Gated Fusion and Dual Attention Network for Road Extraction from Satellite Imagery

Automatic recognition and extraction of roads from high-resolution satellite images is a crucial task in remote sensing and computer vision. With the continuous development of remote sensing technology, more ground object information is contained in images, making it more challenging to extract roads due to increased interference. This paper proposes a gated fusion and dual attention network with an encoder-decoder structure to address this problem. The full-stage gated fusion module in skip connection selectively fuses feature maps of different scales using the Gated Fusion (GFF) unit, which increases the receptive field of the network and improves the accuracy of road extraction. The context extraction and dual attention module introduce rich global information and simultaneously weights features from both spatial and channel dimensions. This improves the semantic segmentation problem caused by focusing only on local information in current road extraction models. Experimental results on two public datasets show that GFDANet can effectively extract roads in complex scenes.