Abstract: Building is the main feature of urban construction, is one of the basic elements of a city, is an important embodiment of the continuous development of urbanization, is the main place of human production and life. Therefore, how to effectively manage and supervise buildings is crucial. With the improvement of remote sensing image acquisition capability and widely spread applications, how to quickly and accurately extract buildings to provide a basis for subsequent applications become an urgent problem to be solved. In this paper, a method of building extraction by using remote sensing images based on foreground-aware is proposed by analyzing and combining advanced technologies such as depth learning. Firstly, the basic features are extracted by employing the improved ResNet from the input remote sensing image. Then, the pyramid feature map is obtained by using two-way FPN; the construction of relevant context association is achieved with the using of the foreground and geospatial scene modelling. After that, the input feature map is enhanced and the gap between foreground features and background features is enlarged, thus the foreground feature differentiation is improved. Finally, the efficient and accurate automatic extraction of buildings from remote sensing image is realized.

Key words: foreground perception, remote sensing image, building extraction, convolutional neural network, ResNet network structure