Scholarlink

<p>When sweat is transported by liquid in garment fabrics, moisture absorption and quick drying performance of fabrics will directly affect the comfort of wearer. Therefore, accurately characterizing the transport of liquid water in garment fabrics is important. Currently, there are many solutions for measuring liquid water transport in fabrics by machine vision, which utilize fixed attention mechanism to segment the wet zone in fabrics. However, these methods generally with higher calculation cost suffer from various problems such as low segmentation accuracy and poor robustness due to fabric texture and light reflection. Here, a strategy to flexibly adjust attention is proposed by adding <i>Modified Select Kernel</i> (<i>MSK</i>) unit and using <i>Cross-Shapes Windows Transformer</i> (<i>CS Win</i>) for <i>Convolution-BatchNorm-LeakyReLU</i> (CBL) in <i>YOLOv5</i>. <i>MSK</i> unit is used to adaptively allocate feature weights, while <i>CS Win</i> dynamically tunes attention by replacing the last two ConvModule modules of <i>YOLOv5</i> backbone. In addition, a higher-accuracy segmentation method named detection and segmentation (<i>DAS</i>) is proposed, which can avoid the interference of dry zone via inputting box-position prompt from improved <i>YOLOv5</i> output into segmentation algorithm. The effectiveness experiments show that improved <i>YOLOv5</i> can greatly raise detection accuracy of small- and medium-sized wet zones in fabrics, the highest <i>P</i><sub><i>mAP</i></sub> (50–95) can reach 94.0%, and regarding <i>DAS</i>, the segmentation accuracy can be 94.08%. It is dual effective that the results are verified using the improved <i>YOLOv5</i> or <i>DAS</i> method. Finally, some suggestions for segmentation model and fabric selection are proposed. The source code for the complete system can be found at <a href="https://github.com/dontlikeEncoding/DAS2">https://github.com/dontlikeEncoding/DAS2</a>.</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

An Efficient Segmentation Method for Fabric Wetted Region with Improved YOLOv5 Through a Flexible Attention Strategy

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • Yucheng Wang,
  • Huiping Wang,
  • Hang Mao,
  • Suwei Gao,
  • Yuan Zeng,
  • Shujing Li,
  • Rangtong Liu

摘要

When sweat is transported by liquid in garment fabrics, moisture absorption and quick drying performance of fabrics will directly affect the comfort of wearer. Therefore, accurately characterizing the transport of liquid water in garment fabrics is important. Currently, there are many solutions for measuring liquid water transport in fabrics by machine vision, which utilize fixed attention mechanism to segment the wet zone in fabrics. However, these methods generally with higher calculation cost suffer from various problems such as low segmentation accuracy and poor robustness due to fabric texture and light reflection. Here, a strategy to flexibly adjust attention is proposed by adding Modified Select Kernel (MSK) unit and using Cross-Shapes Windows Transformer (CS Win) for Convolution-BatchNorm-LeakyReLU (CBL) in YOLOv5. MSK unit is used to adaptively allocate feature weights, while CS Win dynamically tunes attention by replacing the last two ConvModule modules of YOLOv5 backbone. In addition, a higher-accuracy segmentation method named detection and segmentation (DAS) is proposed, which can avoid the interference of dry zone via inputting box-position prompt from improved YOLOv5 output into segmentation algorithm. The effectiveness experiments show that improved YOLOv5 can greatly raise detection accuracy of small- and medium-sized wet zones in fabrics, the highest PmAP (50–95) can reach 94.0%, and regarding DAS, the segmentation accuracy can be 94.08%. It is dual effective that the results are verified using the improved YOLOv5 or DAS method. Finally, some suggestions for segmentation model and fabric selection are proposed. The source code for the complete system can be found at https://github.com/dontlikeEncoding/DAS2.