Research On Asphalt Pavement Crack Image Detection Algorithm Based On Digital Image Processing

At present,China’s road traffic development has transitioned from the period of centralized construction to the phase of equal emphasis on construction and maintenance,and it is urgent to pay attention to the detection and prevention of pavement diseases.As the initial manifestation of pavement diseases,crack diseases are of great significance in highway maintenance.However,the traditional pavement crack detection algorithm can not be applied to a variety of pavement conditions,the algorithm is not universal,and it is often difficult to achieve the desired results.Therefore,it is necessary to carry out further research on the automatic detection algorithm of pavement cracks to make it more accurate,efficient and stable.This thesis combines emerging deep learning algorithms,with pavement crack detection and recognition as the main line,focusing on the noise reduction processing of pavement crack images,image sample expansion,crack semantic segmentation and crack target detection.First,for the mixed noise existing in the road surface crack image,a hybrid noise reduction method based on the combination of variable window shape median filtering and adaptive wavelet layered threshold noise reduction is proposed,while filtering out the noise in the image,the detail edge information of the crack is kept as much as possible.Secondly,in view of the insufficient sample size of pavement crack images in deep learning,a generative adversarial network model that can generate a resolution of 224 × 224 is proposed.Using noise-reduced pavement crack images as the real sample set,a large number of high-quality crack images with a high degree of reality are generated to expand the data set,which ensures the accuracy of subsequent image processing tasks.Next,for the segmentation of small targets such as cracks,an improved Deep Lab V3 + semantic segmentation model is proposed.To extract the cracks from the background more completely,we have improved its DCNN module,and added another fusion of deep semantic information and low-level features to the decoding structure.At the same time,we analyzed the impact of training parameters on model accuracy,and selected the best optimizer and initial learning rate from it.Then,for the classification and location of cracks,an optimized Faster R-CNN target detection model is proposed.We have optimized the structure layer of the original model,and selected the best 9 anchor points by comparing the evaluation results of 30 different anchor parameter combinations on the test set.Pavement crack images after semantic segmentation are divided into three categories: transverse cracks,longitudinal cracks and alligator cracks.Finally,a method for calculating the geometric parameters of linear cracks and alligator cracks is proposed.The mean error between the calculated geometric parameters and the real values is further proved that the algorithm in this paper has good practicality and feasibility for the recognition of road crack images.In summary,the automatic detection algorithm of road crack image based on digital image processing proposed in this paper has the characteristics of strong versatility and good detection effect,which can provide reference for the further study of the road crack detection system and provide a scientific basis for the maintenance decision-making of management department.