.png){kind=icon}
.png){kind=icon}
Return
Continuous Reliability-Calibrated Alignment for Remote Sensing Image-Text Retrieval
.png){kind=icon}
PRE
AI
B
P
H
DOI：10.1080/01431161.2026.2691984.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Remote Sensing Image-Text Retrieval (RSITR) is the task of learning a shared representation to measure the semantic similarity between remote sensing (RS) images and their textual descriptions. This technology is critical for applications such as disaster assessment and urban management. However, this task is highly challenging due to the varying degrees of alignment reliability between information-dense RS images and their sparse textual descriptions, ranging from explicit mismatches to weak correspondences, thereby fundamentally limiting model retrieval accuracy and robustness. Mainstream embedding-based methods typically rely on discrete supervision paradigms that fail to handle this continuous spectrum of reliability alignment. To resolve this, we propose the continuous reliability-calibrated alignment (CRCA) framework, which pioneers a reliability-aware learning paradigm. The core of our approach is a new supervision framework featuring two key innovations: reliability-calibration weighting (RCW) module, which assigns continuous weights to each pair, with a confidence-gated triplet fusion (CGTF) loss for stable and discriminative learning. To provide a robust foundation for RCW’s assessment, we first employ attentive token condensation (ATC) to purify features by filtering background noise. Furthermore, to deepen the model’s fine-grained semantic understanding, we introduce a Text-guided visual reconstruction (TVR) auxiliary task that compels the model to learn robust local region–word correspondences. Extensive experiments on the RSICD and RSITMD benchmarks demonstrate that CRCA achieves highly competitive performance, with remarkable mR scores of 38.62% and 50.80%, respectively.
Keywords:
Remote Sensing Image-text Retrieval (RSITR)
reliability-aware learning
adaptive sample weighting
fine-grained representation
masked visual-language modelling
Journal
IF:
2.6
Papers:
1.2W
Citations:
2.7W
Organization
Cited Papers