Research On Cross Dimensional Interactive Significance Detection Of Road Cracks

In recent years,China’s economy has developed greatly,and expressway has become an indispensable scheme for people to go out.The crack is the biggest hidden danger of the road.In order to ensure people’s travel safety,regular crack inspection on the expressway has become an essential link.For road crack detection,this paper mainly includes:(1)road crack image acquisition,preprocessing,and making road crack data set(crack data set);(2)According to the characteristics of different kinds of road cracks,different modules are designed to construct a new road crack significance detection model:GCLNet network;(3)The network performance is evaluated and analyzed on four standard data sets and crack data sets.The road crack image is collected by UAV acquisition system.Because the collected images are blurred and can not be directly used for network training,preprocessing operation is needed.Deblurring the collected pictures,and then cutting them to obtain a standard picture with a resolution of 320*320.Then label and label the standard pictures with labelme labeling software to obtain the.json format file.After further operation,the Mask data set is obtained.Finally,the Mask data set is expanded by data expansion methods such as rotation and mirror image to generate road crack data set(crack data set).Aiming at the characteristics of various road cracks and the small number of images under different complex backgrounds,gclnet network extracts a large number of crack basic texture information through the basic feature extraction module.Then the CF in the effective feature module performs channel and spatial cross-dimensional interaction to enhance and form high-level features.In addition,the basic feature highlight area is enhanced through the CL module to form a global context feature.Finally,through the aggregation module,the three road crack characteristics of basic,advanced and global context are aggregated to form a significant crack prediction map.In order to embody the advanced nature of GCLNet network,experiments were conducted on four standard datasets:ECSSD,HKU-IS,DUT-OMRON and PASCAL-S.Compared with other 9 advanced networks,GCLNet network has the highest quantitative evaluation on ECSSD data set,with Fβ reaching 0.948 and Sm reaching 0.929,indicating the effectiveness of GCLNet network in extracting significant objects in complex background.In addition,in the road crack data set,compared with the other three advanced networks:GCPANet,U2Net and F3Net,the Fβ and Sm of GCLNet network reach 0.720 and 0.790 respectively,which are increased by 3.6%and 8.9%respectively compared with the second highest network.Thus,the generalization ability of gclnet network is further verified.Figure[39]table[6]reference[84]...