Scholarlink

<p>Palm-leaf manuscripts (PLMs), essential Asian cultural artifacts, necessitate accurate species identification for their preservation. Traditional destructive analysis or manual inspection prone to errors hampers large-scale conservation efforts. To address this, we present PLNet, a lightweight framework integrating Lion-optimized EfficientNetB2 with domain-specific augmentation strategies. Lion optimizer enhances training stability, while task-driven augmentation simulates manuscript degradation. PLNet achieves exceptional performance on a diverse test set: 99.07% accuracy and 99.21% F1-score, outperforming ResNet50, DenseNet169, and YOLOv11-cls. Crucially, PLNet operates with high efficiency, utilizing only 7.7 million parameters and processing a folio in 0.31 s. To enable practical application, we developed PLNet-GUI for large-scale identification. Validated on 142,679 PLMs from 8 countries, the framework demonstrates robust real-world applicability, reveals regional species distribution patterns, and establishes a method shift in heritage conservation by replacing destructive practices with AI-driven preservation.</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 lion-optimized efficientnet framework for non-destructive and rapid plant species identification in palm-leaf manuscripts

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月30日

查看全文
  • Xiao Yang,
  • Song Chen,
  • Lin Tan,
  • Yue Wang,
  • Feng Gao,
  • Zhimin Zhang,
  • Xiao Zhou,
  • Hongmei Lu

摘要

Palm-leaf manuscripts (PLMs), essential Asian cultural artifacts, necessitate accurate species identification for their preservation. Traditional destructive analysis or manual inspection prone to errors hampers large-scale conservation efforts. To address this, we present PLNet, a lightweight framework integrating Lion-optimized EfficientNetB2 with domain-specific augmentation strategies. Lion optimizer enhances training stability, while task-driven augmentation simulates manuscript degradation. PLNet achieves exceptional performance on a diverse test set: 99.07% accuracy and 99.21% F1-score, outperforming ResNet50, DenseNet169, and YOLOv11-cls. Crucially, PLNet operates with high efficiency, utilizing only 7.7 million parameters and processing a folio in 0.31 s. To enable practical application, we developed PLNet-GUI for large-scale identification. Validated on 142,679 PLMs from 8 countries, the framework demonstrates robust real-world applicability, reveals regional species distribution patterns, and establishes a method shift in heritage conservation by replacing destructive practices with AI-driven preservation.