R2H-CCD: Hyperspectral Imagery Generation from RGB Images Based on Conditional Cascade Diffusion Probabilistic Models

Hyperspectral imaging can capture detailed spectra of materials, but due to imaging conditions and equipment limitations, the cost of hyperspectral data collection is extremely high. Leveraging low-priced and easy-obtainable RGB images to generate hyperspectral images (HSIs) has gradually become a trend. Benefiting from the advantages of the denoising diffusion probability model (DDPM) in natural image super-resolution tasks, we consider the structural characteristics of HSI to design a novel spectral super-resolution algorithm, named R2H-CCD, which is a hyperspectral imagery generation method from RGB images based on conditional cascade diffusion probabilistic models. More specifically, the algorithm takes RGB image as conditional input and synthesise corresponding HSI from pure noise through a stochastic iterative denoising process. In addition, to obtain high-fidelity HSIs, it adopts U-Net structure to iteratively refine the generated images in an end-to-end training manner. The experiments on HFD100 dataset show the effectiveness and superiority of the proposed method.