Cooperation of multi-task segmentation and a graph convolutional network for object vector boundary extraction in remote-sensing imagery

Abstract

ABSTRACT Identifying and vectorizing the object in the image is an important part of producing high-precision vector maps. Deep learning can automatically extract vector boundaries accurately, but it still does not satisfy the application requirements for boundaries. Clearer boundaries and more concise vector points are also important components that cannot be neglected in vectorization. Taking buildings as the research object, we introduce a cooperative neural network of multi-task segmentation and graph convolution to improve the extraction of buildings by strengthening the boundaries and strategically selecting key points. We design a multi-task neural network to extract and optimize the vector boundaries, whose key points can be selected and refined with a graph convolutional network. In addition, to improve the coherence between features and jointly multi-information, we design a mutual-supervision loss for our method. Our experimental results show that our method effectively extracted buildings and outperformed several equal methods on the different public datasets.