Concrete Crack Detection Based on Hybrid Residual Network and Graph Convolutional Network

Crack is an important sign of degradation of health and reliability of civil infrastructure. It is of great significance to detect cracks automatically to maintain civil infrastructure. Many computer vision based concrete crack detection methods had been proposed, but currently, the proposed methods did not consider the relationship between categories when generates classification parameters, and ignored the global correlation between labels. To solve the problem, a concrete crack detection method based on hybrid residual network and graph convolutional network is proposed. Firstly, the crack features extraction network was constructed by using ResNet-101 to generate crack feature map. Then, the label matrix of crack feature maps and the adjacency matrix according to the co-occurrence relationship between labels of crack images were constructed and generated respectively. Finally, the concrete crack detection network was constructed by using the graph convolutional network to detect concrete cracks. In order to verify the detection result, a comparative experiment on BCD and SDNET2018 data sets was conducted. The experimental results show that this method has better accuracy compared to other methods, such as CNN, SSENet and Inception v3.