The Graph-structure based Graph Convolutional Network (GCN) integrates GCN  with spectral feature learning  and object classification . It is a popular deep learning framework for discovering feature representations in a graph-structured data. Convolutional neural network (CNN) is a class of well-known neural networks used for 2D image recognition, despite their reduction in resolution, the model's performance of CNN is still much better than SIFT or HOG.
In , the idea of enhanced convolutional neural network for image classification is adopted. The difference is that the layer in the normal convolutional neural network is replaced with an aggregation function to obtain rich feature representations for graph-structured data. With the help of aggregation function, GCN elevates the structural feature learning to the semantic domain.
An Attention Model (AM)  for graph-structure data is proposed. Graph-structured data includes not only the node data but also the edge data such as the edge weights. In , the authors analyze the graph-structure attention model to realize multi-instance multi-label classification. However, the methods above cannot generate label information, while the labels still play an important role in the concrete pavement distress detection problem. It is necessary to integrate multi-label classification with attention model to achieve the purpose of concrete pavement distress detection.
This paper will integrate a graph neural network for the feature extraction branch of the concrete pavement distress detection method. It breaks the concept of GCN into two components: an aware layer and an aggregation layer. The graph neural network component requires the input of a directed graph labeled data in a feature vector. And the aware layer is the transformation of the feature vector to output a feature vector representing the nodes of the graph. The aggregation layer combines this output so as to get a more accurate representation of the nodes. Finally, the aggregation layer will convert the whole graph feature vector back into a graph feature vector. As a result, the feature extracted by the aware layer can be used in the graph and the graph aggregation. We built a graph neural network that accepts a sequential data, such as an image, and outputs the same, in the form of a graph. d2c66b5586