Building extraction using remote sensing images plays an important role in urban planning, land use investigation and other fields. However, the buildings in the image are of various types and sizes, which brings great challenges to automatic extraction. In order to solve the problem of voids in large-scale buildings and missing detection in small-scale buildings in remote sensing image extraction, this paper designs a method that combines multi-scale features with non-local computation. The method adopts encoder-decoder structure. Firstly, Res2Net50 is used as the encoder to improve the multi-scale feature extraction capability, and then a non-local computing module is introduced in the decoder part to obtain context information to further improve the extraction results of buildings with different scales. The results indicate that IoU and F₁ values of the proposed method on the WHU building dataset reache 89.65% and 94.55%, respectively, , which is 1.52% and 0.86% higher than that of the original UNet and proves the effectiveness of the proposed method.