Scholarlink

<p>Aiming at the problems of low segmentation accuracy and inaccurate mask boundaries in multi-scale clothing instance segmentation, a multi-scale clothing image instance segmentation method based on improved YOLOv8-seg is proposed. First, a Spatial Pyramid Pooling Fast Cross Stage Partial Connections (SPPFCSPC) module is utilized to optimize the backbone network, which prevents different feature layers from learning repetitive gradient information to enrich the feature representation; Second, a boundary refinement strategy is introduced to supplement information lost during the feature extraction process, combining low-level location information with high-level semantic information to enhance the mask boundary quality; Then, a dual attention mechanism is designed to focus on the channel and spatial clothing features information, improving the network’s ability to extract essential features; Finally, Partial Convolution (PConv) and Grouped Shuffle Convolution (GSConv) are integrated to reduce model parameters while maintaining segmentation accuracy. Experimental results demonstrate that the proposed method outperforms the existing advanced approaches on the ModaNet dataset, achieving significant accuracy improvements in multi-scale clothing instance segmentation. The highest accuracy is achieved by using the SPPFCSPC module, the boundary refinement strategy, and the dual attention mechanism. Specifically, the accuracy is improved by more than 1.9%, 1.7%, and 1.8% for small-scale, medium-scale, and large-scale clothing, respectively. Although the optimization increases model complexity and reduces speed, the partial and grouped shuffle convolutions are used to reduce the number of parameters by 0.66&#xa0;M and improve the frames per second by 5.9 img/s. The code is available at <a href="https://github.com/ygp-cloud/SBA-YOLO">https://github.com/ygp-cloud/SBA-YOLO</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

Multi-scale clothing image instance segmentation method based on improved YOLOv8-seg

  • 期刊论文

  • 发表日期：

    2025年12月27日

查看全文
  • Meihua Gu,
  • Guiping Yao,
  • Xiaoxiao Dong,
  • Yang Pan

摘要

Aiming at the problems of low segmentation accuracy and inaccurate mask boundaries in multi-scale clothing instance segmentation, a multi-scale clothing image instance segmentation method based on improved YOLOv8-seg is proposed. First, a Spatial Pyramid Pooling Fast Cross Stage Partial Connections (SPPFCSPC) module is utilized to optimize the backbone network, which prevents different feature layers from learning repetitive gradient information to enrich the feature representation; Second, a boundary refinement strategy is introduced to supplement information lost during the feature extraction process, combining low-level location information with high-level semantic information to enhance the mask boundary quality; Then, a dual attention mechanism is designed to focus on the channel and spatial clothing features information, improving the network’s ability to extract essential features; Finally, Partial Convolution (PConv) and Grouped Shuffle Convolution (GSConv) are integrated to reduce model parameters while maintaining segmentation accuracy. Experimental results demonstrate that the proposed method outperforms the existing advanced approaches on the ModaNet dataset, achieving significant accuracy improvements in multi-scale clothing instance segmentation. The highest accuracy is achieved by using the SPPFCSPC module, the boundary refinement strategy, and the dual attention mechanism. Specifically, the accuracy is improved by more than 1.9%, 1.7%, and 1.8% for small-scale, medium-scale, and large-scale clothing, respectively. Although the optimization increases model complexity and reduces speed, the partial and grouped shuffle convolutions are used to reduce the number of parameters by 0.66 M and improve the frames per second by 5.9 img/s. The code is available at https://github.com/ygp-cloud/SBA-YOLO.