计算机科学
遮罩(插图)
遥感
人工智能
过程(计算)
特征(语言学)
计算机视觉
匹配(统计)
模式识别(心理学)
图像检索
迭代重建
语义特征
面子(社会学概念)
特征提取
歧管(流体力学)
透视图(图形)
语义学(计算机科学)
图像(数学)
标识
DOI:10.1080/01431161.2026.2740244
摘要
Remote Sensing Image-Text Retrieval (RSITR) still faces challenges in semantic recovery and fine-grained discrimination when visual or textual information is incomplete. Most existing methods give limited attention to sample-level differences in reconstruction reliability. They usually depend on single-path reconstruction and fixed masking ratios, which makes them less suited to scenes with different levels of semantic complexity. To reduce this problem, this paper proposes a Synergistic Reconstruction and Progressive Alignment (SRPA) framework. The Content-aware Adaptive Masking (CAM) module uses a lightweight MLP to estimate latent semantic density and predict a sample-specific masking ratio. This design keeps a better balance between reconstruction difficulty and feature preservation across heterogeneous datasets. The Quality-Guided Bimodal Reconstruction (QGBR) module then builds a self-correction mechanism by combining unimodal self-reconstruction with cross-modal guided reconstruction. This process encourages the reconstructed features to stay close to the original representations. The Hierarchical Progressive Alignment (HPA) module uses cross-modal momentum queues to enlarge the negative sample pool. It also follows a staged alignment process, moving from global manifold stabilization to local decision-boundary refinement. This helps the model distinguish visually similar remote sensing samples. Experiments on the RSICD, RSITMD, and UCM datasets show that SRPA achieves competitive retrieval performance and captures complex spatial-semantic relations in different remote sensing scenes.
科研通智能强力驱动
Strongly Powered by AbleSci AI