Abstract: In order to improve the detection accuracy of road pavement cracks, a dual encoder pavement crack detection network based on Sobel operator bridging was proposed to address the issues such as the polymorphism of road cracks and the complexity of noise. The dual encoder consisted of two parts: original image encoding and gradient encoding, in order to solve the problem of single encoder easily losing gradient information. Firstly, the original image encoding results were computed by bridging the Sobel operator in eight directions to generate the encoding information for gradient encoding. Then, the original image encoding results and the gradient encoding results were passed through a multi-scale edge information compensation module to enhance the edge information of the cracks. Finally, dynamic channel graph convolution was introduced to obtain the topological relationships existing between channels to highlight the semantic features of important channels. The research results indicate that the proposed method achieves better results on three benchmark datasets such as DeepCrack, CamCrack789 and CFD. The composite metric ODS is 87.75%, 85.05% and 78.83% for DeepCrack, CamCrack789 and CFD datasets, respectively.

Key words: highway engineering; pavement crack detection; dual encoder; Sobel operator; edge information compensation; dynamic channel graph convolution

摘要： 为提高道路路面裂缝的检测精度，针对路面裂缝的多态性和噪声复杂等问题，提出了一种基于Sobel算子桥接的双编码器路面裂缝检测网络，双编码器由原图编码和梯度编码两部分组成，以解决单编码器容易丢失梯度信息的问题。首先，原图编码结果通过桥接Sobel算子计算8个方向产生梯度编码的编码信息；然后，将原图编码结果与梯度编码结果通过一个多尺度的边缘信息弥补模块，以增强裂缝的边缘信息；最后，引入动态通道图卷积获得通道之间存在的拓扑关系，以突出重要通道的语义特征。研究结果表明：所提出的方法在DeepCrack、CamCrack789和CFD这3个基准数据集上取得较好的结果；综合指标ODS在DeepCrack、CamCrack789和CFD数据集分别为87.75%、85.05%、78.83%。

关键词: 道路工程；路面裂缝检测；双编码器；Sobel算子；边缘信息弥补；动态通道图卷积