Scholarlink

<p>Feature extraction is crucial for pattern recognition tasks, but traditional methods such as PCA and LDA can only handle single feature set. In practical applications, a single pattern often has multiple feature representations. Existing combination strategies (serial and parallel fusion) neglect feature correlations, while canonical correlation analysis (CCA) and its variants can only model correlations between dual feature sets, forcibly decomposing multiple feature sets into two feature sets. While multiset canonical correlation analysis (MCCA) addresses multi-set correlation modeling, it suffers from inefficient solutions to the multivariate eigenvalue problem (MEP) and existing algorithms like power successive overrelaxation (P-SOR) exhibit extreme sensitivity to relaxation factors. To address these issues, this paper proposes a supervised MCCA framework based on the power-symmetric successive overrelaxation (P-SSOR) algorithm (SMCCA/PSSOR). P-SSOR achieves stable solutions by reducing relaxation factor sensitivity. This framework enhances discriminative power by simultaneously maximizing multiple set global correlation and the local inter-class divergence. Experiments on the COIL-100, CENPARMI, MFD, and AR datasets demonstrate that this method outperforms classical multiple set fusion approaches in terms of accuracy.</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

Multiple feature fusion using supervised multiset canonical correlations with power-symmetric successive overrelaxation

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Jing Yang,
  • Xizhan Gao,
  • Quansen Sun,
  • Zhimin Liu

摘要

Feature extraction is crucial for pattern recognition tasks, but traditional methods such as PCA and LDA can only handle single feature set. In practical applications, a single pattern often has multiple feature representations. Existing combination strategies (serial and parallel fusion) neglect feature correlations, while canonical correlation analysis (CCA) and its variants can only model correlations between dual feature sets, forcibly decomposing multiple feature sets into two feature sets. While multiset canonical correlation analysis (MCCA) addresses multi-set correlation modeling, it suffers from inefficient solutions to the multivariate eigenvalue problem (MEP) and existing algorithms like power successive overrelaxation (P-SOR) exhibit extreme sensitivity to relaxation factors. To address these issues, this paper proposes a supervised MCCA framework based on the power-symmetric successive overrelaxation (P-SSOR) algorithm (SMCCA/PSSOR). P-SSOR achieves stable solutions by reducing relaxation factor sensitivity. This framework enhances discriminative power by simultaneously maximizing multiple set global correlation and the local inter-class divergence. Experiments on the COIL-100, CENPARMI, MFD, and AR datasets demonstrate that this method outperforms classical multiple set fusion approaches in terms of accuracy.