Fine-Grained Information Guided Super-Resolution for Mural Hyperspectral Image

Hyperspectral image (HSI) supports non-invasive pigment analysis and digital restoration of ancient murals, yet its spatial resolution is fundamentally constrained by acquisition hardware and fragile heritage environments. Existing hyperspectral super-resolution (SR) approaches are largely developed for remote sensing imagery and assume homogeneous textures, linear spectral mixtures, and large-scale training data. These assumptions do not hold for mural imagery, where fine brushstroke details coexist with smooth plaster substrates and mineral pigments exhibit nonlinear spectral mixing. Consequently, directly applying conventional HSI-SR models often leads to over-smoothed textures or spectral distortion. We propose FIGNet, a mural-oriented hyperspectral SR framework explicitly designed to address these domain mismatches. First, FIGNet introduces an adaptive gated spectral–spatial frequency decomposition to explicitly separate structural background and high-frequency cultural details, enabling differential processing aligned with mural texture characteristics. Second, FIGNet performs joint spectral–spatial contextual refinement to preserve pigment signatures during sub-pixel reconstruction. Finally, a hybrid optimization objective incorporating spectral–spatial variation regularization and adaptive frequency-domain supervision is designed to stabilize reconstruction under limited training data and reduce spectral inconsistency. Extensive experiments on two mural HSI datasets under multiple upscaling factors demonstrate that FIGNet achieves superior spatial detail recovery while maintaining spectral fidelity, validating the importance of frequency-aware and spectrally coupled modeling for cultural heritage HSI-SR.