REMOTE SENSING IMAGES FEATURE LEARNING BASED ON MULTI-BRANCH NETWORKS

Remote sensing (RS) images feature learning, plays a crucial role in many RS images application, and attracts scholars' attention. However, since RS images contain complex contents, how to extract robust features that can fully represent RS images becomes an important and tough task. In this paper, we develop a feature learning method based on multi-branch networks, named M-Net, which consists of fine-grained branch and coarse branch. Considering the objects within RS images are diverse in type and resolution, the fine-grained branch is developed to capture rich object-level information. First, the RS images convolutional features are extracted by fine-grained branch. Second, through encoding the score maps which can highlight the important regions, the fine-grained structure mapping are obtained. Finally, the object-level features are generated by transforming the convolutional features through mapping. The coarse branch is developed to transform the obtained object-level features into global structure for representing images The positive experimental results counted on RS benchmark data set demonstrate that the proposed M-Net can learn more powerful features.