Mixed X-Ray Image Separation for Artworks With Concealed Designs.

Area of Science:

• Art Conservation Science
• Computer Vision
• Digital Image Processing

Background:

• X-ray imaging reveals hidden layers in paintings, such as underdrawings or earlier compositions.
• Analyzing these complex X-radiographs is challenging due to the superposition of multiple features.
• Existing methods often require extensive labeled data for image separation.

Purpose of the Study:

• To develop a self-supervised deep learning model for separating mixed X-ray images of paintings.
• To differentiate between surface and concealed features within a single X-radiograph.
• To enable the analysis of hidden artistic content without prior knowledge of separated layers.

Main Methods:

• A novel self-supervised deep learning network comprising analysis and synthesis sub-networks.
• The analysis sub-network utilizes learned coupled iterative shrinkage thresholding algorithms (LCISTA) via algorithm unrolling.
• The synthesis sub-network employs linear mappings for image reconstruction.
• The model learns without requiring pre-separated X-ray image datasets.

Main Results:

• Successful separation of X-ray images into distinct layers representing surface and concealed artwork.
• Demonstration of the method's efficacy on a real-world case study: Goya's 'Doña Isabel de Porcel'.
• Validation of the self-supervised approach in accurately isolating hidden painting details.

Conclusions:

• The proposed self-supervised deep learning method offers an effective solution for X-ray image separation in art analysis.
• This technique advances the non-invasive study of concealed artistic elements and revisions.
• The approach holds significant potential for art history research and conservation efforts.