Research On Pavement Defect Detection Algorithm Based On Deep Learning

The road is one of the major ways of modern transportation.It plays an important role in the economic development of a nation and is the most basic and extensive transport infrastructure.Due to vehicle crushing or natural disasters,the pavement is prone to cracks,potholes,and other defects during long-term use.The occurrence of defects will lead to the reduction of driving comfort and driving safety,and also shorten the service life of the road.Therefore,defect treatment is an important part of road maintenance.However,there are still some shortcomings in the existing research on pavement defect detection,such as the problem of missed detection of small-size defects when detecting pavement defects,and the problem of the inaccuracy of segmentation results caused by various noises when segmenting pavement cracks.Therefore,taking common pavement defects as the research object,using image processing algorithms and deep learning technology,a target detection algorithm for pavement defects and a semantic segmentation algorithm for pavement cracks are designed in this thesis.The main work of this thesis can be summarized as follows:Ⅰ.A pavement defect detection algorithm based on RFB and Soft-NMSAiming at the problem of missed detection caused by insufficient recognition ability for small targets and close distance between defects in current pavement defect detection tasks,a pavement defect detection algorithm based on RFB and Soft-NMS on the basis of the one-stage object detection network Yolov4 is designed in this thesis.Firstly,a detection branch on the original Yolov4 model for small targets is increased in this algorithm to improve the detection ability of the network to detect small-sized defects.Then,the original spatial pyramid pooling block of Yolov4 is replaced by the RFB module which has a larger receptive field to enhance the feature fusion ability of the network and maximumly retain the small pavement defect in the feature map.Soft-NMS is also used to improve the non-maximum suppression algorithm to reduce the missed detection of defect targets in the detection stage.Finally,the prior anchor is redesigned by the K-means++clustering algorithm to better match the size of anchors to the targets.Experiments on the RDD2020 data set demonstrate that the recognition accuracy of pavement defects can be improved by the designed detection algorithm and a F1-Score of 0.637 is achieved,which proves that the detection ability of the algorithm for small pavement defects is improved.Ⅱ.A pavement crack segmentation algorithm based on CGANAiming at the problem of inaccurate crack segmentation caused by the complex background of pavement and various disturbances in the current pavement crack segmentation task,a pavement crack segmentation algorithm based on CGAN on the basis of the semantic segmentation model U-net3+is designed in this thesis.U-net3+with the attention module is used in the generator to generate segmented images for pavement cracks.Crack features are highlighted and noise features are suppressed in both channel and space dimensions by the attention module,and more complementary crack features are obtained through cross-dimensional feature fusion.The original image is stitched with the manual annotation of cracks and the generated segmented image as the input of the discriminator.The PatchGAN method is used in the discriminator.Moreover,a weighted hybrid loss function is designed to improve the segmentation accuracy by exploiting the difference between the generated and annotated images.Through alternating gaming training of the generator and the discriminator,the segmentation image of cracks generated by the generator is very close to the actual segmentation image,thus achieving the effect of crack detection.The experiment results using the Crack500 datasets demonstrate that various disturbances can be eliminated by the designed method and superior performance in pavement crack segmentation with complex backgrounds is achieved.Ⅲ.The software design of the pavement defect detection systemThe software of the pavement defect detection system is designed and the software function is realized in this thesis.Firstly,to speed up the running speed of the network when it is called,Tensor RT is used to accelerate the reasoning of the designed algorithm model.After that,according to the structure of the pavement defect detection system,the software of the detection system is designed.The pavement defect detection algorithm and the pavement crack segmentation algorithm designed in this thesis are adopted.The interface design of the pavement defect detection system is realized by using PyQt5,which provides a graphical interface for users to operate and use conveniently.In summary,aiming at the problems existing in the current pavement defect detection algorithm,a pavement defect detection algorithm and a pavement crack segmentation algorithm are designed respectively in this thesis.The superiority of the two algorithms is verified by experiments.The software design of the pavement defect detection system is completed,including using Tensor RT to accelerate the model and using PyQt5 to design the user interface.