Research On Expressway Crack Detection Based On YOLOv5s

With the development of expressway in China,the maintenance of expressway has received extensive attention.The appearance of pavement diseases will greatly affect the service life of expressway,so the realization of high-quality maintenance of expressway is a particularly important part of intelligent traffic management.As for the intelligent maintenance of expressway,the current research focus is mainly on digital image processing.However,due to the constraints of equipment performance and background environment,the collected image data are uneven,which leads to its inability to meet the actual engineering needs.In recent years,with the rapid development of deep learning,the method of target detection based on neural network has achieved good results in image detection.Based on the current research status of expressway crack detection,this thesis analyzes the characteristics of the collected crack image data in depth and carries out systematic research based on the current mainstream target detection algorithm YOLOv5 s.The main research contents of expressway crack detection are as follows:In order to weaken the influence of various environmental factors in the expressway crack data collection,the crack image preprocessing algorithm is first studied.The crack image is first grayed out,and the interference of pavement marker lines and other factors on the crack object is eliminated by using the gray-scale threshold segmentation method.Then,the adaptive median filter algorithm is used to denoise the image.Finally,the restricted histogram equalization method is used to enhance the expressway crack image to improve the contrast of the image.By analyzing the characteristics of expressway cracks,the original algorithm is improved to make it more effective in the expressway crack detection task.Firstly,the EIOU loss function is introduced to replace the original GIOU loss function when the prediction box and the real box overlap.At the same time,the channel attention mechanism is introduced on the basis of the original model algorithm to further improve the detection accuracy of the model.In order to reduce the size of the model and improve the detection speed of the model,the model pruning and sparse training methods are introduced,and the most suitable clipping rate is selected through the comparison of experiments with different clipping rates.The effectiveness of the improved method is further demonstrated through the ablation experiment by removing one or only retaining one of the above methods.