Hyperspectral recovery from RGB images using Gaussian Processes

Hyperspectral cameras preserve the fine spectral details of scenes that are generally lost in the traditional RGB cameras due to the gross quantization of radiance. These details are desirable in numerous imaging applications, nevertheless the high cost of hyperspectral hardware and the associated physical constraints currently limit the pervasive use of hyperspectral imaging. We take a computational approach to construct hyperspectral images using the RGB cameras of known spectral response, and assuming a prior over the imaged scene. Our approach first clusters training hyperspectral image patches and infers a set of Gaussian Processes (GPs) to represent the naturally smooth reflectance spectra of materials in each cluster. The GPs and the clusters are then transformed to match the spectral quantization of the RGB camera. A patch from the test RGB image is assigned a matching cluster and it is represented by the transformed GPs for that cluster. The computed representation codes are combined with the original GPs to construct the desired hyperspectral signatures. The approach infers the Gaussian Processes using a model inspired by the Beta-Bernoulli Process and it encourages those to be non-negative to better approximate the positive reflectance values. We present the analytical expressions for the Bayesian inference over the proposed model. Thorough evaluation using three hyperspectral datasets demonstrates the effectiveness of the proposed technique.