Research And Implementation Of Pavement Early Diseases Detection Algorithm Based On Deep Learning

With the rapid development of logistics,freight transportation and other industries,the increase in transportation volume,and the impact of environmental factors,road damage and diseases,mainly cracks and potholes,are increasing.The traditional manual detection methods have the disadvantages of low efficiency,low safety,and subjectivity in the detection results,which can no longer meet my country’s increasing maintenance highway detection requirements.In addition,the traditional manual detection method has the disadvantages of low efficiency,low safety,and subjectivity in the detection results,which can no longer meet the increasing requirements of maintenance highway detection.In addition,the automatic system detection algorithms used by national and provincial roads are susceptible to road conditions and illumination,and there are still shortcomings and shortcomings in the accuracy and scalability of the algorithm.Moreover,the detection system has not been popularized on urban and rural roads,so early road diseases Inspection is still a practical engineering problem that the highway management department urgently needs to solve.Therefore,in order to achieve accurate,fast and low-cost detection,an early disease detection algorithm based on deep learning is researched and improved,and the overall detection process of the software system is introduced,aiming at improving the efficiency of urban and rural road inspection and reducing the cost of later road repairs.The main content is depicted as follows:(1)Datasets.At present,there are few public data sets related to pavement diseases in cities and towns,and they mainly focus on crack detection.The detection results don’t cover common types of early pavement diseases,which can’t meet the actual road scene detection requirements.Therefore,by using mobile smart phones,private car driving recorders and other tools to obtain a large number of pictures of pavement diseases in the urban and townships of Fuyang,and divide the common early road damage diseases into 4 categories:(1)horizontal cracks;(2)longitudinal cracks;(3)crocodile cracks;(4)potholes.To avoid the over-fitting phenomenon of the algorithm model during the training process and improve the overall detection performance of the model,geometric transformation and color space enhancement are used to increase the number of datasets.(2)Algorithm model.Firstly,for the detection speed of the algorithm model,the YOLOv3 target detection algorithm is used,and the backbone network of YOLOv3,Dark Net-53,is replaced with a lightweight network Shuffle Netv2.Secondly,for the detection accuracy of the model,measures were taken to increase the size of the DW convolution kernel,and the Repeat parameter was modified.Finally,aiming at the parameter scale of the algorithm model,improving the running performance of the algorithm model on the mobile terminal,using the convolution kernel clipping strategy to select a reasonable clipping rate to clip and compress the three detection branch parameters of the YOLOv3-Shuffle Netv2 model.Experimental results show that common early diseases on the road surface can be detected by improved algorithms with correct,speed and low cost.(3)Interactive interface of software system.Based on wx Form Builder,the interactive interface of the detection software system is designed and implemented.The interface includes the functions of detection source,detection of road disease types,and detection result recording.By calling the Open CV module,pictures,videos and real-time detection can be realized,and the common early disease detection on the road can basically be completed after the test.Finally,the function of each module of the whole system software is introduced the functions of each module of the system software are introduced.