Optimizing the eigenvectors of spectral data with an initial clustering approach: Spectral versus colorimetric criteria

Abstract

AbstractClustering of spectral data based on the correlation distance coefficient of the reflectance spectra and the color difference values is presented to minimize the information loss caused by the compression of spectral data when the first three most important eigenvectors are considered. The suggested criteria used a distance function to partition the reflectance spectra of Munsell chips into subgroups using a k‐means algorithm. The cumulative energy contributions of the first three eigenvectors of the groups are calculated and compared to those that are computed based on the 10 principal hues of Munsell color system as well as nonclustered spectra. The performance of the suggested methods is evaluated by the colorimetric and spectrophotometric errors between the ground truth and the reconstructed spectra. While the clustering methods yield significantly better results than nonclustered data, the clustering approach based on the correlation distance coefficient of spectral data provides the best overall performance for all considered measures compared to colorimetric clustering strategies.