A Gaussian Process Model for Color Camera Characterization: Assessment in Outdoor Levantine Rock Art Scenes.

Area of Science:

• Digital imaging and color science.
• Computational modeling and machine learning.

Background:

• Accurate colorimetric camera characterization is vital for digital preservation of cultural heritage.
• Traditional methods, like polynomial regression, are widely used but may have limitations in capturing complex color transformations.

Purpose of the Study:

• To introduce and validate a novel Gaussian process (GP) based approach for colorimetric camera characterization.
• To compare the performance of GP models against traditional second-order polynomial models.

Main Methods:

• Application of Gaussian processes (GPs) for multivariate nonlinear function modeling.
• Colorimetric characterization using raw image data from two Single Lens Reflex (SLR) cameras.
• Leave-one-out cross-validation (LOOCV) to evaluate predictive performance using CIE XYZ residuals and Δ E a b * color differences.

Main Results:

• Gaussian process models achieved significantly lower residuals and Δ E a b * color differences compared to polynomial models.
• Δ E a b * values below 3 CIELAB units were attained, indicating high accuracy.
• Both models proved suitable for practical applications in cultural heritage documentation, with GP showing superior performance.

Conclusions:

• Gaussian process-based colorimetric characterization offers superior accuracy and is recommended for cultural heritage documentation.
• The study highlights the importance of camera selection due to noise variations affecting characterization outcomes.
• The findings support the use of advanced machine learning techniques for precise digital archiving.