Road Crack Detection Research Based On Image Segmentation Algorithm

Road crack reflects the quality of highway pavement,which directly affects the traffic safety and service life of highway.Therefore,early detection of highway road cracks and timely maintenance can effectively prevent the further deterioration of the crack,and keep driving comfortably of the vehicle on the highway pavement.It is of great significance to study the road crack detection algorithm to ensure the bearing capacity of the road surface and improve the ride comfort of the vehicle.The road crack image has the characteristics of many noises and complex topological structure.It is difficult to explore an algorithm that can accurately segment the crack area.First,the paper proposes a road crack approach based on gradient boost decision tree method,which makes full use of the facts that the visually salient road crack edge contains abundant diverse local structure.The paper takes the fixed size image blocks as training sample,and computes a number of features,then reduces them as an input feature vector.At the same time,the paper maps binary matrix of the correspondent ground truth edge to discrete score.Finally,the paper adopt the gradient boosting decision tree method to train the model,and the diverse ground truth binary matrix will be stored in the nodes of all decision trees.While applying to road crack detection,each pixel in the tested image will be calculated independently the likelihood whether it is a part of edge by the trained model.Then our proposed approach combines the results into global reasoning.After removing the noise point,the trained model can detect the road crack detection successfully.Next,the paper introduces the deep learning method into road crack detection.The deep and abstract features of the image are extracted by convolutional neural network.These high-level and low-level features will be combined.Then the approach increases the number of deconvolution layers,the same time reduces the step.After that the interpolated reconstruction is performed through deconvolution operation on the extracted feature map to obtain segmentation results.Compared to convolution neural network,this model can fully use the spatial information of road crack images.Experimental results demonstrate the state-of-the-art accuracy and high efficiency of the proposed approach compared with traditional convolution neural network methods.Finally,this thesis summarizes the work,and points out the direction for further research in the future.