Scholarlink

<p>Traditional authentication of Yue Kiln celadon, an archaeologically significant early Chinese porcelain, relies heavily on expert connoisseurship and therefore remains vulnerable to subjectivity, limited scalability, and reduced efficiency when subtle inter-dynasty variations must be evaluated. To address the combined challenges of data scarcity, fine-grained discrimination, and deployment constraints, this study develops and validates HDBN-LGA, a lightweight deep-learning framework integrating a hybrid dual-branch network, a Local-Global Attention mechanism, and dynamic multi-scale feature fusion. The model was trained and evaluated on a curated dataset of 3,000 high-resolution images, expanded to 9,000 through augmentation, spanning Tang, Five Dynasties, and Bei Song Yue Kiln celadon. HDBN-LGA achieved an average classification accuracy of 95.7%, outperforming ViT-Base (93.5%), EfficientNet-B7 (94.8%), ResNet-50 (93.1%), and SIFT/SURF (82.9%), while using only 12&#xa0;M parameters and supporting 50 ms inference per image on an NVIDIA Tesla V100 GPU. Compared with large-scale vision models, the proposed framework therefore offers a stronger balance of accuracy, speed, and parameter efficiency, making it more suitable for resource-constrained museum and archaeological applications. Robustness experiments under noise and blur further confirmed its stability against image degradation. By combining detail-sensitive attention, adaptive multi-scale fusion, and lightweight architectural design, HDBN-LGA provides an efficient, objective, and practical methodology for automated dynasty identification and cultural heritage digitization.</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

Research on dynasty identification of Yue Kiln Celadon using multi-scale feature fusion and deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月1日

查看全文
  • Changyong Zhu,
  • Xue Bai,
  • Chunxia Zhang,
  • Jiumei Shen

摘要

Traditional authentication of Yue Kiln celadon, an archaeologically significant early Chinese porcelain, relies heavily on expert connoisseurship and therefore remains vulnerable to subjectivity, limited scalability, and reduced efficiency when subtle inter-dynasty variations must be evaluated. To address the combined challenges of data scarcity, fine-grained discrimination, and deployment constraints, this study develops and validates HDBN-LGA, a lightweight deep-learning framework integrating a hybrid dual-branch network, a Local-Global Attention mechanism, and dynamic multi-scale feature fusion. The model was trained and evaluated on a curated dataset of 3,000 high-resolution images, expanded to 9,000 through augmentation, spanning Tang, Five Dynasties, and Bei Song Yue Kiln celadon. HDBN-LGA achieved an average classification accuracy of 95.7%, outperforming ViT-Base (93.5%), EfficientNet-B7 (94.8%), ResNet-50 (93.1%), and SIFT/SURF (82.9%), while using only 12 M parameters and supporting 50 ms inference per image on an NVIDIA Tesla V100 GPU. Compared with large-scale vision models, the proposed framework therefore offers a stronger balance of accuracy, speed, and parameter efficiency, making it more suitable for resource-constrained museum and archaeological applications. Robustness experiments under noise and blur further confirmed its stability against image degradation. By combining detail-sensitive attention, adaptive multi-scale fusion, and lightweight architectural design, HDBN-LGA provides an efficient, objective, and practical methodology for automated dynasty identification and cultural heritage digitization.