Application Of Convolutional Neural Network And Generative Sample Augmentation Method In Crack Detection In Pavement Image

With the continuous construction of China’s road network,the demand for road maintenance is increasing day by day.Pavement crack detection is of great significance for scientific maintenance of pavement and the strategy for developing a quality transportation.Although deep learning-based crack detection methods are mainstream methods in the field of pavement image crack detection,the detection performance of most methods relies too much on pixel-level annotation and it is difficult to avoid the problem of imbalance of positive and negative samples.As a result,it is difficult to obtain satisfactory test results in practical applications.Therefore,exploring a new crack detection method in pavement image,overcoming the pixel-level labeling dependence,effectively amplifying positive samples,and solving the problem of imbalance between positive and negative samples has important theoretical and practical significance for improving detection performance.The main work of this thesis is as follows:1.A staged crack detection method based on visual attention and convolutional neural network is proposed.The method consists of three stages: salient area detection,image patch classification and regional growth post-treatment.Firstly,the significant area is obtained by the graph-based visual saliency(GBVS)detection model,and the coarse detection of crack area is realized.Then,based on the image patch classification method of convolutional neural network,the coarse detection results are eliminated the false and retained the true.Then,the final crack detection results are obtained through regional growth post-treatment.On the premise that only image-level annotations are required,accurate detection of cracks from coarse to fine can be achieved,which improves the problem of annotation dependence to a certain extent.2.Aiming at the problem of sample imbalance,a positive sample augmentation method based on autoencoder(AE)and deep convolutional generative adversarial networks(DCGAN)is proposed.The entire model consists of two encoders,a generator and a discriminator.By minimizing the adversarial loss,reconstruction loss,and minimizing the feature distance between the input image and the generated image in latent space,the optimal deep generation model is obtained,so as to automatically generate a sufficient number of positive sample images of cracks,and significantly improve the problem of sample imbalance.3.The performance of the two methods is tested separately using public datasets.First,verify the performance of the staged crack detection method based on visual attention and convolutional neural network in the crack detection experiment of pavement images,and prove its effectiveness;secondly,verify the performance of the staged crack detection method incorporating positive sample augmented images,verifying the feasibility of the positive sample augmentation method based on AE and DCGAN model in solving the problem of imbalance between positive and negative samples.