Crack Detection Of Concrete Pavement Based On Deep Learning

With the continuous improvement of our country’s highway transportation construction,the transportation network extends in all directions,and the focus of transportation construction gradually shifts to road maintenance.As an important material,concrete is widely used in road construction.Due to the long-term impact of load bearing,harsh environment erosion and construction process quality,the concrete structure facilities are prone to cracks and diseases on the road surface and shorten the service life of concrete facilities.Huge potential safety hazards.Traditional detection methods rely on manual visual detection or use non-destructive sensors for auxiliary detection.The detection personnel are required to have relevant technology and experience.However,this detection method is inefficient,expensive,and highly subjective,which is not easy Accurate measurement of cracks can easily lead to misjudgment of crack types.In a complex background,the problems of difficult image detection of small cracks,more noise interference and easy loss of crack width information are likely to occur.Algorithms based on traditional digital image processing have weak denoising ability and poor generalization effect.In response to this problem,in this subject,the deep learning method is used to identify concrete cracks,and the semantic segmentation method is used to detect cracks.The main work content and innovations of this paper include:(1)U-net’s framework structure is used to embed the residual network layer in it to solve the network degradation problem caused by the deepening of the number of network layers,and the Batch Normalization layer to improve the gradient dispersion problem.Incorporating deep separable convolution and high-scale transposed convolution to realize the transfer of feature information from shallow to deep.Add an improved attention mechanism,extend the length of the U-net feature vector,expand the network depth,and add a layer fusion module constructed by the largest pooling layer,small-scale depth separable convolution and upsampling layer at the bottom of the U-net frame.The experimental results show that under objective standards,the accuracy value of the neural network reaches 0.9908,the value of jaccard reaches 0.7973,the value of recall reaches 0.9176,the value of Precision reaches 0.9226,and the F1-score reaches 0.9198,which reduces the phenomenon of false segmentation.Complete the segmentation of small cracks and obtain more detailed crack width information.(2)Construct a dual-input interactive connection network model based on multi-feature fusion to detect cracks.The model is based on the "encoder-decoder" structure,which connects the two mirror-symmetric networks "Chain 1" and the network "Chain 2" to achieve cross-transmission of feature information and the fusion of high-level feature information.Dual input introduces more low-level feature information,and adds a pyramid pooling layer built by multiple transposed convolutions with a convolution kernel size of 1×1.The experimental results show that,under objective standards,the accuracy value of the neural network reaches 0.9930,the value of jaccard reaches 0.8399,the value of recall reaches 0.9373,the value of Precision reaches 0.9389,and the value of F1-score reaches 0.9382.Experimental simulation results show that the network reduces the phenomenon of false segmentation and obtains more precise crack width information.The two concrete crack detection models in this paper are both proposed based on convolutional neural networks,using a series of methods such as feature fusion,so that the crack pixels can be more accurately classified and located.Therefore,the two concrete crack detection models proposed in this paper can obtain more accurate detection results under the premise of providing calculation speed and accuracy.