Detection Of Concrete Surface Crack Based On Deep Convolutional Neural Network

In recent years,artificial intelligence technology with deep learning as its core has developed rapidly.In the field of civil engineering,image recognition technology in artificial intelligence is widely used in structural damage detection.The traditional crack detection based on concrete images uses image processing,which requires high image preprocessing technology,and the detection results are easily affected by factors such as light and noise.Convolutional neural network does not require artificial feature extraction to detect concrete cracks.When the number of training samples of the network is sufficient,its detection accuracy and robustness are better than image processing technology,and it has a good application prospect.This paper systematically studies many factors in the process of detecting cracks using convolutional neural networks,and tests many better convolutional neural network models.The main research contents and conclusions of this article are as follows:(1)Four issues of data set partitioning,network type,test method,and crack marking method during the detection of concrete cracks by convolutional neural network.A better concrete crack detection framework is proposed.The framework uses a semi-automatic sample labeling method to build a training data set that does not discard samples,trains the network on this data set,and finally uses the center tag or the filter tag to complete the fracture localization.Studies have shown that using this framework saves time in sample classification and improves crack detection accuracy.(2)The effects of the Alex Net,Inception-v3,Res Net＿18,Res Net＿50,Xception and Dense Net201 network models on the detection of concrete cracks were explored.Research shows: Dense Net201 has the best crack detection result,followed by Inception-v3,and Alex Net and Xception has the worst crack detection result.From the perspective of saving network computing time,Res Net＿18 and Res Net＿50 are combined to build a Res Net18＿50 network,which improves the result of crack detection and saves time costs.(3)Using FCN-8s,FCN-16 s and FCN-32 s in the semantic segmentation network to perform pixel-level detection of concrete cracks.Research shows: FCN-8s is better for detecting concrete cracks.Dilated convolution is introduced to improve FCN-8s.The test results show: The improved network improves the accuracy of crack detection and has good adaptability to the interference of the rough surface of concrete.The improved FCN-8s network is tested on new data set,the result shows that the new network is better than FCN-8s in detecting ordinary concrete crack pictures.