Scholarlink

<p>The protection of traditional residential landscapes faces significant challenges due to complex environmental factors such as diverse architectural styles, intricate structural details, varying lighting conditions, and occlusions from natural elements like vegetation. These complexities lead to difficulties in manual annotation, low efficiency, and inconsistent accuracy in architectural element identification and style classification. To address these issues, this paper implements an art knowledge management system for traditional residential landscape identification and classification based on the YOLOv5 object detection model and the ResNet-50 classification network, combined with natural language generation (NLG) technology. First, YOLOv5 is employed to identify and locate architectural elements (e.g., windows, doors, roofs) in traditional residential images. Subsequently, ResNet-50 classifies the style and era of the detected elements, with performance enhanced through data augmentation and optimizer adjustments. Finally, a template engine extracts relevant knowledge from an art knowledge base, and NLG technology generates concise and natural art description text, enabling automatic knowledge generation and display. Experimental results demonstrate that our method achieves a mAP of 0.96, an F1 score between 0.96 and 0.97, accuracy up to 0.97, recall of 0.92, and classification error below 0.3. In cross-cultural architectural style classification, the accuracy difference (Δ Acc) is only 3.6%, effectively overcoming cultural differences. Moreover, under mixed noise interference, the classification accuracy reaches 82.5%, indicating strong robustness. The annotation consistency between our system and expert annotations for traditional residential buildings across different regions remains above 96%. These results confirm that the system can automatically identify architectural elements, generate style- and era-related knowledge, and provide robust technical support for the digital protection and innovation of traditional residential landscapes.</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

Artificial intelligence-driven art knowledge management system: protection and innovation of traditional residential landscape

  • 期刊论文

  • 发表日期：

    2026年7月18日

查看全文
  • Gengrui Li,
  • Wei Dang,
  • Dongxu Yang

摘要

The protection of traditional residential landscapes faces significant challenges due to complex environmental factors such as diverse architectural styles, intricate structural details, varying lighting conditions, and occlusions from natural elements like vegetation. These complexities lead to difficulties in manual annotation, low efficiency, and inconsistent accuracy in architectural element identification and style classification. To address these issues, this paper implements an art knowledge management system for traditional residential landscape identification and classification based on the YOLOv5 object detection model and the ResNet-50 classification network, combined with natural language generation (NLG) technology. First, YOLOv5 is employed to identify and locate architectural elements (e.g., windows, doors, roofs) in traditional residential images. Subsequently, ResNet-50 classifies the style and era of the detected elements, with performance enhanced through data augmentation and optimizer adjustments. Finally, a template engine extracts relevant knowledge from an art knowledge base, and NLG technology generates concise and natural art description text, enabling automatic knowledge generation and display. Experimental results demonstrate that our method achieves a mAP of 0.96, an F1 score between 0.96 and 0.97, accuracy up to 0.97, recall of 0.92, and classification error below 0.3. In cross-cultural architectural style classification, the accuracy difference (Δ Acc) is only 3.6%, effectively overcoming cultural differences. Moreover, under mixed noise interference, the classification accuracy reaches 82.5%, indicating strong robustness. The annotation consistency between our system and expert annotations for traditional residential buildings across different regions remains above 96%. These results confirm that the system can automatically identify architectural elements, generate style- and era-related knowledge, and provide robust technical support for the digital protection and innovation of traditional residential landscapes.