On the Direct Problem of Spectral Reflection Prediction by Artificial Neural Networks: Framework Selection

Abstract

Digital image processing involves different methods for color and spectral prediction during printing systems characterization and profile making, which requires significant amount of computations. Continuing complication of models leads to high demands on computing power, however this does not always contribute to the convenience and accuracy of forecasts. We offer an easy-to-operate method of spectral prediction by artificial neural networks. We train networks by color recipes, however we use the spectral density instead of spectra as a target values. This provides simplification of the simulation and high accuracy measured in terms of color difference CIE Lab dE2000 that is confirmed during the experimental verification. In this paper we describe the framework selection with the minimal volume of training set and discuss the results of experiments.