Neural Networks for Transformation to Spectral Spaces

This work is concerned with mapping between the CMYK colour space and spectral space using Artificial Neural Networks (ANNs). The dimensionality of the spectral space is high (typically 31) leading to a large number of weights (or free parameters) in the network. This paper explores the hypothesis that a computational advantage can be obtained, in these cases, by treating the reflectance at each wavelength as being independent of the reflectance at any other wavelength; the implication of this hypothesis is that instead of using a single large ANN, it is possible to use, for example, 31 separate networks, each of which maps to one dimension of the 31-d spectral space. The results showed that as the number of training samples is reduced the advantage of the population of single-wavelength networks over the standard neural network approach increased.