Scholarlink

<p>Murals are considered a magnificent expression of artistic creativity and are located amidst the picturesque landscapes of temples and palaces in Rajasthan. However, their beauty has faded over time due to the relentless forces of nature and human neglect. To breathe new life into these masterpieces through image inpainting, it is crucial to accurately identify the damaged areas. Conventional methods for replicating mural damage are limited to simple dust and spot-like damage and are not realistic. In response to this challenge, the present study introduces an integrated framework that combines deep learning with advanced image processing techniques to generate realistic mural damage. To detect deterioration in mural, this study explicitly examines nine distinct architectures based on eight different backbones. For a thorough comparison, the current investigation also explores the use of a probabilistic ensemble and a generative adversarial network (GAN)-based segmentation network. Overall, Ensemble and GAN achieve an average Structural Similarity Index Measure (SSIM) score of 0.901 and 0.919, respectively, on the test dataset. Qualitatively, the ensemble model demonstrated superior performance for coarse-like damage, while the GAN-based model provided more semantically accurate fine-mask predictions. Following this, an experimental validation is conducted using Thangka murals, demonstrating the efficacy of the proposed framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning-Based Framework for Damage Introduction and Its Segmentation in Ancient Mural Paintings

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Anshul Kumar Yadav,
  • Dhiraj

摘要

Murals are considered a magnificent expression of artistic creativity and are located amidst the picturesque landscapes of temples and palaces in Rajasthan. However, their beauty has faded over time due to the relentless forces of nature and human neglect. To breathe new life into these masterpieces through image inpainting, it is crucial to accurately identify the damaged areas. Conventional methods for replicating mural damage are limited to simple dust and spot-like damage and are not realistic. In response to this challenge, the present study introduces an integrated framework that combines deep learning with advanced image processing techniques to generate realistic mural damage. To detect deterioration in mural, this study explicitly examines nine distinct architectures based on eight different backbones. For a thorough comparison, the current investigation also explores the use of a probabilistic ensemble and a generative adversarial network (GAN)-based segmentation network. Overall, Ensemble and GAN achieve an average Structural Similarity Index Measure (SSIM) score of 0.901 and 0.919, respectively, on the test dataset. Qualitatively, the ensemble model demonstrated superior performance for coarse-like damage, while the GAN-based model provided more semantically accurate fine-mask predictions. Following this, an experimental validation is conducted using Thangka murals, demonstrating the efficacy of the proposed framework.