基于深度学习的裂缝识别与量化分析研究
Crack Recognition and Quantitative Analysis Based on Deep Learning
范存君 1金松燕 1金楠 1施钟淇 1伍永靖邦 1郝新田1
作者信息
- 1. 深圳市城市公共安全技术研究院有限公司,深圳市城市灾害数字孪生重点实验室,广东深圳 518023;城市安全发展科技研究院(深圳),广东深圳 518023;城市安全风险监测预警应急管理部重点实验室,广东深圳 518023
- 折叠
摘要
裂缝是混凝土结构中常见的表观损伤之一,对结构性能的评估具有重要的意义.利用计算机视觉技术对混凝土结构表面进行裂缝识别与量化已得到广泛研究,然而,基于深度学习的裂缝识别技术依赖于大规模的裂缝数据集进行训练.为此,提出基于风格迁移网络的数据增广方法,利用少量裂缝数据和各种背景图像数据,构建大规模的、复杂背景的裂缝数据集,并通过训练YoloV8网络模型实现裂缝的识别与分割,并根据裂缝特性,对识别结果中的孤立、微小区域进行滤除.在此基础上,基于已知参考标志物进行裂缝宽度量化分析,试验结果表明裂缝宽度计算误差基本控制在20％以内.
Abstract
Cracks are a common form of surface damage in concrete structures and have significant implica-tions for assessing structural performance.The use of computer vision techniques for crack recognition and quantification on the surface of concrete structures has been widely studied.However,deep learning-based crack recognition techniques rely on large-scale crack datasets for training.To address this issue,the paper proposed a data augmentation method based on style transfer networks.A large-scale,complex-background crack dataset was constructed by using a small amount of crack data and various background image data.The YoloV8 network model was trained to achieve crack recognition and segmentation.Based on the crack characteristics,isolated and tiny areas in the recognition results were filtered.Based on this,crack width quantification analysis was performed based on known reference markers,and the experimental results showed that the calculation error of crack widths was basically controlled within 20％.
关键词
深度学习/数据扩充/裂缝识别/裂缝宽度Key words
deep learning/data augmentation/crack recognition/crack width引用本文复制引用
基金项目
国家重点研发计划(2022YFC3800304)
深圳市科技计划资助(ZDSYS20210929115800001)
深圳市科技计划资助(KJZD20231023093057005)
青年人才托举工程(2023QNRC001)
出版年
2024
NETL
NSTL
万方数据