Research On Pavement Crack Detection Based On Deep Learning

In recent years,with the increase of global road ownership,the research of pavement crack detection has also been widely concerned.Although the pavement crack detection algorithms proposed in recent years show excellent performance,most of the detectors will detect cracks,even"all black",and the crack edge detection results are generally smooth.Based on the encoder decoder network,in this paper,a continuous attention mechanism and a multi-layer convolutional pyramid structure are proposed.In order to solve the problem of crack continuity,this paper integrates continuous attention mechanism into the encoder network.When the contrast between the pavement crack and the surrounding area is low or the crack itself is relatively thin,the detector will usually identify this section of crack as the background,so that the crack in the detection result appears fracture or"all black".After introducing the attention gating mechanism composed of continuous attention mechanism between each encoder layer,the features extracted from the current encoder layer can be weighted with the attention parameters before entering the next layer to enhance the crack features and suppress the background features.After continuously passing the attention gating mechanism,the crack features are continuously strengthened,so the loss of crack features is reduced when entering the encoder layer.Thus,the integrity of these crack areas can be guaranteed during decoding.In order to solve the integrity of crack edge information,a multi-layer convolutional pyramid structure is integrated between encoder and decoder.To deal with the loss of edge information,this paper uses the feature pyramid structure in edge detection.Pyramid structure can fuse multi-scale information to ensure the integrity of global features.Therefore,in this paper,the output feature map of each layer of the encoder is constructed into a feature pyramid structure,and then a series of convolution pooling operations are used to fuse the features as the output features of the encoder and the input features of the decoder.In order to fully integrate the two modules of continuous attention mechanism and multi-layer convolutional pyramid structure,this paper adopts the pseudo twin input and dense connection mechanism.On the one hand,one end of the pseudo twin input generates attention parameters,on the other hand,it connects with the other end of the input through dense connection,that is,it is connected with the coding network for weighting.The propagation path of each attention parameter can be regarded as a pyramid structure feature fusion process.In addition,the decoder is also very important for feature integrity.In this paper,residual blocks are integrated into the decoder layers.The residual blocks are generated by the current decoder layer and the corresponding encoder layer to guide the next decoder layer.Through quantitative experimental analysis,segmentation and edge detection indicators are used for evaluation.Compared with the latest method,the mean intersection over union of crack200,crack500 and crack forest datasets is increased by 26.51%,7.74%and21.61%respectively.On the crack forest dataset,the accuracy,recall andF₁-measure are improved by 3.63%,9.57%and 8.51% respectively.