Research And Implementation Of Concrete Crack Detection Method Based On Deep Learning

With the rapid growth of concrete infrastructure construction in China,more and more attention has been paid to the health inspection and maintenance of the facilities.Cracks will inevitably appear in concrete structures during construction or long-term use,and if these cracks are not detected and repaired in time,they will not only affect the external appearance,but also the normal use of concrete structures,so the detection of concrete cracks is crucial.The current manual detection method is inefficient and vulnerable to subjective factors.With the rapid development of computer vision,researchers have started to use images to complete crack detection.Traditional methods of image processing techniques require human intervention and are susceptible to factors such as light intensity.In view of this,this thesis uses a deep learning approach to solve the problem of concrete crack detection,and the main research includes three aspects:(1)Object detection method based on deep learning.An improved concrete crack detection method based on YOLOv4 network is proposed to address the problems such as high weight and complexity of YOLOv4 network.Firstly,the backbone feature extraction network of YOLOv4 is replaced with Mobile Netv2,and the standard convolution of the deep network is replaced with a deep separable convolution to make the network lighter.Then the SENet attention mechanism is incorporated to compensate for the loss of detection accuracy caused by network light-weighting.Finally,the dataset is constructed by acquiring crack images and data augmentation,labeled in a unique way,and clustered using the K-means algorithm to generate anchors suitable for crack detection.Experiments show that this method greatly reduces the network weight,parameter amount and calculation amount while maintaining a high detection level,and shortens the detection time,which can meet the needs of various concrete crack detection tasks.(2)Semantic segmentation method based on deep learning.An improved concrete crack detection method based on U-Net network is proposed for the problems of fine crack leakage and false detection in complex concrete background.Firstly,dilated convolution is introduced to form a parallel convolution module,which enables the decoder to better recover the original image information.Then the SENet attention mechanism is incorporated for the parallel convolution module to strengthen the important feature channels and suppress the useless feature channels.The experiments show that the precision,recall and F1 value of the method are improved and can meet the needs of various concrete crack segmentation.(3)Concrete crack detection system.Crack detection methods using deep learning can achieve better detection results,and in order to make them better applicable to various realistic scenarios,a concrete crack detection system based on web-side implementation is designed.The whole system integrates object detection function and semantic segmentation function,which can import system images from the front-end and return detection results through back-end processing for user-friendly operation and use.