Rotation-Invariant Deep Embedding for Remote Sensing Images

Endowing convolutional neural networks (CNNs) with the rotation-invariant capability is important for characterizing semantic contents of remote sensing (RS) images since they do not have typical orientations. Most existing deep methods learning CNN models are based on design proper or pooling layers, which aims at predicting correct category labels rotated RS equivalently. However, a few works focused embeddings in framework metric modeling fine-grained relationships among embedding space. To fill this gap, we first propose rule that should be closer to each other than those any (including belonging same class). Then, maximize joint probability leave-one-out image classification and rotational identification. With assumption independence, such optimization leads minimization novel loss function composed two terms: 1) class-discrimination term 2) term. Furthermore, introduce penalty parameter balances these terms further final Rotation-invariant Deep images, termed RiDe. Extensive experiments conducted benchmark datasets validate effectiveness proposed approach demonstrate its superior performance when compared state-of-the-art methods. The codes article will publicly available https://github.com/jiankang1991/TGRS_RiDe .