Color error tendencies in spectral super-resolution

Abstract

We explored a machine learning approach towards spectral super-resolution (SSR) which aims to reconstruct high-dimensional spectral information from low-dimensional RGB information. Different models were designed and trained on data augmented ensemble and were made to predict novel datasets. It was observed that despite having low root-mean-square-error, rendered colors from recovered spectra were perceptually different from their original color. We therefore computed perceptual color error metrics to supplement performance measure. Error tendencies based on input pixel chromaticity revealed the non-correlation of spectral and color errors hence, reinforcing the need for colorimetric evaluation in future SSR research.