FPZNet: Fuzzy Perception Zero-Shot Learning for Infrared and Visible Image Fusion

Human cognitive systems excel at making approximate decisions in complex and uncertain environments, a capability particularly evident in visual perception. This inherent fuzzy decision-making ability has profound implications for multimodal image fusion, where the scarcity of ground-truth data for infrared and visible light integration presents a fundamental challenge for traditional deep learning approaches. Here we introduce a zero-shot transformer architecture that mirrors human cognitive flexibility by implementing fuzzy decision-making mechanisms for pixel-level fusion weight determination. Our approach circumvents the limitations of conventional pretraining requirements through a sparse attention mechanism that selectively preserves only 15-20% of the most salient cross-modal interactions, effectively filtering out redundancy and noise. To address the computational challenges of high-resolution data, we incorporate low-rank approximation techniques that reduce the complexity from quadratic to linear, capturing over 95% of cross-modal information using a projection dimension of merely 128. The method demonstrates remarkable stability across multiple benchmark datasets, achieving a peak signal-to-noise ratio of 65.904 dB under diverse environmental conditions. Our findings challenge the prevailing assumption that dense attention patterns are essential for effective feature integration, revealing instead that multimodal fusion inherently operates in a low-dimensional space. This biomimetic approach to zero-shot learning not only advances our understanding of cross-modal feature interactions but also provides a more generalizable framework for real-world applications where ground-truth data is scarce.