Scale-Semantic Joint Decoupling Network for Image-Text Retrieval in Remote Sensing

Image-text retrieval in remote sensing aims to provide flexible information for data analysis and application. In recent years, state-of-the-art methods are dedicated “scale decoupling” “semantic strategies further enhance the capability of representation. However, these previous approaches focus on either disentangling scale or semantics but ignore merging two ideas a union model, which extremely limits performance cross-modal models. To address issues, we propose novel Scale-Semantic Joint Decoupling Network (SSJDN) image-text retrieval. Specifically, design Bidirectional Scale (BSD) module, exploits Salience Extraction Map (SEM) Suppression (SSM) units adaptively extract potential features suppress cumbersome at other scales bidirectional pattern yield different clues. Besides, Label-supervised Semantic (LSD) module by leveraging category semantic labels as prior knowledge supervise images texts probing significant semantic-related information. Finally, Semantic-guided Triple Loss (STL), generates constant adjust loss function improve probability matching same image text shorten convergence time model. Our proposed SSJDN outperforms numerical experiments conducted four benchmark datasets.