The least squares fitting is a commonly used method of output colour device characterization. For this purpose, polynomial models are fitted to the data obtained by measuring the device responses to inputs. As it implies the model form, i.e. the polynomial terms to be chosen prior to fitting the model, this choice affects the model prediction power. Terms of a given order in higher order models can be formed in a large number of ways by combining different powers over the variables in cross-product terms, all resulting in the same order. In this research, optimal models for the eight processes were determined by performing backward elimination on their maximum models. The backward elimination is the procedure of eliminating terms or blocks of terms from the maximum model, one at a time, and comparing each reduced model to the maximum model in terms of sum of squared differences. The evaluation was performed on the same independent test used to determine maximum models. Partial F test was used to compare performances of different models. The maximum and optimized models were, in addition to statistical evaluation, also evaluated psychophysically by transforming an image and evaluating it visually. The evaluation showed that the reduced model performed better than the maximum model. This paper is a part of a larger study which has the aim to investigate whether the significance of model terms can be related to some statistic calculated from the characterization data. This would allow choosing the appropriate terms and building models adapted to particular devices. |