Research On Pavement Damage Detection Method Based On Deep Learning

With the rapid development of China’s transportation industry,China has formed a spe-cific scale of road traffic network.But the vast traffic network has brought challenges to road supervision and pavement maintenance.The existing pavement defects are mainly cracks and some irregular defects.These defects reduce the service life of highways to a large ex-tent and pose a threat to traffic safety.Therefore,detecting and repairing these defects on time plays a vital role in highway safety.This thesis proposes a deep learning-based pavement defect detection method to address these problems,which can quickly and efficiently detect pavement defects in real time.The main research content includes five parts: pavement defect image acquisition method,pave-ment defect target detection,pavement defect image segmentation,pavement defect quan-titative evaluation,and pavement defect automatic detection system design.A pavement defect image acquisition method is proposed for the shortcomings of existing image acqui-sition methods,such as low cost and high efficiency.The technique introduces perspective transformation and combines image processing techniques such as target detection and im-age cropping to make the pavement defect image acquisition fast and efficient.The target detection dataset PD-Dataset and image segmentation dataset CRACK2000 are established by this method,which contains various pavement defect types such as cracks,potholes,and patches.They have complex background interference,which is more consistent with the actual road conditions and representative.The optimal YOLOv5 network with 91% m AP is used to detect pavement defect location in the target detection.U-MDN(U-shaped multiscale dilation convolutional network)is proposed in the image segmentation part.u-MDN uses U-Net as the backbone network and joins the designed U-MDM to complete the extraction of the depth features of the pavement defects.The experiments show that U-MDN outperforms U-Net in terms of Precision,Recall,F1-score,etc.Finally,various defects are quantified and calculated based on the results obtained from target detection and image segmentation,including the length,width,and coverage area of various linear cracks,the minimum exter-nal rectangular size of mesh cracks,and other quantization parameters.And based on these quantified parameters,the pavement condition index PCI is calculated,and finally,the road health level is obtained to complete the pavement inspection.The automatic detection of pavement defects is significant for pavement condition as-sessment.To provide a theoretical basis for pavement maintenance,the automated detection system designed in this paper can accurately locate and classify the location and category of pavement defects.Moreover,the proposed U-MDN can achieve pixel-level segmentation of defect areas in the image so that the obtained segmentation results are directly used for quantitative pavement evaluation.Therefore,the pavement image can now output the corre-sponding repair suggestions after this system,enhancing the practicality of pavement defect detection.