Research On Pavement Crack Detection Method Based On Multi-scale Convolution Features

With the change of our country’s highway from the construction era to the maintenance era,it is accompanied by the scientific,intelligent and precise maintenance management of the road surface in a large range.At present,there are some outstanding problems in road maintenance management in our country,such as the low degree of maintenance refinement,the low level of management information and intelligence.The main technical bottleneck in this field is the intelligent detection of road diseases and the accurate assessment of damage conditions.Pavement disease detection is an important part of road intelligent maintenance decision-making.As the dominant type of pavement disease,the realization of rapid and accurate automatic pavement disease detection system is very important for the maintenance and monitoring of a wide and complex transportation network system.In the traditional task of pavement disease detection,the use of artificial methods is entirely dependent on the knowledge and experience of experts,which is inefficient,labor-intensive and lack of objectivity in quantitative analysis.Therefore,it is of great theoretical significance and practical value to realize the automatic extraction of pavement cracks in the field of highway pavement quality detection,which is the current research hotspot.However,due to the strong heterogeneity,high topological complexity and large noise similar to the crack texture,it is a great challenge to complete the automatic crack extraction and classification.In view of the existing automatic crack detection algorithm in large-scale application,especially in the complex environment such as wide area and multi road conditions,there are serious problems in the stability of the algorithm.Based on the deep learning theory,this paper realizes the automatic detection and classification of crack pixel level by learning the advanced features of crack representation,realizes the intelligent automatic detection of road diseases,and meets the new requirements of intelligent maintenance and fine maintenance.The main contents of this paper are as follows:1.When using supervised learning methods such as deep learning to extract road diseases,it is necessary to build a training sample data set that is accurate to pixel-level,and the limited samples of labeling sequence limit the training of more complex deep model.Aiming at the problem that the input cost of high-precision crack data set is too large,this paper proposes a semiautomatic crack labeling algorithm based on the improved C-V model by analyzing the basic characteristics of pavement interferences and cracks,which realizes the automatic initial labeling of pavement cracks.Then the initial annotation results are used as samples for manual interactive annotation,which can effectively reduce the workload of manual annotation.2.The automatic crack detection method in complex background.This paper proposes an end-to-end crack extraction network structure based on the full convolution network to achieve pixel level extraction in crack image.In the automatic fracture extraction network,a new multiscale expansion convolution module is proposed to capture the fracture context information in multiple scales.On this basis,the attention mechanism is used to further refine fracture features.The multi-scale attention module can learn fracture features in complex background with stronger discrimination and robustness.In addition,through the fusion of low-level features and high-level features,an effective feature fusion sampling module is proposed to obtain more precise segmentation results.3.Based on the results of pixel level extraction of pavement cracks,according to the definition of pavement disease classification in China,the paper studies the algorithm of connected area marking and the characterization description of each crack type,and divides the disease types into transverse crack,longitudinal crack,tortoise net crack and block crack by definition rules.In this paper,the average width and distance between cracks are given to evaluate the severity of various cracks,and a new algorithm and evaluation mechanism of severity classification are proposed to realize the automatic classification and evaluation of pavement cracks.It provides decision support for intelligent and efficient identification and evaluation of pavement diseases.The method can effectively extract the crack information in complex background.The crack extraction Precision,Recall,F-score and m Io U index can reach 98.84%,97.05%,98.39% and 74.81% respectively.At the same time,the model has strong stability and robustness,which can effectively solve the problem of shadow,stain and exposure noise in the process of data acquisition.Compared with the result of manual classification,the accuracy of this method is higher than 95% for linear fracture and 86% for reticular fracture.