Hyperspectral Image Super-Resolution via Boundary Perception and Topology Inference

Hyperspectral image super-resolution involves fusing low-resolution hyperspectral images with high-resolution multispectral images, providing an effective way to improve the spatial quality of hyperspectral images. Most existing methods are devoted to fully integrating the modalities, achieving advanced performance. However, they primarily emphasize the overall intensity characteristics and neglect precise edge representation, which is inadequate for recovering credible textures. In this paper, we propose a network that aggregates image-level boundary perception and instance-level topological inference for hyperspectral image super-resolution, enhancing detail representation by improving the understanding of boundary knowledge and its contextual instances. Specifically, to estimate and incorporate effective edge information from the high-resolution auxiliary modality, we develop edge perception and enhancement units that intensify the focus on complementary edge details and refine the edge information representation. Furthermore, considering that the inter-instance topological properties are useful cues for edge detail definition, we introduce a topological inference unit to facilitate the interaction of instance-level boundary context, thereby augmenting the comprehension of edge contextual information and implicitly improving edge region identification. Experimental results on natural and remote sensing datasets demonstrate that the proposed method outperforms other state-of-the-art peers both visually and metrically.