Adaptive global training set selection for spectral estimation of printed inks using reflectance modeling

Area of Science:

• Color Science
• Computer Vision
• Machine Learning

Background:

• Spectral estimation performance relies heavily on training sample selection.
• Existing global training schemes focus on reducing sample size or selecting representative samples.
• Printed ink reflectance estimation is crucial for in-line print inspection quality assessment.

Purpose of the Study:

• To introduce a novel global training sample selection scheme for printed ink reflectance estimation.
• To model a large population of realistic printable ink reflectances for training.
• To improve the efficiency and accuracy of spectral estimation in print inspection.

Main Methods:

• Developed a global training scheme using a large dataset of realistic printable ink reflectances.
• Employed a recursive top-down algorithm to prune training samples that do not improve performance.
• Utilized linear least-square regression (pseudoinverse-based estimation) for spectral estimation.
• Conducted experiments using real camera response data from a 12-channel multispectral camera system.

Main Results:

• The proposed global training scheme outperforms state-of-the-art algorithms in estimation quality.
• The method efficiently handles large datasets, a significant advantage for practical applications.
• Reflectance modeling is demonstrated as a viable and convenient approach for generating extensive training sets.
• Experiments validated the scheme's effectiveness with real-world multispectral camera data.

Conclusions:

• The novel global training sample selection scheme significantly enhances spectral estimation for print inspection.
• This approach offers a computationally efficient and accurate method for handling large-scale training data.
• Reflectance modeling provides a practical solution for creating robust training datasets in the print industry.