Continuous Spectral Reconstruction from RGB Images via Implicit Neural Representation

Existing spectral reconstruction methods learn discrete mappings from spectrally downsampled measurements (e.g., RGB images) to a specific number of spectral bands. However, they generally neglect the continuous nature of the spectral signature and only reconstruct specific spectral bands due to the intrinsic limitation of discrete mappings. In this paper, we propose a novel continuous spectral reconstruction network with implicit neural representation, which enables spectral reconstruction of arbitrary band numbers for the first time. Specifically, our method takes an RGB image and a set of wavelengths as inputs to reconstruct the spectral image with arbitrary bands, where the RGB image provides the context of the scene and the wavelengths provide the target spectral coordinates. To exploit the spectral-spatial correlation in implicit neural representation, we devise a spectral profile interpolation module and a neural attention mapping module, which exploit and aggregate the spatial-spectral correlation of the spectral image in multiple dimensions. Extensive experiments demonstrate that our method not only outperforms existing discrete spectral reconstruction methods but also enables spectral reconstruction of arbitrary and even extreme band numbers beyond the training samples.