Scholarlink

<p>Cashmere, wool, and modified wool fibers exhibit significant morphological, chemical, and physicochemical similarities, posing challenges in their accurate differentiation. Traditional convolutional neural networks (CNNs) often struggle to capture global and local features at a single scale and fail to effectively distinguish feature importance across regions. This paper proposes an improved ConvNeXt model for classifying cashmere, wool, and modified wool images. We introduce a simple parameter-free attention mechanism (SimAM) during feature extraction to focus on key regions, such as fiber boundaries and textures. Additionally, an Enhanced Feature Pyramid Network (EFPN) is utilized to integrate multi-scale features from shallow and deep network stages, enhancing feature representativeness. Experimental results on a self-constructed dataset demonstrate a 3.333% increase in accuracy compared to the original model, verifying the effectiveness of our approach in accurately identifying cashmere, wool, and modified wool fibers.</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

Enhanced fiber classification: improved ConvNeXt for cashmere, wool and modified wool identification under limited data conditions

  • 期刊论文

  • 发表日期：

    2026年7月4日

查看全文
  • Yaolin Zhu,
  • Yuhuan Cao,
  • Meihua Gu,
  • Gang Hu,
  • Hong Li,
  • Yunhong Li

摘要

Cashmere, wool, and modified wool fibers exhibit significant morphological, chemical, and physicochemical similarities, posing challenges in their accurate differentiation. Traditional convolutional neural networks (CNNs) often struggle to capture global and local features at a single scale and fail to effectively distinguish feature importance across regions. This paper proposes an improved ConvNeXt model for classifying cashmere, wool, and modified wool images. We introduce a simple parameter-free attention mechanism (SimAM) during feature extraction to focus on key regions, such as fiber boundaries and textures. Additionally, an Enhanced Feature Pyramid Network (EFPN) is utilized to integrate multi-scale features from shallow and deep network stages, enhancing feature representativeness. Experimental results on a self-constructed dataset demonstrate a 3.333% increase in accuracy compared to the original model, verifying the effectiveness of our approach in accurately identifying cashmere, wool, and modified wool fibers.