<p>The historical rock-cut structures in the Cappadocia region are unique repositories of cultural heritage with their priceless murals and motifs. Some of the murals and motifs in these structures have lost their visual integrity due to the soot layer formed by fires and linseed oil based candles burnt inside. Due to the high cost of physical and chemical restoration processes and the possibility of damaging the original murals and motifs, nondestructive digital restoration methods have recently gained prominence. This study aims to see the back of the soot layer on darkened hematite based motifs. A modified Pix2Pix architecture based on a Conditional Generative Adversarial Network (cGAN) is proposed to preserve the structural integrity of the motifs. In this study, 89 natural rock samples, consistent with the lithological characteristics of the region, were painted with hematite pigment and sooted with linseed oil in a laboratory environment to create a novel dataset. This approach, by directly grounding the training process of the developed model on real world field conditions, has taken restoration performance beyond theoretical simulation. To overcome the blurring and texture loss caused by the standard Pix2Pix network, Structural Similarity Index (SSIM) loss is integrated into the training process, and an optimal balance is established with the L1 error function using <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\lambda _{L1} = 100, \lambda _{SSIM} = 5$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msub> <mi>λ</mi> <mrow> <mi>L</mi> <mn>1</mn> </mrow> </msub> <mo>=</mo> <mn>100</mn> <mo>,</mo> <msub> <mi>λ</mi> <mrow> <mi mathvariant="italic">SSIM</mi> </mrow> </msub> <mo>=</mo> <mn>5</mn> </mrow> </math></EquationSource> </InlineEquation>. Experimental results show that the proposed model exhibits superior performance compared to the standard Pix2Pix model, achieving an increase of up to 0.45 dB in Y-PSNR and 0.027 in SSIM. NIQE and BRISQUE based perceptual quality analyses also confirm that the model produces the most natural looking results for the human eye. This study offers art historians and restorers a highly effective and reliable digital tool to assist in restoration of sooty motifs.</p>

Virtual soot removal on red painted motifs in rock-cut structures: A deep learning approach for preserving structural integrity using a novel dataset

期刊论文
发表日期：

2026年5月15日

查看全文

Bilgin Yazlık,
Egemen Yazlık

摘要

The historical rock-cut structures in the Cappadocia region are unique repositories of cultural heritage with their priceless murals and motifs. Some of the murals and motifs in these structures have lost their visual integrity due to the soot layer formed by fires and linseed oil based candles burnt inside. Due to the high cost of physical and chemical restoration processes and the possibility of damaging the original murals and motifs, nondestructive digital restoration methods have recently gained prominence. This study aims to see the back of the soot layer on darkened hematite based motifs. A modified Pix2Pix architecture based on a Conditional Generative Adversarial Network (cGAN) is proposed to preserve the structural integrity of the motifs. In this study, 89 natural rock samples, consistent with the lithological characteristics of the region, were painted with hematite pigment and sooted with linseed oil in a laboratory environment to create a novel dataset. This approach, by directly grounding the training process of the developed model on real world field conditions, has taken restoration performance beyond theoretical simulation. To overcome the blurring and texture loss caused by the standard Pix2Pix network, Structural Similarity Index (SSIM) loss is integrated into the training process, and an optimal balance is established with the L1 error function using $\lambda _{L1} = 100, \lambda _{SSIM} = 5$ $λ_{L 1} = 100, λ_{SSIM} = 5$ . Experimental results show that the proposed model exhibits superior performance compared to the standard Pix2Pix model, achieving an increase of up to 0.45 dB in Y-PSNR and 0.027 in SSIM. NIQE and BRISQUE based perceptual quality analyses also confirm that the model produces the most natural looking results for the human eye. This study offers art historians and restorers a highly effective and reliable digital tool to assist in restoration of sooty motifs.