Deep Orthogonal Fusion Smoothing Hashing for Remote Sensing Image Retrieval

Fuwei Huang, Yaxiong Chen, Yin Ye, Shengwu Xiong

Abstract: In the face of massive remote sensing image data, this is a challenging missions to retrieve images containing specific content quickly and accurately. With the characteristics of low storage and high efficiency, deep hashing algorithms have been widely used in image retrieval. However, with the crazy growth of the number of remote sensing images, effective retrieval and management of these massive data has become an urgent issue to be addressed. At present, many retrieval methods mainly extract the global features of images, which cannot distinguish images with different visual representations but similar semantic information effectively. To solve this problem, we propose a deep orthogonal fusion method based on global and local feature fusion called DOFSH. First, by designing an effective local feature extraction module, the remote sensing image retrieval system can better adapt to complex ground objects and background scenarios, thereby improving the accuracy and robustness of retrieval. Second, by designing a satisfactory local and global fusion scheme, these two features are integrated into a compact descriptor to achieve single-stage remote sensing image retrieval. Finally, to solve the problem that the uncontrollable distribution of remote sensing image dataset is incompatible with contrast loss, a loss function is designed to adapt to the characteristics of remote sensing image dataset. Comparative experiments demonstrate that our method outperforms existing approaches in terms of retrieval performance on the NWPU-RESISC45, AID, and AISION remote sensing image datasets.