摘要
针对桥梁锈蚀数据难获取、锈蚀病害数据集小的问题,基于生成对抗网络(GAN)对桥梁锈蚀数据集进行扩增,并采用IS和K均值聚类算法验证其有效性.采用扩增后的数据集,按4∶1的比例划分为训练集和验证集,分别对U-Net网络和DeepLab-V3+网络进行训练,对比分析2种网络对于锈蚀分割的精度、召回率及F1分数.结果表明,采用深度卷积生成对抗网络(DCGAN)生成虚拟数据集的IS值达到2.41,分类肘形图类别数为5,与原数据集吻合,可作为扩增数据集以提升模型泛化性;DeepLab-V3+网络模型在验证集上的精度为0.935,召回率为0.952,F1分数为0.943,均显著高于U-Net网络模型.DeepLab-V3+网络在点状锈蚀区域连通与分割方面优于U-Net网络,并实现了片状锈蚀区域分割,为桥梁锈蚀精准识别与分割提供了技术支撑.
To solve the problems that the data of bridge corrosion is difficult to obtain and the dataset of corrosion damage is small,the bridge corrosion dataset was amplified based on generative adversarial networks(GAN).Inception score(IS)and K-means clustering algorithm were adopted to verify the effectiveness of the amplified dataset.The amplified dataset was divided into a training dataset and a validation dataset at a ratio of 4∶1.U-Net network and DeepLab-V3+network were respectively trained by the training dataset.The precision,the recall and F 1 score of the two networks for corrosion segmentation were compared.The results show that the IS value of the synthetic dataset generated by the deep convolutional generative adversarial networks(DCGAN)is 2.41,and the category number of the classification elbow digram is 5,which is consistent with the original dataset.The synthetic dataset can be added as amplified datasets to improve model generalization.The precision,the recall rate,and the F 1 score of the DeepLab-V3+network model are 0.935,0.952,and 0.943 in the validation dataset,respectively,which are significantly higher than those of the U-Net network model.DeepLab-V3+network is superior to U-Net network in connection and segmentation of spot-corrosion area,and realizes segmentation of flake-corrosion area,providing technical support for accurate detection and segmentation of bridge corrosion.
作者
倪有豪
陆欢
季超
茅建校
王浩
徐寅飞
Ni Youhao;Lu Huan;Ji Chao;Mao Jianxiao;Wang Hao;Xu Yinfei(Key Laboratory of Concrete and Prestressed Concrete Structures of Ministry of Education,Southeast University,Nanjing 211189,China)
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2023年第2期201-209,共9页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(51978155,52108274)
江苏省研究生培养创新工程研究生科研与实践创新计划资助项目(SJCX21_0056)。
关键词
深度学习
生成对抗网络
桥梁锈蚀识别
语义分割
deep learning
generative adversarial networks
corrosion detection of bridges
semantic segmentation
作者简介
倪有豪(1987-),男,博士生;联系人:茅建校,男,博士,副研究员,jianxiao@seu.edu.cn.
