Attention Residual U-Net for Building Segmentation in Aerial Images

Abstract

Semantic segmentation of aerial images plays an important role in urban area monitoring. But the diversity of buildings makes segmentation a hard task. To detect buildings from aerial images more precisely, this paper proposes a pixel-level segmentation method, named Attention Residual U-net (ARU-net). ARU-net adds two major part into the framework of U-net, i.e. attention path and residual connection, focusing on feature reuse. Attention path utilizes attention mechanism to capture spatial feature details. Residual connection implies the semantic information flow through a 1×1 convolution similar to the residual form. ARU-net can be trained end-to-end. Experiments are conducted to evaluate the effectiveness of the proposed model on the Inria Aerial Image Labeling Dataset. Results indicate that ARU-net outperforms other baselines with an accuracy of 93.84% and intersection over union (IoU) of 60.90%.