摘要
本文针对混凝土常见病害,提出了一种改进Unet网络对混凝土多病害进行检测的方法,建立了一个包含混凝土裂缝、泛碱、剥落、露筋、孔洞数据集;再裁剪原始图像,选取10000张小块图像进行随机翻转,共产生13200张病害数据集用于研究;最后,将数据集对改进的Unet网络进行训练、验证和测试,实现了像素级病害区域识别。为验证网络模型性能,将模型与原Unet、基于VGG16的Unet、FCN和基于ResNet50的FCN进行了比较,结果表明所改进的网络模型检测效果优于其他模型。为测试模型鲁棒性,利用滑动窗口算法对不同环境下的混凝土病害进行全局检测,取得了较好的检测效果。
Aiming at the common diseases of concrete, this paper proposes an improved Unet network to detect multiple diseases of concrete, and establishes a data set containing concrete cracks, efflorescence, spalling, exposed tendons, and holes;afterwards, the original images are cropped and 10000 pieces are selected. The block images are randomly flipped to produce a total of 13200 disease data sets for research;finally, the data sets are used to train, verify and test the improved Unet network to realize pixel-level disease area recognition. In order to verify the performance of the network model, the model was compared with the original Unet, VGG16-based Unet, FCN and ResNet50-based FCN. The results show that the improved network model detection effect is better than other models. In order to test the robustness of the model, the sliding window algorithm is used to globally detect concrete diseases in different environments, and good detection results have been achieved.
作者
饶勇成
韩晓健
肖飞
孙思其
RAO Yongcheng;HAN Xiaojian;XIAO Fei;SUN Siqi(College of Civil Engineering,Nanjing Tech University,Nanjing 211800,China;Jiangsu Jian Yan Civil Engineering Quality Appraisal Co.,Ltd.,Nanjing 211800,China)
出处
《建筑结构》
CSCD
北大核心
2021年第S02期1439-1445,共7页
Building Structure
关键词
深度学习
语义分割
混凝土
病害检测
deep learning
semantic segmentation
concrete
disease detection