Scholarlink

<p>Enzymatic oxidation (EO) is a core process in Tieguanyin tea processing, playing a decisive role in determining its color, aroma, and taste. However, automatic recognition of EO stages remains a highly challenging task. This study proposes TGR-EOSR, an improved model based on ShuffleNet V2. It is designed to achieve high recognition accuracy across six key EO stages. Specifically, the coordinate attention mechanism is introduced to enhance the model’s focus on high-level features, such as the distribution of brown spots on tea shoots. A multi-scale feature extraction module is employed to improve the capture of low-level features, including leaf color and morphological details. Additionally, the combined use of ReLU6 and residual connections suppresses excessively large activations, enhances robustness, and accelerates convergence. The Ghost module is incorporated to reduce parameters and enable a lightweight design. The experimental results demonstrate that TGR-EOSR substantially outperforms ShuffleNet V2. It achieves an EO stage recognition accuracy of 93.93% and an F1-score of 93.95%, with only 3.902&#xa0;M parameters. The proposed model reduces reliance on the subjective expertise of tea masters and enhances the standardization and quality stability of tea production. It provides a practical and feasible solution for intelligent tea processing.</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

Enzymatic oxidation stage recognition method for Tieguanyin tea based on improved ShuffleNet V2

  • 期刊论文

  • 发表日期：

    2025年10月22日

查看全文
  • Wei Chen,
  • Shengwei Xia,
  • Yuyan Huang,
  • Qiwei Fang,
  • Yongkuai Chen

摘要

Enzymatic oxidation (EO) is a core process in Tieguanyin tea processing, playing a decisive role in determining its color, aroma, and taste. However, automatic recognition of EO stages remains a highly challenging task. This study proposes TGR-EOSR, an improved model based on ShuffleNet V2. It is designed to achieve high recognition accuracy across six key EO stages. Specifically, the coordinate attention mechanism is introduced to enhance the model’s focus on high-level features, such as the distribution of brown spots on tea shoots. A multi-scale feature extraction module is employed to improve the capture of low-level features, including leaf color and morphological details. Additionally, the combined use of ReLU6 and residual connections suppresses excessively large activations, enhances robustness, and accelerates convergence. The Ghost module is incorporated to reduce parameters and enable a lightweight design. The experimental results demonstrate that TGR-EOSR substantially outperforms ShuffleNet V2. It achieves an EO stage recognition accuracy of 93.93% and an F1-score of 93.95%, with only 3.902 M parameters. The proposed model reduces reliance on the subjective expertise of tea masters and enhances the standardization and quality stability of tea production. It provides a practical and feasible solution for intelligent tea processing.