The Inverse Problem of Spectral Reflection Prediction by Artificial Neural Networks: Neugebauers Primaries vs. Recipes

Abstract

Reliable color reproduction can only be achieved by establishing a precise mutual correspondence between the spectral reflectivity of the print and the amount of ink used. Therefore, digital image processing for color reproduction often requires significant calculations used to characterize printing devices. Such calculations involve the implementation of different mathematical models. Most reflection spectrum prediction models use certain mathematical methods to predict the reflection coefficient for a blend of colorants that are characterized by certain characteristics of the absorption and scattering of light. However, until now, few efforts have been made to create an inverse model for predicting dye values based on the observed spectrum. In this paper, we attempt to apply the machine learning approach to solve the inverse problem in the field of spectral reflection prognosis. We put forward the assumption that the prediction of the initial colorant formulation from the available spectral data is possible by analogy with the work of the color perception system in humans. In the work, we compare two approaches in order to define whether it is better to predict the recipes themselves or the Neugebauer primaries. The models were created in Matlab and have shown tolerable prediction accuracy.