AN END-TO-END ADVERSARIAL HASHING METHOD FOR UNSUPERVISED MULTISPECTRAL REMOTE SENSING IMAGE RETRIEVAL

Unlike natural images, remote sensing images are usually multispectral. And the lack of sufficient labeled data puts a limit on supervised learning for remote sensing image retrieval. In this paper, we proposed a novel method for unsupervised multispectral remote sensing image retrieval. The proposed method makes use of the unsupervised representation learning ability of GAN. Meanwhile, a new reconstruction loss exploits the latent codes in GAN to make the final output informative and representative. Transfer learning and color histogram method are used to generate an estimated similarity matrix to further guide the training. Hash constraints can make the output codes binary and compact. In the testing stage, the hash codes of multispectral images can be computed in an end-to-end manner. Experiments on a multispectral remote sensing image dataset, EuroSAT, show the superiority of the proposed method over other state-of-the-art methods.