Inpainting of damaged temple murals using edge- and line-guided diffusion patch GAN.

Area of Science:

• Digital Art Restoration
• Computer Vision
• Artificial Intelligence

Background:

• Ancient mural paintings are invaluable cultural heritage but susceptible to degradation.
• Current digital restoration methods often fail to address large, naturally degraded areas effectively.
• Existing techniques may use inadequate masking strategies, leading to suboptimal repair outcomes.

Purpose of the Study:

• To develop a novel AI architecture for reconstructing extensive naturally degraded regions in mural paintings.
• To maintain intrinsic details, prevent color bias, and preserve the artistic quality of the murals.
• To improve upon existing image inpainting techniques for cultural heritage preservation.

Main Methods:

• Integration of generative adversarial networks (GANs) and diffusion models.
• Utilizing a whole structure formation network (WSFN) with image, line drawing, and edge map inputs.
• Employing a semantic color network (SCN) with gated convolution for textural inpainting.
• Incorporating a diffusion mixture distribution (DIMD) discriminator.
• Extending the receptive field for effective large-area inpainting.

Main Results:

• The proposed model demonstrated superior performance in quantitative analysis compared to state-of-the-art methods.
• Achieved high scores in metrics such as SSIM (0.8853), MSE (0.0021), PSNR (29.8826), and LPIPS (0.0426).
• Successfully reconstructed large, naturally degraded areas while preserving mural details and artistic excellence.

Conclusions:

• The novel GAN and diffusion model architecture offers a significant advancement in digital restoration of ancient murals.
• The method effectively addresses the challenge of large-area inpainting, maintaining fidelity and aesthetic quality.
• This approach provides a promising solution for preserving cultural heritage through advanced AI techniques.