Scholarlink

<p>The scarcity of high-value timber and the rise of market fraud highlight the need for efficient wood species identification. Existing methods, including X-ray–based deep learning approaches, remain constrained by low training efficiency and limited dataset sizes, and frequently suffer from severe overfitting in small-sample scenarios. This paper presents the FARR framework, a frozen-feature learning paradigm that integrates multi-block feature aggregation, SE-enhanced residual connections, and randomized autoencoders to better utilize pre-trained representations under small-sample and limited training resource conditions. The proposed cross-layer feature optimization mechanism effectively addresses the lack of hierarchical complementarity, feature enhancement, and robustness in existing frozen-feature methods. Experiments show that FARR achieves 99.86% accuracy and improves training efficiency by 20–68<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation> over partial fine-tuning, while maintaining strong robustness and cross-domain generalization. This work provides an efficient and reliable solution for intelligent wood identification and offers new insights into deep learning under small-sample scenarios.</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

FARR: An efficient frozen-feature learning framework for wood species identification with applications to texture recognition

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Tao Yang,
  • Rigui Zhou,
  • Pengju Ren,
  • Hongpeng Wang

摘要

The scarcity of high-value timber and the rise of market fraud highlight the need for efficient wood species identification. Existing methods, including X-ray–based deep learning approaches, remain constrained by low training efficiency and limited dataset sizes, and frequently suffer from severe overfitting in small-sample scenarios. This paper presents the FARR framework, a frozen-feature learning paradigm that integrates multi-block feature aggregation, SE-enhanced residual connections, and randomized autoencoders to better utilize pre-trained representations under small-sample and limited training resource conditions. The proposed cross-layer feature optimization mechanism effectively addresses the lack of hierarchical complementarity, feature enhancement, and robustness in existing frozen-feature methods. Experiments show that FARR achieves 99.86% accuracy and improves training efficiency by 20–68 \(\times \) × over partial fine-tuning, while maintaining strong robustness and cross-domain generalization. This work provides an efficient and reliable solution for intelligent wood identification and offers new insights into deep learning under small-sample scenarios.