In this work, a novel approach to color differencing in a spectral space is presented. The algorithm is based on well known pattern recognition technique-kernel methods, which include polynomial, Gaussian radial basis function (RBF) and sigmoid kernels. This article starts with a brief overview of several existing methods created both for color and spectral data. The performance of novel measures is tested against the Munsell Matte spectral dataset, and a spectral database of metameres. The results of tests obtained for kernel methods are compared with those produced by twelve conventional similarity metrics, i.e., Correlation Coefficient, Exponential Similarity, Maximum–Minimum methods, etc. The assumption behind experiments is that methods should model the behavior of human observers in the task of color differencing. The tests show that the proposed Gaussian RBF kernel metric performs significantly better, compared to the rest of the measures.