Research On Road Pavement Defect Detection Method Based On Data Enhancement

With the mass construction and use of the road traffic system,the maintenance of the road has become more important than ever,among which the detection of road pavement defects is the top priority.Large amount of research has been carried out to detect multiple types of defects in road pavements,but the existing methods can only work in normal illumination environments and the detection performance is limited.As a matter of fact,road pavement defect detection often is carried out low-illumination working environments such as in tunnels,under flyovers and emergency night-time inspection and repair.In addition,mobile devices are becoming the mainstream road pavement defect detection equipment,and existing high-calculus methods cannot meet the hardware configuration requirements of the equipment.Therefore,it is of great practical significance and research value to study high accuracy,high real-time,lightweight road pavement defect detection methods that can adapt to different illumination working environments.To address the above problems,this thesis proposes a data augmentation-based road pavement defect detection method,and verifies the comprehensive performance of this method through extensive experiments and analysis.The specific research is as follows:(1)Acquiring new images and enhancing the original images with various data enhancement methods to supplement and improve the open normal illumination pavement defect data set,a high-quality normal illumination pavement defect data set RDD-W is obtained to meet the needs of this research.(2)To address the problem of missing of images due to the low-illumination and the difficulty of manual annotation,this thesis proposes a simulated low-illumination pavement defect image generation method—Cycle GAN-CLAHE from the perspective of data enhancement,which can transform the style of the annotated normal illumination pavement defect dataset to RDD-W low-illumination style.The obtained dataset RDD-B can be used to train road pavement defect detectors which can effectively improve the detector’s ability to detect multiple types of defects in lowillumination working environment.And the data set RDD-WB containing the normal illumination pavement defect images and the simulated low-illumination pavement defect images is obtained by combining the data set RDD-W with the data set RDD-B.This data set is used to train road pavement defect detectors that can adapt to different illumination working environments.(3)Considering the characteristics of road pavement defect images and the attention mechanism,this thesis proposes the C3-BoT module based on the selfattention mechanism,the C3-BoT module and an another attention module,NAM,are introduced into the YOLOv5 s network,so that the network can suppress the interference information in the image while establishing the dependency between the overall information of the image and focusing on the defect information,which improves the detection performance of the road pavement defect detector.The GSConv module,can efficiently sample and assign features,is introduced into the YOLOv5 s network to achieve a lightweight network structure,making the network-based road pavement defect detector more light while improving the detection performance.The experimental results show that the proposed method for generating simulated low-illumination pavement defect images can generate high-quality simulated lowillumination pavement defect images with good usability and practicality.Based on the comparison of the method presented in this thesis with the YOLOv5 s method,the obtained road pavement defect detector meets the lightweight requirement while improving the average detection accuracy of four types of defects: transverse cracks,longitudinal cracks,mesh cracks and potholes in different illuminated working environments by 1.7%,reaching 75.9%.