Hybrid Attention Cascaded U-net For Building Extraction from Aerial Images

Abstract

Building extraction from aerial images is an effective way to monitor the urban changes and make city planning. To detect more easily missed small objects and improve accuracy, this paper proposes a simple but powerful deep neural network named Hybrid Attention Cascaded U-net (HAC U-net). In the new model, attention unit (AU) is designed to take the place of some skip connections between encoder and decoder. Receptive fields of different sizes are captured by AU. Through this replacement, more spatial and contextual information on different scales can be obtained. HAC U-net combines two encoder-decoder components to form a cascaded framework. Such cascaded architecture can deepen the depth of network and abstract advanced features. HAC U-net is the first model to flexibly decompose large-kernel convolutions, use residual connection internally and AU at different stages. Therefore, memory consumption is greatly reduced without significantly influencing the accuracy. Experimental results show that HAC U-net outperforms other baseline models with an accuracy of 93.90% and intersection over union (IoU) of 61.92%.