Supervised Discriminative Transformer Hashing for Large-Scale Remote Sensing Image Retrieval
Jiajun Zhu, Xingbo Liu, Xuening Zhang, Xiushan Nie · IEEE Geoscience and Remote Sensing Letters · 2025
With the advancement of remote sensing technology and the exponential growth of remote sensing visual data, efficiently retrieving remote sensing images from extensive databases has become increasingly important. Deep hashing, which combines the advantages of deep learning and hashing techniques, has emerged as a significant research direction in remote sensing image retrieval (RSIR). However, remote sensing images often contain substantial amounts of complex background information that are unrelated to the target. This noise can obscure or interfere with the target features, making it challenging for the model to effectively distinguish the target objects. To address these challenges, we propose a novel method called Supervised Discriminative Transformer Hashing (SDTH) for large-scale RSIR task, designed to enhance the retrieval of remote sensing images by extracting more distinguishable features. Our approach utilizes the Swin Transformer V2 architecture to improve feature extraction capabilities, thereby acquiring richer global contextual information and multi-scale features. To mitigate the impact of image noise and generate more discriminative hash codes, we propose to integrate batch-hard triplet loss with symmetric cross entropy loss. Experimental results on three benchmark datasets for remote sensing demonstrate the effectiveness and superiority of the proposed method.