基于轻量级可变形卷积实例分割的桥梁病害无人机成像识别研究  

UAV Imaging Recognition of Bridge Diseases Based on Lightweight Deformable Convolution Instance Segmentation

作　　者：杨达赖 唐德密 王超伦 李晓光 李千禧 彭雄 YANG Dalai;TANG Demi;WANG Chaolun;LI Xiaoguang;LI Qianxi;PENG Xiong(Huazhong of CCCC First Highway Engineering Group Co.,Ltd.,Wuhan,Hubei430019,China;School of Civil Engineering,Hunan University of Science and Technology,Xiangtan,Hunan411201,China)

机构地区：[1]中交一公局集团华中工程有限公司,湖北武汉430019 [2]湖南科技大学土木工程学院,湖南湘潭411201

出　　处：《施工技术(中英文)》2025年第6期170-177,共8页Construction Technology

基　　金：教育部“春晖计划”合作科研项目:基于机理⁃数据驱动的无人飞机桥梁“体检”大数据智能评估方法研究(HZKY20220354)。

摘　　要：利用无人机桥梁病害检测机器视觉系统,可获取大量桥梁混凝土表面病害数据。针对桥梁病害尺寸细长特点,引入实例分割YOLOv7_Seg(You Only Look Once V7_segmentation)网络,并对网络进行轻量级可变形卷积改进,拓宽小目标缺陷的感受野,有效增强小尺度缺陷特征,实现露筋、破损、蜂窝麻面、裂缝4种桥梁病害的分割与识别;以跨江大桥为例,利用无人机采集病害数据,并使用训练好的模型进行实例分割。结果表明,所提出的系统和方法具有较高识别准确性及广阔应用前景。The instance segmentation YOLOv7_Seg(You Only Look Once V7_segmentation)network is introduced and lightweight deformable convolution improvement is made to broaden the receptive field of small target diseases,effectively enhance the characteristics of small⁃scale diseases,and achieve the segmentation and recognition of four types of bridge diseases,including exposed reinforcement,damage,honeycomb and pitted surface,and cracks.Taking a cross river bridge as an example,the model training is completed using unmanned aerial vehicle to collect diseases data for instance segmentation.The results show that the proposed method has high recognition accuracy and wide application prospect.

关 键 词：桥梁 病害 无人机 可变形卷积 实例分割 

分 类 号：U446[建筑科学—桥梁与隧道工程]