CaFEN: A Causal Feature Embedding Network for Visible-Infrared Image Registration

Due to the significant cross-modal differences in texture, intensity, and structural representation, image registration in multi-view computer vision remains a critical yet challenging task, particularly for visible and infrared images. Traditional methods primarily follow the “detect-then-describe” paradigm, but their performance deteriorates under extreme conditions. To address this, we propose a novel cross-modal image registration framework called the causal feature embedding network (CaFEN), which enhances two key aspects: causal inference for robust feature extraction and cosine alignment loss for improved feature alignment. Specifically, the causal feature weighting module (CFM) mitigates confounding factors in the feature extraction process, ensuring more stable and semantically meaningful representations. Meanwhile, the cosine alignment loss enforces consistency between corresponding features across different modalities, improving cross-modal matching accuracy. Experimental results on benchmark datasets demonstrate that CaFEN achieves superior accuracy and robustness in visible-infrared image registration.