A Method for Pavement Crack Segmentation Based on Multi-scale Cavity Convolution Structure
In order to realize the automatic detection of road cracks and improve the problems of segmentation discontinuity and false segmentation of noisy background in the existing crack segmentation model,a crack segmentation model MAC-UNet based on multi-scale cavity convolution structure is proposed.Taking UNet as a basic network,the multi-scale cavity convolution structure is firstly proposed to replace the double convolution structure in the encoder and decoder,that improves the segmentation performance of the network for complex topology structure.Then,the cross-attention mechanism is constructed,and the pyramidal attention module is used to replace the jump connection between the encoder and decoder.The spatial characteristics lost due to pooling are retained,and the channel attention is increased to guide the multi-scale information to be effectively integrated into the decoder characteristics,so that when the crack is restored,the details are more abundant,and the positioning is more accurate.Finally,the experimental comparison with FCN,PSPNet and other 5 methods is carried out on the road crack data set CFD and GAPS384.Compared with FCN,PSPNet and other 5 methods on the road crack dataset CFD and GAPS384,MIOU and Kappa coefficient increase by 8.4%and 8.52%respectively on the CFD dataset compared with UNet.On the GAPS384 dataset,it is improved by 6.84%and 8.23%respectively,and the segmentation of road cracks is clearer and more complete.The result shows that compared with the mainstream segmentation algorithm,the proposed algorithm has obvious advantages in recognition accuracy.Under the conditions of uneven illumination,various noise interference and different background gray levels,the proposed model can still obtain stable detection results and can deal with complex crack segmentation problems.The visual crack detection error is small,which meets the actual engineering requirements,and the model volume is small,which has certain engineering application value.
road engineeringcrack identificationdeep learningroad crackcavity convolutionmulti-scale features
NETL
NSTL
万方数据
