Research On Crack Detection In High Resolution Image Based On Deep Learning

Traffic facilities are an important part of public infrastructure,and the automatic detection of road cracks is an important research field in the construction of intelligent transportation infrastructure.In order to ensure the safety and reliability of infrastructure,crack detection is very important.The traditional manual detection method is susceptible to the influence of environment,road shadow,strong light and road degradation and other factors,leading to the corresponding reduction of accuracy.Moreover,the manual detection of cracks is cumbersome and inefficient,requiring a large amount of manpower,material resources and time to complete the detection.In order to reduce the work cost and improve the detection efficiency and quality,it is very important to realize the automatic detection of cracks.In this paper,deep learning framework is used to study the recognition of high-resolution concrete crack images.The main research contents are as follows:First of all,this paper systematically introduces the research purpose and application significance of the subject,and summarizes the contact and non-contact concrete crack detection methods.In view of the mainstream of current development,the status quo of data-driven crack detection methods at domestic and international are described in detail.Then,the basic composition structures of the convolutional neural network are introduced,and the composition details and functional characteristics of the basic modules are expounded from the perspective of theoretical and mathematical principles.Secondly,image data sets are made according to the training requirements of deep learning algorithm,then the data sets are enhanced and annotated for subsequent network training.For crack detection problem to determine the corresponding evaluation index and model of network using Early-Stop strategy training after will complete the online training on the validation set using sliding window algorithm for crack detection experiments.The connected domain noise reduction method is adopted to further optimize the test results,the algorithm framework can accurately extract the crack location information,after optimization is test results maximum suppresses the false positive error.Following,from the network structure analysis Yolo-v4 and CenterNet networks two one stage objection detection network applied to crack the advantage of target recognition.The structural parameters of the two networks were initialized by the transfer learning strategy,and the parameters of the two networks were trained by the single-image multi-target labeling strategy.The advantages and disadvantages of the two single-stage networks in crack target detection were measured by the evaluation index and visual expression.Then,based on the structural characteristics of Yolo-v4 network,the influence of input sampling size on network detection effect was analyzed by experimental method,and the crack location of high-resolution images were accurately detected by adjusting the network input scale.Finally,in order to solve the complexity of labeling and improve the quality of detection,a kind of high-resolution image crack detection network based on weak supervision is designed.The training of the network only needs to label the image level of the data set,thus greatly reducing the cost and difficulty of data set production.The network takes into account the local similarity of the crack topology in the image,and obtains multi-scale context information by using PDC module,so that the patch-level detection accuracy can reach more than 95%.Using full convolutional structure,the network can be applied to images of any resolution.The crack image can be processed effectively and accurately through one detection,and the time consumed by using sliding window can be avoided.A center skeleton extraction algorithm was also introduced in the post-processing process to provide more accurate crack location information by refining the first-stage detection mask.