The Research Of Pavement Cracking Detection Technology Based On Mask R-CNN

In recent years,scholars have done a lot of research on the detection of road diseases based on pixel-level segmentation technology.Crack feature processing is carried out by analyzing the extracted image characteristics.The statistical significance of crack characteristics is classified according to its correlation.Existing automatic crack detection technology is more effective but more expensive(based on automatic vehicle detection technology),while the lower-priced scheme is less effective and time-consuming.Relying on professionals to check and evaluate road conditions is the easiest way,but not accurate enough.In addition,this approach requires significant human and time costs,especially in highway area work or complex weather conditions that put inspectors at risk.Combined with the above,automated road detection and management systems can efficiently detect and classify different types of road diseases.At present,most road maintenance services and transportation agencies have adopted automated road management and evaluation systems,which makes it possible for computer vision technology to develop and use in the field of road engineering,and related development programs can be combined with intelligent traffic systems to improve management efficiency.The road cracks concerned in this paper are considered to be the most common road defects,and researchers have put a lot of effort into the research of crack detection algorithms,especially in the detection and classification of cracks based on digital image processing.If the cracks in the road surface are not properly treated,it may lead to more serious road diseases.Therefore,it is of great significance to the road management organization to effectively detect the cracks in the road surface and to carry out reasonable maintenance before the structural diseases occur on the road surface.With the development of technology and the remarkable achievements of deep convolution neural network,it is possible to detect road diseases based on deep learning.For the road crack detection technology,the first need for better effect of disease detection and classification algorithm to locate individual diseases.Second,the algorithm must be able to distinguish between different types of crack data,including alligator,longitudinal and transverse cracks,which is considered to be the most common phenomenon of road diseases.Based on the above,with the excellent performance and widespread use of Mask R-CNN in the field of computer vision,the road crack detection technology based on Mask R-CNN is likely to be studied in this paper.In the face of the lack of resources and road diseases in Africa,the realization of this technology will greatly improve the level of road maintenance in The African region.In the experiment,we used two research methods: method one: two experimental comparisons.For the first time under the Windows operating system,the Tensorflow model framework is used with NVIDIA GTX 1650 and 8GB RAM hardware systems.The second is under the Ubuntu 16.04 operating system using CUDA,cu DNN,Docker NVIDIA GTX1050 and 8GB RAM.Both operations were performed using GPUs and the model was trained for 12 iteration cycles.Method two: Use NVIDIA 1650 GTX and 8GB of RAM under the Windows 10 operating system,combining Tensor Flow and Python.In the evaluation,the intersection threshold on the hemset is set to 0.5.The number of nodes is set at 80.In this paper,two convolutional neural network methods,especially Mask R-CNN as object detection and classification methods,and the use of boundary box information to show breakage and its type.Combining boundary box regression losses with classified losses for models,end-to-end processing of detection and classification issues.At present,there are fewer public data sets for road cracks.The real road crack data from manual selection of the experimental training data set has the characteristics of pavement images in different stages of degradation,so that the data set can be well represented.Regarding the results of a comprehensive experiment on challenging road damage detection,this experiment uses an evaluation data set to assess the accuracy of The Mask R-CNN application on standard images for road cracks.Based on the results,84% and more than 85% of recall rates and accuracy were achieved in the best detectable categories.The model advantages presented in this paper are,first,fast label data sets,followed by higher accuracy.Of course,the model has some drawbacks.The first is the training time of the two methods,which can be improved to some extent by reducing the data set.This was done primarily to reduce computational costs and to obtain results with fewer subsampling and a deeper architecture.In addition,the resources consumed by the feature extraction network remain limiting factors for the depth of the network and the size of its layers.Because convolution operations on the GPU can be greatly faster,it is best to limit the size of the network to the right range.