Research And Implementation Of Road Surface Damage Analysis Algorithm Based On Deep Learning

In recent years,a considerable amount of data has been accumulated in the field of road infrastructure.How to analyze them effectively so that their value can be reflected is a problem that urgently needs to be considered and resolved.Roads have a great impact on people's daily lives.If road surface damage can be detected as early as possible and reasonable road maintenance programs can be formulated,then maintenance costs can be greatly reduced.Deep learning is a mainstream solution to the problem of object detection and image classification.The task of road surface damage detection is suitable to be solved by deep learning algorithms.Relevant research in recent years shows that mainstream target detection algorithms can achieve good results on road surface damage detection tasks,and Faster R-CNN has achieved the best detection accuracy.However,they all ignore the high inter-class similarity characteristic of the road surface damage dataset.In addition,the super-resolution technology can better enrich the details of low-resolution images,and it shows high use value in many fields.Therefore,the super-resolution technology can be applied to road surface damage detection tasks.Based on the above background,this topic has mainly done the following three tasks:(1)A road surface damage detection algorithm Faster R-CNN_INF based on convolutional neural network is proposed.Faster R-CNN INF uses Faster R-CNN as the basic network,and optimizes the basic network from the network structure,parameter configuration,and dataset characteristics.The most important optimization part is to propose IouNmsFilter,which is an improved NMS and candidate bounding box-based IoU filter module.IouNmsFilter improves the F1 value of the road surface damage detection by rationally utilizing the rich IoU information between similar road surface damage categories..(2)A road damage detection algorithm combined with super-resolution technology is proposed.The task of Faster R-CNN_INF is limited to damage location,and then the cropped suspected damage area image is optimized by super-resolution technology and then sent to the image classification to predict its category,and finally get better damage detection results(3)A road foundation big data analysis platform was designed and the corresponding prototype system was implemented.The big data analysis platform allows front-line engineers to use common machine learning algorithms,ETL algorithms,and the road damage detection algorithm proposed in this paper to assist in solving real-world problems.