A symmetrical parallel two-stream adaptive segmentation network for remote sensing images

Segmentation of remote sensing images plays an important role in various civil applications. Although some achievements of artificial intelligence have been made in the past, the current challenge of remote sensing image segmentation is mainly the inadequate capture of global and local features, which leads to poor target feature extraction. This paper proposes a parallel two-stream adaptive remote sensing image segmentation network with symmetric semantic reasoning and context awareness, which enhances the feature extraction ability and further improves the segmentation accuracy. The proposed network consists of a main stream and a subordinate flow. Specifically, main stream is mainly used to extract local features from remote sensing images. The subordinate flow obtains the global feature information of the image. In the two-stream network coding stage, a hierarchical aggregation module is proposed to achieve the purpose of mining the global and local features of remote sensing images. In addition, to further improve the discriminate power of multi-scale features, an adaptive semantic reasoning module is proposed to extract multi-scale features. Experiments are carried out on two commonly used data sets, and the experimental results prove the effectiveness of the proposed network.