The forward problem of spectral reflection prediction: Mutual match between framework selection and the training set volume

Abstract

Novel technology of the color reproduction is closely related to a variety of color prediction techniques. Along with the deterministic models, those based on the use of artificial neural networks have recently begun to appear. An important problem in the application of a neural network approach is the choice of the training set and network training algorithms. In our work, we describe the results of a computational experiment where various configurations of the artificial neural networks along the different volumes of training subsamples were simultaneously determined providing a satisfactory accuracy of the spectral reflection prediction. The results show that each subsample might be mapped to a particular network configuration.