AI and Machine Learning Methods for n-Colorant Printer Characterization

Printer color characterization maps colorant amounts to colorimetry. We compare 13 machine learning methods with polynomial regression on nine characterization datasets with three to seven colorants, and 13 newsprint press runs, using the CIEDE2000 color difference (ΔE00). The planned (pre-specified) comparison fits every model in XYZ. A cube-root (CIELAB-like) fitting space was added later and applied to every method. In the main comparison, Gaussian process regression has the lowest median on all nine datasets. On the three five- and seven-ink datasets, a degree-4 polynomial fitted in cube-root space comes within a factor of 1.1 to 1.7 of it. On the chart with the largest margin, where 84% of patches use at most four inks, a degree-5 cube-root polynomial has a lower median but twice the maximum error. A cube-root support vector regressor tuned inside each training fold has a lower median on two of the three, but a worse error tail on both. Machine learning characterizes n>4 colorant systems well, and so does polynomial regression, given the right fitting space and degree. The cube-root Gaussian process is the most reliable fixed configuration: close to the best in median, 95th percentile and maximum error, but not always the best.