Multi-Modal Dictionary Learning for Image Separation With Application in Art Investigation

Area of Science:

• Art investigation
• Image processing
• Computational imaging

Background:

• X-ray imaging is crucial for art investigation, but separating signals from double-sided paintings is challenging due to similar signal characteristics.
• Existing source separation methods struggle with the complex morphological features present in such X-ray scans.

Purpose of the Study:

• To develop a novel source separation method for unmixing X-ray scans from double-sided paintings.
• To leverage multi-modal data (photographs and X-rays) for improved signal separation.

Main Methods:

• A coupled dictionary learning framework was developed to integrate information from front-and back-side photographs with X-ray data.
• The framework captures common and modality-specific features using parsimonious representations.
• Convex optimization procedures were formulated for accurate X-ray separation, with options for single- and multi-scale analysis.

Main Results:

• The proposed method accurately separates X-ray signals from complex, overlapping sources.
• Multi-scale dictionary learning significantly improved separation performance.
• Training dictionaries to ignore craquelure enhanced the visual quality of separated images.

Conclusions:

• The novel coupled dictionary learning approach outperforms state-of-the-art methods for X-ray separation in art investigation.
• This method provides a powerful tool for analyzing layered artworks and understanding their hidden details.