Anchor Map-Guided Dynamic Fusion for Multimodal Occluded Image Recognition Under Limited Sample Conditions
Chao Li, Jiacheng Ni, Ying Luo, Siyuan Zhao, Qun Zhang · IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 2026
With the rapid advancement of multimodal remote sensing fusion technology, significant progress has been made in the application of multimodal image target recognition in fields such as target reconnaissance and environmental detection. However, in real-world scenarios, the targets to be identified are often subject to occlusion or camouflage, leading to a notable decline in recognition performance. To address this challenge, we propose an anchor map-guided dynamic fusion method for multimodal occluded image fusion and recognition under limited sample conditions. First, we propose a learnable anchor map module that synthesizes a global feature template from unobstructed multimodal inputs, providing an ideal reference for subsequent occlusion assessment. Second, we design an anchor map-guided occlusion-aware module that performs hierarchical evaluation via global difference measurement and local boundary detection to accurately quantify the occlusion level of each modality. Third, we propose a cross-modal Mamba fusion module with adaptive gating. It leverages a state-space model to transform spatial features into sequential representations, and dynamically assigns fusion weights based on occlusion degree and feature reliability via an adaptive gating mechanism. Finally, comparative experimental results on multimodal image datasets demonstrate that our method achieves optimal performance in occluded target recognition tasks compared to existing approaches.