Segmentation of Buildings in Aerial Photographs Using Graph Convolutional Networks

The paper considers the problem of segmentation of buildings in aerial photographs. Graph convolutional networks are used for segmentation of buildings. The spectral convolution of the graph is calculated. Graphs are constructed on the basis of image superpixels. The image is pre-segmented into superpixels using simple linear iterative clustering. The color of the superpixels is averaged over all points in the region. Superpixels are the vertices of the graph. Adjacent regions are connected by edges. A binary image mask is used during training. A graph is also constructed for the mask. The structures of graphs of the original image and the mask are the same. The color of a superpixel is predicted on the basis of the classification of nodes of the graph. The network consists of five convolutional layers. For training, 5000 images of buildings are used. Graph neural networks use aggregated information about nodes from the neighborhood of the image in a convolutional manner. The results of the proposed method are presented. The accuracy of the method is comparable to known solutions. The advantage of the presented method is that training occurs on a relatively small set of images.