Image Separation With Side Information: A Connected Auto-Encoders Based Approach.

Area of Science:

• Art investigation
• Digital image processing
• Machine learning

Background:

• X-radiography is crucial for art analysis, revealing hidden details and artist techniques.
• X-ray imaging of double-sided paintings produces mixed images, complicating analysis.
• Separating these mixed X-ray images is essential for detailed art historical research.

Purpose of the Study:

• To develop a novel method for separating mixed X-ray images of double-sided paintings.
• To utilize visible color images (RGB) to reconstruct individual X-ray images.
• To create a self-supervised learning algorithm that does not require pre-separated image pairs.

Main Methods:

• Proposed a new Neural Network architecture based on 'connected' auto-encoders.
• Encoders utilize convolutional learned iterative shrinkage thresholding algorithms (CLISTA) via algorithm unrolling.
• Decoders use linear convolutional layers to reproduce RGB and mixed X-ray images.
• Employed a self-supervised learning approach for training.

Main Results:

• Successfully separated mixed X-ray images into two distinct simulated X-ray images.
• The proposed connected auto-encoder architecture demonstrated superior performance.
• Outperformed existing state-of-the-art X-ray image separation methods in art investigation.
• Validated on double-sided wing panels of the Ghent Altarpiece.

Conclusions:

• The developed neural network effectively separates mixed X-ray images of double-sided paintings.
• This method offers a significant advancement for art investigation and analysis.
• The self-supervised approach simplifies the data requirements for training.