Generative Deep Learning-Based Thermographic Inspection of Artwork

Area of Science:

• Materials Science
• Computer Vision
• Art Conservation Science

Background:

• Infrared thermography is a key nondestructive testing method for artwork.
• Raw thermograms often have noise and limited data, hindering accurate defect detection.
• Advanced data enhancement techniques are needed to improve thermographic analysis.

Purpose of the Study:

• To develop and evaluate deep learning methods for enhancing thermographic data for artwork inspection.
• To improve defect detectability in artwork using enhanced thermograms and principal component analysis.

Main Methods:

• Proposed a defect inspection method combining Principal Component Analysis (PCT) with deep learning.
• Utilized Spectral Normalized Generative Adversarial Network (SNGAN) for data augmentation.
• Employed Convolutional Autoencoder (CAE) for thermographic image quality enhancement.

Main Results:

• SNGAN integration improved signal-to-noise ratio by 1.08%.
• CAE integration enhanced signal-to-noise ratio by 8.73%.
• Both deep learning methods, when combined with PCT, improved defect detectability compared to PCT alone.

Conclusions:

• Deep learning-based thermographic data enhancement significantly improves defect detection in artwork.
• CAE shows superior performance in enhancing image quality and defect visibility.
• The proposed integrated approach offers a promising solution for nondestructive artwork inspection.