摘要
机场跑道裂缝形态多样、方向各异、长短不一且粗细不均,通常不具有统计规律。现有的各类裂缝分割算法难以在此类复杂场景中落地。针对上述问题,提出了联合self-attention与axial-attention的机场跑道裂缝分割网络(CSA-net),通过引入自注意力模块、轴向注意力模块、可变形卷积模块,提取裂缝的局部特征和全局语义特征。通过transformer decoder还原特征图的原始尺寸,融合了不同尺度间的分割结果,保留尽可能多的细节信息,使得CSA-net有更好的分割精度。在机场跑道实拍的数据集上进行的测试表明,针对裂缝的像素级分割指标F1-score达到了78.91%,高于目前各类裂缝分割算法。
The cracks on the airport pavement have various shapes,different directions,different lengths and uneven thickness,and they usually have no statistical laws.Existing various crack segmentation algorithms were difficult to implement in such complex scenes.In response to the above problems,a crack segmentation network that combines self-attention and axial-attention(CSA-net)was proposed.By introducing a self-attention module,an axial-attention module,and a deformable convolution module,the local and global semantic features of the fracture were extracted.The network restored the original size of the feature map through the transformer decoder,and at the same time integrated the segmentation results at different scales,and retained as much detailed information as possible,so that the network had better segmentation accuracy.The method had been tested on a dataset of real shots of airport runways,and the pixel-level segmentation index F 1-score for the cracks reached 78.91%,which was higher than the current various crack segmentation algorithms.
作者
李海丰
范天啸
黄睿
侯谨毅
桂仲成
LI Haifeng;FAN Tianxiao;HUANG Rui;HOU Jinyi;GUI Zhongcheng(College of Computer Science and Technology,Civil Aviation University of China,Tianjin 300300,China;Key Laboratory of Smart Airport Theory and System,Civil Aviation University of China,Tianjin 300300,China;Chengdu Guimu Robot Co.Ltd.,Chengdu 610100,China)
出处
《郑州大学学报（理学版）》
CAS
北大核心
2023年第4期30-38,共9页
Journal of Zhengzhou University:Natural Science Edition
基金
国家重点研发计划课题(2019YFB1310400)
天津市教委科研计划项目(2021KJ036)。
