Research On Crack Detection Algorithm Of Pavement Image Based On Deep Learning

Highway crack disease is a common pavement problem.If it is not discovered and dealt with in time,it is easy to cause traffic accidents.At present,the detection of road cracks is mainly done manually,but the disadvantages are high cost and low detection efficiency.Therefore,the advantages of applying drones to daily inspections on roads are simple,flexible and efficient,and can effectively reduce blind spots in inspections.At present,small target detection technology in aerial images has always been a difficult point in the detection field.The reason is that under low light,the details of the road image are severely lost and the contours are blurred.The picture background environment is complicated and the target size is small.This paper uses the detection method of deep learning to identify the road cracks.Compared with the traditional algorithm,deep learning can better realize the feature learning and the detection accuracy is higher.The research work of this paper is as follows.Most of the current road crack data sets use crack images with clear targets and a single background.Therefore,the trained model will have a problem of low recognition accuracy when detecting road crack images under complex backgrounds.Drones are used to collect crack images on the road surface under different lighting environments,and then the number of road crack images collected by data augmentation is expanded.And manually mark the cracks in the image to complete the construction of the data set.In a low-light environment,the acquired image will have the problem of blurred target contours and loss of details.First,convert the RGB color space of the image to the HSV color space,and perform a global logarithmic transformation on the light and dark component V;then use an improved weighted guide filter to perform edge-preserving filtering;finally,the color space is converted back to RGB through color restoration,and the result is Enhanced image.Through experimental comparison with other image enhancement algorithms,it can be proved that this algorithm has a significant effect on low-illuminance image enhancement.The enhanced image has no distortion,and the crack edge information is highlighted.In view of the small size of road cracks in aerial images,the traditional Faster R-CNN single feature detection method cannot accurately extract crack feature information.A road image crack detection algorithm based on improved Faster R-CNN is proposed.VGG16 is used as a feature extraction network,and a multi-scale fusion strategy is used to output feature maps of different scales from multiple convolutional layers.By combining the L2-Normalization method for fusion,the final feature map for detection contains more target location information and stronger semantic information.Introduce online difficult sample mining algorithm to make the training more sufficient and optimize the resolution ability of the detection model.The soft non-maximum suppression algorithm is used to avoid missed detection of adjacent and overlapping targets,which can improve the accuracy of detecting multi-scale cracks in road images.Experiments show that the algorithm can quickly and accurately identify road cracks.