Asymmetric Cross-Modal Hashing With Missing-Modality Feature Compensation for Remote Sensing Image Retrieval

Abstract

Cross-modal hashing plays a crucial role in large-scale remote sensing (RS) image retrieval. Existing methods often assume that both query and database images are single-modal, each from a different modality. In practice, however, database images are accumulated over a long period and are typically available in multiple modalities, while query images are captured in a short time and often contain only a single modality. This asymmetry creates significant challenges for traditional cross-modal hashing approaches, which cannot effectively handle missing or uneven modality information. To address this problem, we propose an asymmetric cross-modal hashing (ACMHash) with a missing-modality feature compensation framework. ACMHash consists of a cross-modal hashing network with a missing-modality feature compensation module and a semantic-centric Hamming alignment objective function. The hashing network leverages a prototype bank to bridge the semantic gap between query and database representations, while the objective function preserves semantic similarity in Hamming space and ensures robust cross-modal hash consistency. Extensive experiments on benchmark RS image retrieval datasets demonstrate that ACMHash outperforms state-of-the-art methods, achieving superior retrieval accuracy and robustness under asymmetric cross-modal scenarios.