Encoder-decoder structure based on conditional random field for building extraction in remote sensing images

The application of building extraction involves a wide range of fields, including urban planning, land use analysis and change detection. It is difficult to determine whether each pixel is a building or not because of the large difference within the building category. Therefore, automatic building extraction from aerial images is still a challenging research topic. Although deep convolutional networks have many advantages, the networks used for image-level classification cannot be directly used for pixel-level building extraction tasks. This is caused by successive steps larger than one in the pooling or convolution layer. These operations will reduce the spatial resolution of feature maps. Therefore, the spatial resolution of the output feature map is no longer consistent with that of the input, which cannot meet the task requirements of pixel-level building extraction. In this paper, we propose a encoder-decoder structure based on conditional random field for building extraction in remote sensing images. The problem of boundary information lost by unitary potential energy in traditional conditional random field is solved through multi-scale building information. It also preserves the local structure information. The network consists of two parts: encoder sub-network and decoder sub-network. The encoder sub-network compresses the spatial resolution of the input image to complete the feature extraction. The decoder sub-network improves the spatial resolution from features and completes building extraction. Experimental results show that the proposed framework is superior to other comparison methods in terms of the accuracy on open data sets, and can extract building information in complex scenes well.