Convolutional Neural Network-based Crack Detection Method For Cement Pavement In Remote Sensing Images

Cracks of varying sizes and depths can occur over long periods of time due to natural and man-made conditions,and these cracks can cause some disruption to the comfort and safety of vehicles during travel,so the task of capturing cracks and identifying and detecting them is urgent.Improving the accuracy of crack detection remains a challenge due to the fact that cracks have no fixed shape and their texture characteristics can change during photography due to the effects of lighting.Traditional pavement crack detection models have certain problems.For example,crack edge information can be lost during filtering or pooling of the image;shadows in the image due to low contrast of the surrounding pavement or poorly selected angles of the shot.As a result,research into the accurate detection of cracks in pavements has received a great deal of attention.(1)In order to improve the integrity of crack detection edge information,this paper proposes a new network structure based on convolutional neural network,which fuses contextual information between the encoder and decoder through feature pyramid operation,so as to come to realize the transmission of feature information from higher to lower layers and ensure the integrity of crack edge information.In order to address the problem of unbalanced contribution of easy and difficult samples to the loss resulting in the network not learning effectively from misclassified samples during the training phase,samples are reweighted by nesting them in a hierarchical manner during the network training,which is used to balance the contribution of easy and difficult samples to the loss,thus improving the accuracy of crack detection.After evaluation,we found that the method outperformed the state-of-the-art crack detection,edge detection and semantic segmentation methods on the three crack datasets,with an average cross-merge ratio improvement of 17.0%,15.5% and12.1% respectively.(2)To improve the detection accuracy of pavement cracks,we added a Dropout layer,a multilayer output fusion technique and a bridging unit to the convolutional neural network.The Dropout layer is used to increase the generalisation capability of the network,the multilayer output fusion technique to improve the accuracy of the model detection,and the bridging unit uses multiple topologies to optimise the sensing range of the network so as to capture the various subtle changes in the cracks more effectively and to better extract local and contextual information.To validate the performance of the designed network,the results of this paper show an average cross-merge ratio improvement of 4.3%,9.4% and 3.7% on the Crack500,Crack200 and Pavement images datasets,respectively,compared to the U-Net network.Based on convolutional neural networks,the proposed method for detecting road cracks in this paper can achieve fast and accurate detection of road crack locations,providing a basis for determining the risk level of road cracks,as well as for road maintenance and road safety.