基于可变形卷积与特征融合的机场道面裂缝检测算法

Airport Pavement Crack Detection Algorithm Based on Deformable Convolution and Feature Fusion

机场道面裂缝具有形态多变、宽度狭小、长短不一、且空间走势呈自由曲线的不规则特征,现有算法检测效果不佳。针对此问题,本文构建了一种基于可变形卷积与特征融合的神经网络(Deformable convolution and feature fusion neural network,DFNet)模型。首先由可变形卷积模块来强化特征提取网络对裂缝形态特征的学习;然后经多尺度卷积模块捕获不同感受野下裂缝的全局信息;最后通过特征融合模块来提取裂缝不同层次的特征,通过融合裂缝低级特征与高级特征,实现对机场道面裂缝的准确分割。在采集的实际机场道面裂缝数据集上,与其他6种现有算法进行了对比实验,本文算法在像素级分割的F1-Score上达到了90.95%,效果优于全部对比算法。DFNet算法提高了对机场道面裂缝检测的能力,实验结果表明本文算法较好地达到了工程实际要求。

Due to the irregular features of the airport pavement cracks with variable shapes,narrow widths,different lengths,and free-form spatial trends,the existing algorithms perform not well. To solve this problem,this paper constructs a neural network model based on deformable convolution and feature fusion(DFNet). Firstly, the deformable convolution module is used to enhance the learning of fracture morphological characteristics by feature extraction network. Secondly,the multi-scale convolution module captures the global information of the fracture under different receptive fields. Finally,the feature fusion module is used to extract the characteristics of different levels of the fracture,through combining low-level and high-level features of cracks to achieve accurate segmentation of cracks on airport pavements. On the collected actual airport pavement crack data set,comparative experiments are carried out with other six existing algorithms. The proposed algorithm reaches 90.95% on the F1-Score of pixel-level segmentation,which is better than those of all other compared algorithms. The DFNet algorithm improves the ability of detecting cracks in airport pavements. Experimental results show that the proposed algorithm better meets the actual engineering requirements.