Spectral prediction method based on the transformer neural network for high-fidelity color reproduction

Area of Science:

• Color Science
• Computer Vision
• Machine Learning

Background:

• Color distortion is a common issue in digital transmission and reproduction.
• Existing spectral prediction methods lack accuracy in halftone reproduction.

Purpose of the Study:

• To develop an accurate spectral prediction model for CMYK halftone reproduction.
• To improve color reproduction quality and spectral accuracy.

Main Methods:

• Established a halftone dataset of four-color (CMYK) ink mixtures.
• Utilized a transformer network to model spectral features of mixed inks.
• Proposed forward color formulation and reverse spectral prediction models (SC-Former).

Main Results:

• The SC-Former model demonstrated superior performance over traditional and data-driven models.
• Achieved higher accuracy in spectral prediction for halftone reproduction.
• Showcased improved color reproduction effects in experiments.

Conclusions:

• The SC-Former model offers a significant advancement in high-fidelity color reproduction.
• This research provides a robust solution for spectral prediction challenges in printing.