Pavement Disease Detection System Based On Deep Learning

Highway transportation is a very important component of economic and social development in recent years,so pavement disease detection is an important thing for highway maintenance,which is the key fact for highway transportation to save facility and highways.Only when the location,type and damage degree of pavement diseases are found in time,maintainers could choose the appropriate method to repair the highway,so as to eliminate the hidden danger.In China,the technology of road disease detection has not been applied extensively,and maintainers are often required to patrol for detection in traditional method.This method takes a long time,and it is inefficient and full of danger.To resolve the above problems,a pavement disease detection system based on deep learning was reported in this dissertation,the main research work is as follows:(1)Pavement diseases is analyzed deeply.In this paper,the types,definition and evaluation criteria of pavement diseases are studied in detail,and the image features of pavement diseases are analyzed.The analysis shows that these images have many noises,monotonous and uneven background and few feature points,and it is also hard to complete Image mosaic.According to these characteristics,this paper projects the overall process of pavement disease detection system.(2)The module of image acquisition and image mosaic is studied.Due to the few feature points of images and the difficulty of stitching,this paper adopts the multi camera structure: two industrial cameras are used to collect road information and roadside information synchronously.Then the perspective matrix was calculated by the extracted features of roadside images,and applied to the mosaic of road images.In these processes,the phase correlation method is used to calculate the coincidence degree of the collected images,and the redundant images are filtered to improve efficiency.(3)The image data set is constructed and calibrated.A total of 1467 images with pavement diseases were collected as data sets,the minimum external rectangular was used to accomplish the classification and positioning of pavement diseases.In this data set,881 training sets,295 verification sets and 291 test sets are used for network training,model fitting,network evaluation and testing separately.(4)The pavement disease detection and location system based on Faster R-RCNN is built to realize the recognition and location of pavement disease types in images.In order to achieve this,the convolution module of neural network is optimized by fitting the values around the filled image in this paper.After this,the loss and accuracy of the training set are calculated.Our experiment shows that the optimized convolution neural network could improve the accuracy of target detection results.(5)Pavement diseases are located.Roadside pile information is an important part of collected images,which can be used for spatial positioning calibration and cumulative error elimination.In these steps,the industrial camera is calibrated firstly,and the pixel equivalent is calculated to measure the special position of the pavement disease in the mosaic image,so as to realize the spatial location of these disease.