Explainable spectral super-resolution based on a single RGB image : a thesis presented in partial fulfilment of the requirements for the degree of Doctor of Philosophy (PhD) in Electronics and Computer Engineering at Massey University, Manawatu, New Zealand

Abstract

Hyperspectral imaging offers fine spectral measurements of target surfaces, finding utility in various fields. However, traditional hyperspectral systems grapple with high-cost issues. On the other hand, conventional RGB cameras, which provide relatively coarse measurements of surface spectra, are widely accessible. Consequently, the recovery of spectral information from RGB images has emerged as a popular approach for low-cost hyperspectral imaging, a venture also known as single-image spectral super-resolution. Yet, existing methods, mostly rooted in deep convolutional neural networks, tend to suffer from limited interpretability. In our research, we propose an explainable method for single-image spectral super-resolution. This method relies on the RGBPQR colour space, a low-dimensional spectral data model representing the spectrum. Leveraging the RGBPQR spectral model, we can transform the spectral reconstruction task into a regression problem. To tackle the metamerism issue, we analysed existing spectral super-resolution networks and discovered that these networks often depend on local textural information as context to mitigate metamerism. Informed by this insight, we utilized features extracted from multiscale local binary patterns as contextual information to design our explainable method. Furthermore, in this study, we discussed the error measurements and loss functions employed in this research area and proposed a new error measurement that can represent performance more accurately. We also endeavoured to put forward a method for quantitatively measuring the ability to resolve metamerism, a critical problem in spectral super-resolution. Through our research, we offered a simple, low-dimensional, and explainable spectral super-resolution solution.