Arbitrary-Shaped Building Boundary-Aware Detection with Pixel Aggregation Network

Abstract

Large-scale building extraction is an essential work in the field of a remote sensing image analysis. The high-resolution image extraction methods based on deep learning have achieved state-of-the-art performance. However, most of the previous work has focused on region accuracy rather than boundary quality. Aiming at the low-accuracy problems and incomplete boundary of the building extraction method, we propose a predictive optimization architecture, BAPANet. Notably, the architecture consists of an encoder-decoder network, and residual refinement modules responsible for prediction, and refinement. The objective function optimizes the network in the form of three levels (pixel, feature map, and patch) by fusing three loss functions: binary cross-entropy, intersection over-union, and structural similarity. The five public datasets' experimental results show that the extraction method in this article has high region accuracy, and the boundary of buildings is clear and complete.