Scholarlink

<p>To address the challenge of low instances segmentation accuracy in ethnic costumes image caused by occlusion, varying shooting angles, and model poses, this paper proposes a multi-scale feature fusion-based ethnic costumes image instance segmentation method, MRA-Net. In the down-sampling stage, a multi-scale feature extraction method is employed. Moreover, an attention block composed of multiple attention mechanisms is introduced to enhance the model's capacity for capturing detailed features. Inspired by the UNet++ network model's use of dense skip connections to reduce differences in semantic feature mapping, the skip connections are redesigned to reduce the model's computational load and parameters while fusing semantic information across different layers, thereby improving the reusability of feature information. Finally, a multi-scale fusion bilinear interpolation up-sampling method is applied to further enhance multi-scale information representation while preserving image details, thus improving segmentation boundary accuracy. Experimental results show that the proposed method achieves foreground segmentation accuracy (FBD) of 98.21% and prediction instance accuracy (SBD) of 82.41% on the ethnic costumes image dataset, effectively adapting to the complex task of ethnic dress image segmentation. Our code and models will be available at <a href="https://github.com/FanYingJie-star/MRA-Net">https://github.com/FanYingJie-star/MRA-Net</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

MRA-Net: an instance segmentation method based on multi-scale feature fusion for ethnic costumes images

  • 期刊论文

  • 发表日期：

    2025年1月27日

查看全文
  • Yingjie Fan,
  • Bin Wen,
  • Hongfei Deng

摘要

To address the challenge of low instances segmentation accuracy in ethnic costumes image caused by occlusion, varying shooting angles, and model poses, this paper proposes a multi-scale feature fusion-based ethnic costumes image instance segmentation method, MRA-Net. In the down-sampling stage, a multi-scale feature extraction method is employed. Moreover, an attention block composed of multiple attention mechanisms is introduced to enhance the model's capacity for capturing detailed features. Inspired by the UNet++ network model's use of dense skip connections to reduce differences in semantic feature mapping, the skip connections are redesigned to reduce the model's computational load and parameters while fusing semantic information across different layers, thereby improving the reusability of feature information. Finally, a multi-scale fusion bilinear interpolation up-sampling method is applied to further enhance multi-scale information representation while preserving image details, thus improving segmentation boundary accuracy. Experimental results show that the proposed method achieves foreground segmentation accuracy (FBD) of 98.21% and prediction instance accuracy (SBD) of 82.41% on the ethnic costumes image dataset, effectively adapting to the complex task of ethnic dress image segmentation. Our code and models will be available at https://github.com/FanYingJie-star/MRA-Net.