Using Bilinear-Siamese architecture for remote sensing scene classification

The key challenge of remote sensing (RS) scene classification is that features generated by similar scenes are difficult to distinguish. To solve the problem, we present a Bilinear-Siamese architecture to learn to distinguish the subtle discriminative features between similar scenes. Specifically, a pair of images are sent to the feature extraction module. Then, the extracted paired features are sent to two branches: 1) A fully connected (FC) layer to generate the normal classification results. 2) A bilinear mix module and a FC layer to generate the bilinear mixed classification results. Finally, we introduce a discriminative fusion method to fuse the aforementioned classification results for final output. Noted that, the contrast loss of Bilinear-Siamese architecture improves the ability to distinguish similar scenes based on metric learning. In addition, we introduce the additional bilinear loss to improve the generalization and the robustness of our network. We conduct extensive experiments on benchmark RS datasets to demonstrate the effectiveness of our network and the experimental results show that the performance of the proposed method surpasses other existing methods.