Advancing RGB-IR Image Fusion: Exploiting Hyperbolic Geometry

Abstract

Infrared and visible image fusion is essential for remote sensing applications, especially for obtaining high-quality imagery of terrestrial environments. Hierarchical feature information is crucial for image fusion as it captures the intricate relationships between different modalities, which are vital for producing detailed and accurate composite images. However, most existing methods operate within the confines of Euclidean space, which, due to its inherently “flat” geometric nature, often struggles to effectively measure the similarities and differences between modalities, thus failing to maintain their distinctiveness. Hyperbolic space, with its constant negative curvature, excels at leveraging these hierarchical structures. It can more effectively gauge the similarities and differences between modalities, preserving their distinctiveness. In this study, we propose a novel fusion method for infrared and visible image fusion in hyperbolic space, named HbFNet. We have developed innovative hyperbolic feature extraction modules, including Hyperbolic Invertible Neural Networks and Hyperbolic Lite Transformer blocks, specifically designed to capitalize on the hierarchical nature of features. Our method emerges as a promising solution for enhancing hierarchical information and elevating the quality of fusion. Extensive experiments across three public datasets have demonstrated that our method outperforms most state-of-the-art image fusion techniques.