Scholarlink

<p>To address the limitations of existing sketch methods for damaged murals, such as insufficient detail preservation, edge blurring, and artifact interference, we propose a mural sketch extraction method based on Multi-Scale Feature Extraction and Gradient Attention Mechanism (MSFE-GAM). The method employs a Multi-Scale Feature Enhancement Module (MSFE) during the downsampling stage to capture cross-scale line details. It integrates gradient information as prior knowledge through a Gradient Attention Module (GAM), guiding the network to focus on critical line regions. In the feature fusion stage, an Adaptive Fusion Module (AFM) dynamically adjusts gradient attention weights via learnable parameters, balancing detail retention and background noise suppression. To address domain distribution discrepancies in damaged murals, an Unsupervised Domain Adaptation (UDA) strategy is proposed, enabling test-sample-driven parameter fine-tuning to enhance adaptability to unseen domain data. Experimental results demonstrate that compared with existing methods, this method improves the SSIM index by more than 7%, increases the AP value by 2.8%, and reduces the RMSE to 0.2071. The generated sketches exhibit rich details and clean backgrounds, particularly outperforming existing methods in complex damage scenarios. Additionally, the UDA strategy enables robustness in cross-domain generalization on unseen samples like Thangka murals, validating its applicability in data-scarce domains. This study provides an innovative, generalizable technical framework for cultural heritage digitization, with ablation experiments and cross-domain analysis confirming its practical utility in addressing damaged murals and data-limited scenarios.</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

Multi-scale feature extraction and gradient attention-based method for sketch extraction of painted cultural relics

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Shenglin Peng,
  • Shan Cui,
  • Xingguo Zhao,
  • Jun Wang,
  • Shuyi Qu,
  • Xianlin Peng

摘要

To address the limitations of existing sketch methods for damaged murals, such as insufficient detail preservation, edge blurring, and artifact interference, we propose a mural sketch extraction method based on Multi-Scale Feature Extraction and Gradient Attention Mechanism (MSFE-GAM). The method employs a Multi-Scale Feature Enhancement Module (MSFE) during the downsampling stage to capture cross-scale line details. It integrates gradient information as prior knowledge through a Gradient Attention Module (GAM), guiding the network to focus on critical line regions. In the feature fusion stage, an Adaptive Fusion Module (AFM) dynamically adjusts gradient attention weights via learnable parameters, balancing detail retention and background noise suppression. To address domain distribution discrepancies in damaged murals, an Unsupervised Domain Adaptation (UDA) strategy is proposed, enabling test-sample-driven parameter fine-tuning to enhance adaptability to unseen domain data. Experimental results demonstrate that compared with existing methods, this method improves the SSIM index by more than 7%, increases the AP value by 2.8%, and reduces the RMSE to 0.2071. The generated sketches exhibit rich details and clean backgrounds, particularly outperforming existing methods in complex damage scenarios. Additionally, the UDA strategy enables robustness in cross-domain generalization on unseen samples like Thangka murals, validating its applicability in data-scarce domains. This study provides an innovative, generalizable technical framework for cultural heritage digitization, with ablation experiments and cross-domain analysis confirming its practical utility in addressing damaged murals and data-limited scenarios.