and average precision of five common diseases are increased from 68.40％to 85.40％,69.87％to 83.59％and 74.64％to 83.72％,respectively.The automated dimension measurement of crack diseases can be realized using the diseases predicted anchor and pixel dimension calcula-tion of improved U-Net algorithm.The improved Faster R-CNN and U-Net algorithm can realize the intelligent identification and quantification of common diseases in concrete bridges,contributing to improve the efficiency of bridge disease detection and promote the intelligence of bridge technical condition assessment.

bridge structuresurface disease recognitioncrack dimension determinationimproved faster re-gion convolutional neural networks(Faster R-CNN)improved U-Net

乔朋、梁志强、段长江、马晨、王思龙、狄谨

展开 >

长安大学建筑工程学院,西安 710061

长安大学公路学院,西安 710061

重庆大学土木工程学院,重庆 400045

桥梁工程 表观病害识别 裂缝尺寸确定 改进Faster R-CNN 改进U-Net

国家自然科学基金资助项目陕西省自然科学基础研究计划资助项目陕西省自然科学基础研究计划资助项目

521926632020SF-3822021JM-181

2024

10.3969/j.issn.1001-0505.2024.03.013

东南大学学报(自然科学版)
东南大学

东南大学学报(自然科学版)

CSTPCD北大核心
影响因子：0.989
ISSN：1001-0505
年,卷(期)：2024.54(3)