基于深度学习的全卷积网络图像裂纹检测

Image Crack Detection with Fully Convolutional Network Based on Deep Learning

为实现复杂背景下裂纹目标的有效检测及降低错误标记,将全卷积网络(FCN)引入图像裂纹检测中,并针对裂纹检测实验中FCN模型存在丢失局部信息和丧失部分精细化区分能力的问题,构建一种Crack FCN模型.首先在增大分辨率的同时,取消全连接层中的Dropout技术,以增大裂纹信息的选择;其次通过加深FCN的网络深度,使整个网络实现递进式特征传递;最后在网络之后添加更高尺度的反卷积层来扩充局部精细细节.在2 156幅自制的裂纹图像数据集上对文中模型、FCN-8s模型以及其他检测方法进行实验的结果表明,Crack FCN网络模型在提高检测精度的同时可以有效地降低错误标记.

In order to effectively detect crack and reduce the error mark under complex background,a FCN fully convolutional network is introduced into the image crack detection in this paper.A crack FCN model based on FCN model is proposed to address the problems that lose local information and the capacity of partial refinement,which are frequently encountered in FCN model in the crack detection experiment.Firstly,with increasing resolution,the Dropout method of the full connection layer is cancelled,such that the increased selection of the crack information is fulfilled.Secondly,the whole network is carried out in a progressive manner by deepening the network depth of the FCN network.Finally,deconvolution layer in higher scale is added to extend the local fine detail based on the constructed network.According to the comparison experiment results,FCN-8s model and other detect methods in the 2 156 self-made crack image datasets illustrate that the accuracy of detection can be improved,whilst the error label can be reduced by using the Crack FCN network model.