Deep learning for terahertz image denoising in nondestructive historical document analysis.

Area of Science:

• Non-destructive analysis
• Cultural heritage preservation
• Advanced imaging techniques

Background:

• Historical documents are fragile and often unreadable due to aging.
• Terahertz (THz) imaging offers non-destructive 3D visualization of hidden contents.
• Standard THz reconstruction algorithms introduce noise and artifacts, limiting image quality.

Purpose of the Study:

• To improve the quality of Terahertz (THz) images for historical document analysis.
• To overcome data scarcity challenges in training deep learning models for THz image enhancement.
• To enable clearer recognition of text in fragile historical documents using deep learning.

Main Methods:

• Utilized unsupervised CycleGAN to generate synthetic noisy-Terahertz (THz) images from clean ones.
• Trained a supervised Pix2pixGAN model using synthetic data for denoising real THz images.
• Applied deep learning-based image enhancement to improve the clarity of historical document scans.

Main Results:

• Achieved 99% character recognition on Xuan paper and 61% on standard paper after Pix2pixGAN denoising.
• Processed images showed perceptual indices (16.83) very close to clean handwriting images (16.19).
• Demonstrated significant improvement in THz image quality for historical document analysis.

Conclusions:

• Deep learning, particularly Pix2pixGAN, effectively enhances Terahertz (THz) images of historical documents.
• The developed method significantly improves the readability of hidden text in fragile artifacts.
• This approach holds substantial value for the non-destructive analysis and preservation of cultural heritage.