Enhancing Embedding Diversity and Robustness for Image-Text Retrieval in Remote Sensing

Remote Sensing Image-Text Retrieval (RSITR) aims to retrieve semantically matched images or textual descriptions. RSITR faces the significant challenge of achieving accurate cross-modal alignment. Existing methods commonly assume that all image-text pairs are correctly aligned, where a single textual description corresponds to one specific image. However, in real-world RSITR scenarios, image-text pairs inevitably exist incorrectly-correlated relationships, where texts contain factual errors in describing image content. Moreover, these datasets also exhibit weakly-correlated relationships, where a single image is correlated with multiple semantically similar texts, leading to the difficulty of cross-modal alignment. In this paper, we propose the Enhancing Embedding Diversity and Robustness (EEDR), which enhances cross-modal alignment by learning both instance-level and feature-level diversity to improve visual-textual correspondence robustness. Firstly, to alleviate the impact of incorrectly-correlated relationships, we propose an Incorrectly-Correlated Feature Rectification (ICFR) module that incorporates large language model (LLM) guidance to refine instance-level visual-textual correspondences. This module introduces a dynamic margin-guided mechanism that quantifies feature discrepancies to adaptively weight the original and LLM-generated auxiliary descriptions, thereby suppressing the interference of incorrect textual descriptions. Secondly, a Weakly-Correlated Feature Decoupling (WCFD) module is proposed to learn diverse discriminative features across modalities. WCFD module learns parameterized normal distributions to generate modality-specific decoupled features, enabling the model to accurately establish visual-textual correspondences under the condition of multiple semantically similar texts. We conduct extensive experiments on benchmark datasets, demonstrating that our approach outperforms state-of-the-art methods. Our code is available at https://github.com/ycharlene/EEDR.