Scholarlink

<p>As fundamental components of Chinese architectural eaves ornaments since the Western Zhou Dynasty, eaves tiles integrate structural protection with artistic value. This study develops an automated dynasty recognition system using an enhanced lightweight EfficientNet-B0. Innovations include integrating Convolutional Block Attention Modules (CBAM) after initial convolution and MBConv6 blocks, replacing Squeeze-and-Excitation (SE) with CBAM, and introducing Widthwise-Transform Convolution (WTConv) to form CWConv blocks. Focal Loss mitigates class imbalance. A comprehensive dataset of 16,312 eaves tile images across eight dynasties from Shang-Zhou to Ming-Qing was constructed using field photography and public museum archives, augmented to address sample scarcity. The optimized model achieves 87.8% accuracy in eight-dynasty classification, surpassing the baseline by 6.3%, with improved precision, recall, and F1-score. It demonstrates robustness for monochromatic, texturally similar tiles under low computational cost, supporting mobile deployment for architectural heritage preservation and historical analysis.</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

A lightweight enhanced EfficientNet model for Chinese eaves tile dynasty classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月6日

查看全文
  • Zongming Liu,
  • Wei Hong,
  • Rui Long,
  • Yue Zhu,
  • Xiaoyu Zhang

摘要

As fundamental components of Chinese architectural eaves ornaments since the Western Zhou Dynasty, eaves tiles integrate structural protection with artistic value. This study develops an automated dynasty recognition system using an enhanced lightweight EfficientNet-B0. Innovations include integrating Convolutional Block Attention Modules (CBAM) after initial convolution and MBConv6 blocks, replacing Squeeze-and-Excitation (SE) with CBAM, and introducing Widthwise-Transform Convolution (WTConv) to form CWConv blocks. Focal Loss mitigates class imbalance. A comprehensive dataset of 16,312 eaves tile images across eight dynasties from Shang-Zhou to Ming-Qing was constructed using field photography and public museum archives, augmented to address sample scarcity. The optimized model achieves 87.8% accuracy in eight-dynasty classification, surpassing the baseline by 6.3%, with improved precision, recall, and F1-score. It demonstrates robustness for monochromatic, texturally similar tiles under low computational cost, supporting mobile deployment for architectural heritage preservation and historical analysis.