Scholarlink

<p>Historical place image classification is a crucial intersection of artificial intelligence and cultural heritage preservation, essential for safeguarding architectural legacies in the digital age. Current AI classification systems often struggle with generalizability across regions and styles, require substantial computational resources, and lack interpretability for heritage professionals. This research introduces HistoNet, a hybrid deep learning framework designed to overcome these limitations. HistoNet integrates convolutional neural networks for local feature extraction and Transformer mechanisms for global context, while also using Mamba state-space models, which operate with linear time complexity O(L), thereby offering substantially higher computational efficiency than Transformer architectures that scale quadratically O(<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {L}^{2}\)</EquationSource> </InlineEquation>). Enhanced with Squeeze-and-Excitation attention blocks and SHapley Additive exPlanations based interpretability, it provides pixel-level attribution maps for more accurate validation of classification decisions. Deployed through a user-friendly web interface with drag-and-drop functionality, HistoNet democratizes access by enabling researchers, conservationists, educators, and heritage professionals in resource-constrained regions to leverage advanced AI tools without requiring high-end hardware or specialized expertise. Testing shows that it achieves 95.66% accuracy on the Architectural Heritage Elements dataset and 96.46% on the Historical Building Recognition dataset, significantly outperforming models such as VGG16 and ResNet50. Moreover, the system demonstrates robust generalization, maintaining stable performance across Malacca’s diverse lighting conditions, weather variations, and architectural styles, thereby ensuring practical reliability in real-world heritage environments. By combining efficiency, accessibility, and robustness, this research makes advanced heritage classification broadly usable and reliable, promoting sustainable development goals in cultural heritage conservation and bridging digital divides.</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 novel deep neural model for efficient and scalable historical place image classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月28日

查看全文
  • Md All Moon Tasir,
  • Mohd Nor Akmal Khalid

摘要

Historical place image classification is a crucial intersection of artificial intelligence and cultural heritage preservation, essential for safeguarding architectural legacies in the digital age. Current AI classification systems often struggle with generalizability across regions and styles, require substantial computational resources, and lack interpretability for heritage professionals. This research introduces HistoNet, a hybrid deep learning framework designed to overcome these limitations. HistoNet integrates convolutional neural networks for local feature extraction and Transformer mechanisms for global context, while also using Mamba state-space models, which operate with linear time complexity O(L), thereby offering substantially higher computational efficiency than Transformer architectures that scale quadratically O( \(\hbox {L}^{2}\) ). Enhanced with Squeeze-and-Excitation attention blocks and SHapley Additive exPlanations based interpretability, it provides pixel-level attribution maps for more accurate validation of classification decisions. Deployed through a user-friendly web interface with drag-and-drop functionality, HistoNet democratizes access by enabling researchers, conservationists, educators, and heritage professionals in resource-constrained regions to leverage advanced AI tools without requiring high-end hardware or specialized expertise. Testing shows that it achieves 95.66% accuracy on the Architectural Heritage Elements dataset and 96.46% on the Historical Building Recognition dataset, significantly outperforming models such as VGG16 and ResNet50. Moreover, the system demonstrates robust generalization, maintaining stable performance across Malacca’s diverse lighting conditions, weather variations, and architectural styles, thereby ensuring practical reliability in real-world heritage environments. By combining efficiency, accessibility, and robustness, this research makes advanced heritage classification broadly usable and reliable, promoting sustainable development goals in cultural heritage conservation and bridging digital divides.