Deep Instance Search Network for Remote Sensing Image Retrieval

Remote sensing image retrieval (RSIR) has always been a hot research topic in the field of remote sensing, and instance-level remote sensing image retrieval (IL-RSIR) is one of the most important challenges. In recent years, although the powerful feature description ability of convolutional neural networks (CNNs) has improved RSIR significantly, the performance is still restricted by the complexity of remote sensing images, such as the same semantic labels but different appearance characteristics. To address these problems, we propose a novel deep instance search network (DISN). It leverages two-level retrieval branches, that is, semantic feature aggregator and keypoint matcher, to integrates the semantic information and the local details of instances and enhances the instance discriminability of feature representations. Experiments on four remote sensing benchmark datasets for IL-RSIR demonstrate that our DISN can outperform some state-of-the-art methods.