Research On Automatic Detection And Evaluation Method For Pavement Diseases Based On Deep Learning

As an important infrastructure supporting the development of the national economy,politics and culture,highways are an important part of the modern transportation system.With the use of a large number of highways in China,and the impact of traffic loads or natural factors,various diseases such as cracks,potholes and subsidence have gradually occurred.These diseases have seriously affected the transportation capacity,service life,and driving safety of highways,resulting in the lower bearing capacity of highways,and increasingly prominent maintenance issues.In the current pavement detection methods,the traditional detection methods mainly rely on manpower.These methods are not only time-consuming and labor-intensive,but also inefficiency,and easily cause traffic jams and safety problems.On the other hand,although the vehicle-mounted pavement detection system has high detection accuracy,it has the disadvantages of expensive detection equipment,difficult data transmission,and non-uniform technical standards.The detection objects are mainly focused on crack diseases,and less detection of other types of pavement diseases.At present,with the increasing maturity of deep learning technologies and the excellent performance of deep neural networks in extracting high-level features of images,the applications of image processing based on deep learning are becoming more widespread.Therefore,it is of great practical significance to use the deep learning technology in the field of artificial intelligence to explore and find a more convenient and efficient method for automatic detection and evaluation of pavement diseases.In this paper,deep neural network is introduced into the research of pavement disease detection and evaluation,and combined with the "Highway Technical Condition Evaluation Standard"(JTG 5210-2018),a unique method and idea is proposed.In this paper,six types of typical frequently occurring diseases in Zhenjiang,including transversal cracks ?longitudinal cracks?patches?block cracks?potholes and pitted surface,were collected through various acquisition equipment and multiple shooting methods.Build and train to realize the classification of pavement images,the detection of pavement diseases,and the evaluation of pavement damage.In addition,based on the actual road scene,the pavement disease detection and evaluation method are tested in this article.The main work of this article is as follows:(1)Constructed a pavement image dataset for deep learningBased on the collection of pavement image and feature analysis,this paper preprocesses the collected image through three aspects: uniform light processing,smooth denoising,and image enhancement to eliminate noise as much as possible and highlight the differences between diseased and non-disease areas.The quality of the image.On the basis of preprocessing,a pavement image data set was constructed by manual annotation,and the data foundation of the convolutional neural network model was obtained.(2)Designed a method for classifying pavement image based on convolutional neural networkIn order to improve the efficiency of pavement disease detection,a pavement image classification method for classifying pavement images of normal pavement without disease and damaged pavement with disease was proposed,and a classification network model was constructed.Through model training and verification,89.7% accuracy rate,86.9% recall rate,88.3% F-Score of normal pavement,91.3% accuracy rate,84.7% recall rate,and 87.9% FScore of damaged pavement were obtained.Experimental results show that the feature learning ability and overall performance of the pavement image classification model are better.(3)Designed a method for detecting pavement disease based on target detection modelBased on the classification,this paper uses the target detection model to realize the disease detection of damaged pavement images.In the feature extraction and recognition phase,for the performance of the network structure's weight initialization method,activation function,optimizer,and Dropout parameters,the optimal parameters are set and applied to the network structure.This article uses 2400 pavement image training and detection models.The recall rate of the images with the presence of pavement diseases is 70.5%,and the recall rate of the images with normal pavements is 71.2%.The accuracy rates for predicting the presence of disease and normal pavements are 82.8% and 55.4%,respectively.The average accuracy rate(AP)is 72.5%,which shows that the method in this paper has reasonableness and feasibility for the detection of pavement images.(4)Designed a method for assessing pavement damage based on semantic segmentationOn the basis of pavement disease detection,a method for evaluating pavement damage is also designed in this paper.Through semantic segmentation,the pixel area of the image of the diseased area is extracted and calculated,and then the pixel area is converted into the actual area through the coordinate unit conversion relationship.Combined with the "Highway Technical Condition Evaluation Standard"(JTG 5210-2018),the pavement damage is quickly and easily evaluation of.By training the SegNet network,this paper obtained a pixel accuracy of 0.8517,an average accuracy of 0.5978,and an average intersection ratio of 0.5840 on the validation set,which illustrates the feasibility of the model for semantic segmentation of pavement disease images.In addition,in order to verify the reliability and practicability of the proposed method in the actual road environment,this paper tests and discusses this method based on the above method.In this paper,pavement images of 4 road sections were collected under different time periods,different road sections,different weather and different lighting conditions,and the different methods were tested step by step according to the method described in this article.The test results show that the method of pavement disease detection and pavement damage assessment proposed in this article not only can detect the type of damage,but also calculate the index more efficiently through area conversion,which provides innovative ideas for the detection and evaluation of pavement damage information.The rapid and efficient detection and evaluation of diseases has certain guidance and application value.On the other hand,the overall recognition accuracy rate of the method proposed in this paper is still a certain gap from the 90% recognition accuracy rate that should be achieved in the "Highway Technology Condition Evaluation Standard"(JTG 5210-2018).The shortcomings of reliability and reliability suggest the direction of future improvement.