Multi-scale building instance refinement extraction from remote sensing images by fusing with decentralized adaptive attention mechanism

The accurate and efficient automatic extraction of building footprints from remote sensing images has a wide range of applications. Since the buildings in remote sensing images have different types, scales, shapes and backgrounds, the existing methods, to varying degrees, suffer from the problems of missing small-scale buildings, blurred contour boundaries, and inability to distinguish individual building instances. Therefore, this paper proposed a multi-scale building instance refinement extraction convolutional neural network (MBRef-CNN) fusing with decentralized adaptive attention mechanism for remote sensing images. First, a feature pyramid network fused with split-attention and adaptive attention mechanism (SA-FPN) was used to learn multi-scale building features. Then, according to the multi-scale features, the region proposal network (RPN) was used to detect the location of individual building instances. Finally, the boundary refinement network (BndRN) was used to iteratively acquire the precise building masks. On WHU aerial imagery dataset, the comparison experiments were conducted with the existing popular segmentation methods. The results show that the accuracy of the proposed method in this paper is higher than the others. Moreover, the MBRef-CNN shows good comprehensive performance in multi-scale building extraction, and has obvious accuracy advantages in small-scale building extraction. © 2023 SinoMaps Press. All rights reserved.