A Structure-aware Spiking Neural Network with Dynamic Membrane Potential Threshold for Infrared-Visible Image Matching

Infrared-visible image matching is particularly challenging in complex scenes with significant modality discrepancies and highly similar textures or structures, where reliable correspondences are difficult to determine and mismatches are prone to occur. Since robust matching mainly relies on sparse and stable structural information rather than appearance consistency, we propose a structure-aware spiking neural network with a dynamic membrane potential threshold for infrared-visible image matching. Unlike conventional artificial neural networks, spiking neural networks generate sparse and event-driven responses, making them inherently more suitable for modeling stable cross-modal structures. Specifically, we construct a hierarchical spiking matching framework to learn multi-scale structure-aware representations from infrared and visible images. To further improve fine-grained cross-modal structural modeling, we address the limited flexibility of existing spiking neurons and propose a novel spiking neuron with a dynamic membrane potential threshold. This neuron adaptively modulates its firing sensitivity according to input responses, enabling more discriminative representation of cross-modal structural information. Extensive experiments demonstrate that DMPT-SNN achieves robust cross-modal correspondence learning, with a 5.98% relative improvement in AUC@10° on the METU-VisTIR Cloudy-Cloudy subset and 7-30% lower matching errors across LLVIP, M3FD, MSRS, and MSIV compared with existing methods.