Algorithm Design Of Classification And Extraction Of Asphalt Pavement Crackings Using Pavement 3D Images

Due to pavement cracking regarded as the early symptom of pavement detections or structural deterioration,prompt treatment of cracking may minimize loss caused by it.As an alternative method to traditional manual inspection in field,automatic pavement distress inspection system becomes a main technical mean both in acadymic circles and industry.The-state-of-art automatic collecting pavement distress systems reach a relatively high level where there is no difficulty to digitize pavement surface distress,but rarely can they report most of distresses happening in road surface even if only one,such as pavement crack,due to lacking of distress extraction algorithms with high performance.As a result,manual intervention is needed at the stage of post-processing pavement distress data.Hence,this paper proposes a hybrid procedure for automatic detection and classification of pavement crack with 3D pavement data.The technologies or theories involved in this paper include digital image processing,convolutional neural network(CNN),tensor voting(TV),crack extraction and enhancement,aiming at improving the accuracy of identifying crack and providing a novel way to classify crack images among five different types.The purpose of these proposed methods is to advance the full automation of pavement distress inspection a little further.A detailed literature review is presented through collecting the research articles done by both domestic and foreign researchers in the field of pavement distress inspection,and then points out some disadvantages of crack segmentation and classification algorithms.As to these aspects needed to be improved,author proposes a series of reseach as following:1.Based on the characteristics of pavement 3D data,a criterion is proposed in order to smooth the step between left and right half images.Using multi-size 2D Gaussian filter,spurious noise is removed,and subsequently a big-size 2D Gaussian filter is employed to rectify the original 3D pavement data.In this way,the negative effects of original pavement 3D data such as large dynamic range,noise caused by vehicle vibration,or imaging system defect,wouldn't happen in the further procedurce.The preprocess methods demonstrated in the Chapter 3 increase significantly pavement 3D image quality.2.Using artificial neural networks theory,four convolutional neural networks are successfully trained to classify pavement distress images,and for the training and validation purpose,an image library is built up constituted of more than 28,000 pavement images.The networks structured with eight layers(except for input layer)are able to distinguish five kinds of pavement images,namely,non-crack,transverse crack,longitudinal crack,block crack and alligator.Via training and validation,it is proven that all the four networks can achieve an expected performance.3.Within the theory of neural network,hyperparameters optimization is a major issue.An orthogonal experiment is designed to optimize the combination of three parameters setting in the neural networks.Since the accuracy of neural network for classifying pavement images,iteration times,and running time during every epoch are the common concern when training neural networks,a study on them is carried out.Sixteen tests are determined with the assistant of SPSS,and executed.When extracting test results,a hybrid convergance criterion is proposed.After this practice,it is proven that the criterion is suitable for extracting test results.Using the method of intuitive analysis on the results,the significance of each factor respect to the accuracy,iteration times,and running time during every epoch is obtained.Meanwhile,the optimal combination of the three hyperparameters is also determined.After vertifying,the neural network structured with the optimal combination of the hyperparameters can reach a high accuracy of 99.5%.4.Based on the classification result,a hybrid procedure is designed to extract and enhance pavement crack from original pavement 3D data.The procedure constitutes four steps: 1)Preprocessing on original pavement 3D data;2)According to the classification result and characteristics in the crack area,a steerable matched filter bank(SMFB)is designed with multi-size and mult-orientation 2D Gaussion filters.A binary crack map with noise can be generated in this step.3)In order to enhance the crack map,Tensor Voting is utilized to fill the gap between crack fragments,resulting in a saliency map,and then the original binary crack map and its saliency map are excuted with OR operation;4)Postprocess on the crack map generated by the Step 3,aiming at remove the noise.The validation of the hybrid procedure is conducted over 200 pavement 3D data collected from different road sites.The results illustrate that the accuracy of crack extraction is between 84% and 97%,and its average value is 88.38%;Recall rate is between 85% and 99%,and its average value 93.15%;F-value is between 85% and 97%,and its average value is 90.68%.As a result,the hybrid procedure has the ability to recognize crack in the asphalt pavement 3D data with high accuracy and recall rate.All the patamaters in the algorithms are fixed,in the other words,the procedure works with full automation and has a good robustness.