Research On The Algorithm Of Crack Detection And Classification Based On RCF

The concrete road will arouse serious diseases which affect the travel safety for the erosion of natural disasters and man-made damage over time.Crack is the early manifestation disease of concrete highway.Intelligent crack detection method can avoid the problems of low detection efficiency,long cycle and high cost caused by manual detection.Therefore,the rapid intelligent crack detection method research has certain research value and application value.In concrete pavement,both the relative merits of domestic and overseas crack detection algorithms and the crack characteristics have been analyzed in this thesis.In view of the shortcomings of existing crack detection algorithms,an improved and innovative method had been proposed.The main research work of this thesis included:1.The fine-tuned Le Net-5 model was used to classify the crack images.The image obtained by the intelligent crack acquisition vehicle was a set of continuous images.The image containing real cracks was the effective crack image,and the proportion of this part of the effective crack image in the total image set was very small.If every image in the image set had to be detected,a lot of time would be spent.To solve this problem,five types of image data sets(pseudo-cracks,cracks,plants,complete surfaces and artificial scratches)were set up.The fine-tuned Le Net-5 model could effectively identify crack images,so as to reduce the time cost of crack detection.2.This method modified the RCF network structure to improve the accuracy of crack detection.When the RCF model detected the crack image,the crack image was not clear and the crack edge was blurred.To solve this problem,the method of layer by layer fusion was adopted from high to low.Firstly,the scale features of the two layers were fused,and the feature images of the lower layers were fused with the fused images.Then,the third layers and fifth layers were fused by jumping fusion.Finally,the sampling layer on the two top layers was connected,and the convolution layer which was connected behind the Crop layer was used to fuse the feature map of the two-layer model to improve the edge blur.3.This method enhanced the feature expression ability of convolution layer through convolution and parameter setting.The depth of the RCF model was too shallow which would result in the insufficient accuracy of crack detection.In terms of this problem,the model adopted the method of horizontal expansion,and built a 5-layer continuous neural network model with convolution layer,Eltwise layer,Deconv layer and Crop layer connection layer.Then,more convolution kernels were used.The Conv3 layers of the model which was connected advance of the convolutional layer with kernel size and channel depth of 64,and the Conv5 layers of the model which was connected advance of the convolutional layer with kernel size and channel depth of 64 could enhance the feature expression ability of the convolutional layer.Finally,the final actual output corresponding to each pixel was obtained through the cross-entropy function.4.Used Cross-entropy loss function to improve the accuracy of fracture detection.In the neural network,the cross-entropy loss function was used to calculate the deviation between the target value and the actual output.In order to improve the precision of crack detection,having used the improved cross entropy function changed the output of neural network into a probability distribution,employed a large number of experiments got optimum value range of positive and negative samples,which could be predicted by cross entropy to calculate the probability distribution as well as the distance between the probability distribution and the real answer,and added balance parameters at the same time to improve the prediction accuracy.