Text-Driven Fusion for Infrared and Visible Images: Achieving Image Scene Adaptation on Hyperbolic Space
Researchers have developed a novel framework for fusing infrared and visible images by leveraging hyperbolic manifold learning. This approach uses text prompts, extracted by BLIP, as anchors in hyperbolic space to align visual attributes. The method naturally encodes hierarchical semantics and avoids metric saturation, leading to improved fusion performance compared to existing Euclidean methods. Notably, the fusion process adapts autonomously to input content at inference time, removing the need for explicit textual input. AI
IMPACT This hyperbolic geometry approach could enable more nuanced image fusion for applications requiring the integration of complementary visual data.