Abstract: Because of the various size of buildings and complicated background in high-resolution remote sensing images,there are some problems such as loss of details and blurring of edges when extracting buildings in remote sensing images,which affect the segmentation accuracy of the model.In order to solve these problems,this paper proposes a two-branch architecture network B2Net with spatial and semantic information branches.Firstly,the cross feature fusion module is provided in the semantic information branch to fully capture the context information to aggregate more multi-scale semantic features.Secondly,in the spatial branch,we combine the atrous convolution and depthwise separable convolution to extract the multi-scale spatial features of the image,and optimize the dilated rate to expand the receptive field.Finally,we use the content aware attention module to adaptively select the high-frequency and low-frequency content in the image to achieve the effect of refining the edges of building segmentation.We train and test the B2Net on two building datasets.On the WHU dataset,compared with the baseline model,the B2Net achieves the best result in precision,recall,F1 score and IoU,which is 98.60%,99.40%,99.30%,and 88.50%,respectively.On the Massachusetts building dataset,the four indicators are 0.9%,1.9%,1.7% and 2.2% higher than BiSeNet,respectively.Experiments show that B2Net can better capture spatial detail and high-level semantic information,improve the segmentation accuracy of buildings in complicated backgrounds,and meet the needs of rapid extraction of buildings.

Key words: Building extraction, Feature fusion, Atrous convolution, Depthwise separable convolution, Content aware attention