Self-training-based spectral image reconstruction for art paintings with multispectral imaging

Area of Science:

• Art conservation science
• Computational imaging
• Spectroscopy

Background:

• Accurate spectral reflectance recovery is crucial for art analysis and conservation.
• Existing methods often suffer from material inconsistencies between training samples and artworks.
• Multispectral imaging offers rich spectral information but requires precise data reconstruction.

Purpose of the Study:

• To develop a self-training method for spectral reflectance recovery of art paintings.
• To improve the accuracy of spectral estimation by minimizing material discrepancies.
• To validate the method's performance against established techniques.

Main Methods:

• Utilized k-means clustering to partition multispectral images and extract training samples directly from art paintings.
• Employed coordinate paper for precise localization of training samples.
• Acquired spectral reflectances indirectly using a spectroradiometer and circle Hough transform for measurement area detection.

Main Results:

• The self-training method demonstrated superior spectral reflectance recovery performance compared to using commercial targets.
• The method's performance was found to be comparable to using painted color targets.
• Validation through simulation and practical experiments confirmed the method's effectiveness.

Conclusions:

• The developed self-training spectral reflectance recovery method accurately reconstructs spectral images of art paintings.
• Directly extracting training samples from artworks mitigates spectral estimation errors due to material differences.
• This approach offers a robust and effective solution for art analysis using multispectral imaging.