基于深度学习的路面裂缝提取

Pavement Crack Extraction Based on Deep Learning

裂缝是主要的路面损坏,路面裂缝自动化提取对于维护和监测路面状况至关重要。针对传统人工检测效率低、缺乏客观性等问题,提出了一种基于深度学习语义分割网络的路面裂缝自动提取方法,实现了由路面图像自动生成裂缝二值图。并且阐释了不同尺度特征对语义分割的好处,并结合裂缝本身细小的特性,在U-Net的基础上增加了大尺度特征提取分支。随后采用激励压缩模块进行两种特征的融合。在CFD(computational fluid dynamics)数据集上的实验表明,该改进算法的F1分数、kappa系数分别可达74.28%和73.83%,相较于其他主流分割网络,提高了约2%。

Crack is the main pavement surface damage.Automatic crack extraction is essential for maintenancing and monitoring pavement surface condition.Aiming at the problem that the use of artificial methods is inefficient and lack of objectivity,this paper proposes a method of Automatic crack extraction based on deep learning semantic segmentation network,which can generate binary images from pavement images.What’s more,this paper explains the benefits of different scale features for semantic segmentation.With the consideration of crack’s characteristics,thin and small,this paper adds a large-scale feature extraction branch on the basis of U-Net.Then the different features are fused by Squeeze and Excitation module.Experiment on CFD dataset shows that the F1 score and kappa coefficient of the improved algorithm can reach 74.28%and 73.83%respectively,which are about 2%higher than state-of-the-art segmentation networks.