Research On Intelligent Cracks Detection Algorithm Of Asphalt Pavement Based On Convolutional Neural Network

Cracking is a common distress on asphalt pavement,and it affect the driving safety and comfort.Finding and fixing a crack in it early stage can prolong the pavement serviceability and save maintenance budget.Therefore,obtaining the crack information timely and accurately is important in pavement maintenance.Convolutional Neural Networks(CNNs)has proved to be the state-of-the-art technology in computer vision.However,it’s still a challenging task to use CNN for pavement crack detect,because quantity imbalance between crack and non-crack pixels and the crack-like pattern in the pavement images impede the CNN to achieve a higher accuracy.To solve the above two problems,this thesis proposes an encoder-decoder network(EDNet)and a policy gradient-based encoder-decoder network(EDNet—PG)respectively.EDNet solved the quantity imbalance between crack and non-crack pixels,and EDNet—PG overcomes the crack-like pattern problem and improves the recognition rate of fine cracks.In addition,the current CNN-based pavement crack detection algorithms require a large amount of labor to calibrate the ground-truth images.Therefore,this thesis proposes a weak supervised learning-based U-Net for crack detection,which directly use the results of a traditional crack detection algorithm as the weak ground-truth images;it reduces the labor cost to calibrate the ground-truth images.This thesis improves the accuracy of crack detection algorithm and contributes to the development of the crack detection field,and carries out a series of research work as follows:(1)This thesis introduces how to obtain the rough ground-truth images for convolutional neural network training by using image processing technologies and the traditional crack detection algorithms.In the image pre-processing technologies section,this thesis introduces how to use smoothing algorithm to denoise the pavement image and how to use image contrast automatically increasing algorithm to improve the contrast of the cracks in pavement image.This paper introduces two traditional crack detection algorithms,they are matched filtering and 3D shadow modeling.In the image post-processing technologies section,the thesis introduces how to use matched filtering and connected components analysis to improve the continuity of small cracks and de-noising the result images.(2)Based on the current development of convolutional neural network,the related theories of convolutional neural network are introduced.This thesis introduces the forward-propagation of the network,which includes the principles of convolution,deconvolution,batch normalization,activation functions and skip-connection.When discussing the training methods of the network,this thesis introduces the definition of several loss functions and their advantages and disadvantages.In the back-propagation part,this thesis introduces how to use backward mode automatic differentiation to calculate the parameter gradient of the network.Finally,the thesis discusses several common optimizers,and how they updating the network parameters is detailed.This section lays a theoretical foundation for the proposed encoder-decoder network(EDNet)and the policy gradient-based encoder-decoder network(EDNet—PG).(3)This thesis presents an encoder-decoder network for pavement crack detection.EDNet effectively solves the quantity imbalance between crack and non-crack pixels.For150 3D testing images,the proposed EDNet achieved 97.36% overall precision,98.24%recall,and a 97.80% F1-score.For 46 2D pavement testing images,an overall precision of96.15%,recall of 99.56%,and F1-score of 97.82% were achieved by EDNet.The comparative study shows that EDNet outperforms other methods.Due to the simple architecture of the EDNet,its processing speed is fast.(4)This thesis proposes to use policy gradient technique to reduce the false-negative errors in networks.EDNet is further trained for only 1 epoch by using policy gradient.The trained network is named as EDNet-PG.Compared with EDNet,EDNet-PG sacrifice some of the precision(no less than 95% on validation images),but it has a higher recall in return.A higher recall implies that the EDNet-PG will be more sensitive to fine crack and produce few false-negative errors.The trained EDNet-PG achieves an overall precision of 95.27%,recall of 99.19%,and F1-score of 97.19% on 150 3D testing images.EDNet-PG greatly reduce the false-negative errors in EDNet,and it anti-noises capability is not weaken too much.Therefore,EDNet-PG further improves the performance of EDNet.(5)This thesis proposes a policy gradient-based CNN weak supervised learning technique for pavement crack detection.This method directly uses the results of a traditional crack detection algorithm as the weak ground-truth images,and the weak ground-truth images in the training set don’t need any manually calibration.Therefore,this method reduces the labor cost to calibrate the ground-truth images.By making the rule for rewarding and punishing the network gradient,the network cost function is designed,and the network has finally achieved good results.However,this method is still in its infancy stage,the performance of the trained U-Net still has a certain gap with that of EDNet and EDNet-PG.The trained U-Net achieves an overall precision of 94.06%,recall of 96.64%,and F1-score of 95.33% on 150 3D testing images.The comparative study shows that trained U-Net outperforms matched filtering algorithm and three-dimensional shadow modeling.