摘要
混凝土表面图像裂纹识别的传统深度学习模型往往需要大量的裂纹图像训练样本,但是在基础设施检测实际场景中,裂纹图像获取和标注困难,无裂纹图像收集则相对便利.提出了基于无裂纹图像单类样本的半监督学习方法,设计了基于自编码网络的混凝土裂纹图像识别模型Crack-Net.该模型的主干网络由自编码网络和对抗网络构成,但是为了进一步扩大裂纹图像的重构误差,CrackNet引入了近邻编码策略.近邻编码模块把输入样本的隐向量用码本中最相近的k个隐向量线性编码,编码后的向量经过解码器重构输入.为了进一步提高CrackNet的识别性能和执行效率,提出了CrackNet-T模型.该模型应用K-Means聚类算法学习紧凑的码本,同时引入阈值策略自动决策输入图像是否进行近邻编码,从而避免无裂纹图像被近邻编码改变在隐空间中的表征.本文方法在公开数据集CCIC上进行实验,结果表明CrackNet-T性能不仅优于经典的异常检测模型,而且与有监督深度学习模型识别性能相当.
Established deep learning models for concrete crack recognition often require a large number of training crack images,whereas crack images are difficult to acquire in actual infrastructure inspection tasks compared with normal images.This paper proposes a semi-supervised learning method that is solely based on normal concrete images and brings forward a concrete crack recognition model based on autoencoder,named CrackNet.This model includes autoencoder network and generative adversarial network.In addition,CrackNet equips with a Neighbor Coding(NC)module in order to enlarge the reconstruction error for crack images.NC module recodes the latent vector of an input image by the knearest latent vectors in a codebook,and then CrackNet reconstruct the input image from the recoded latent vector using the trained decoder.Furthermore,CrackNet-T is proposed to improve the performance and efficiency of CrackNet.On one hand,CrackNet-T obtains a compact codebook using K-Means clustering;on the other hand,it applies a thresholding strategy to automatically decide whether the latent vector of an input image should be recoded with NC,and therefore some normal images can avoid changing their representation in latent space.The proposed models are verified on the public dataset CCIC,and the ex perimental results demonstrate that CrackNet-T outperforms well-established anomaly detection methods and achieves comparable performance with classical supervised deep learning models.
作者
李清奇
LI Qingqi(Loudi Vocational and Technical College,Loudi Hunan 417100,China)
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2020年第2期98-104,共7页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
湖南省娄底市科技创新基金(KJ20182902)。
关键词
图像处理
裂纹识别
混凝土
深度学习
自编码器
生成对抗网络
image processing
crack recognition
concrete
deep learning
autoencoder
generative adversarial network
作者简介
第一作者:李清奇(1970-),男,湖南涟源人,副教授,硕士.研究方向为桥隧工程、计算机辅助设计.email:visint@163.com.
