Scholarlink

<p>Existing self-representative subspace learning frameworks for joint dimensionality reduction have achieved notable success in image clustering. However, these methods often face challenges such as static low-dimensional features, information loss during mapping, and limited capture of data’s full structural detail. To address these issues, we introduce essential graph-embedded dual mapping subspace learning (EGEDMSL). EGEDMSL employs an elementary graph learning (EGL) strategy to minimize information loss while preserving data stability, adaptively capturing diverse structures within the projection space. By reconstructing projective representations in the original dimensional space, EGEDMSL retains intrinsic structural features and enhances robustness. The projection matrix is refined through a generalized Stiefel manifold, and a block-diagonal regularizer enhances the discriminative quality of the subspace. Experimental results on real datasets demonstrate that EGEDMSL significantly improves clustering performance, offering a promising approach for image clustering tasks. The code for EGEDMSL is available at <a href="https://github.com/zgt0910/EGEDMSL">https://github.com/zgt0910/EGEDMSL</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

Essential graph-embedded dual mapping subspace learning for enhanced image clustering

  • 期刊论文

  • 发表日期：

    2025年5月31日

查看全文
  • Gang Zhu,
  • Lixin Han,
  • Hong Yan

摘要

Existing self-representative subspace learning frameworks for joint dimensionality reduction have achieved notable success in image clustering. However, these methods often face challenges such as static low-dimensional features, information loss during mapping, and limited capture of data’s full structural detail. To address these issues, we introduce essential graph-embedded dual mapping subspace learning (EGEDMSL). EGEDMSL employs an elementary graph learning (EGL) strategy to minimize information loss while preserving data stability, adaptively capturing diverse structures within the projection space. By reconstructing projective representations in the original dimensional space, EGEDMSL retains intrinsic structural features and enhances robustness. The projection matrix is refined through a generalized Stiefel manifold, and a block-diagonal regularizer enhances the discriminative quality of the subspace. Experimental results on real datasets demonstrate that EGEDMSL significantly improves clustering performance, offering a promising approach for image clustering tasks. The code for EGEDMSL is available at https://github.com/zgt0910/EGEDMSL.