Scholarlink

<p>Image representation learning techniques aim to extract meaningful features from high-dimensional image data to enhance performance in downstream clustering and classification tasks. Recently, low-rank representation (LRR) methods have shown promise for uncovering the hidden low-dimensional subspace structure embedded in high-dimensional data. However, real-world data often deviates from LRR's idealized assumption that similar samples reside closely in the feature space. Specifically, data corruption can distort the spatial relationships in data, potentially misleading LRR into incorrectly interpreting corrupt samples as similar to samples from different classes if they stay close together, resulting in negative correlation and sub-optimal clustering outcomes. In this paper, we propose a novel method, which uses a low-rank consistency regularization (LCR) to overcome this limitation. LCR is introduced as a dual regularization term into the classical LRR model. The aim is to adaptively find such optimal low-rank representation that significantly minimizes the distance between similar samples in the feature space. Thus, a flexible similarity matrix is introduced simultaneously to adaptively capture an accurate similarity between samples. Unlike existing methods, this similarity matrix is employed directly for clustering by imposing a rank constraint on its Laplacian matrix. Experimental results on multiple benchmark image datasets show that our method is more efficient than state-of-the-art LRR approaches. Additionally, our method exhibits greater robustness to corruption across various experimental conditions.</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

Robust image representation learning via low-rank consistency regularization for subspace clustering

  • 期刊论文

  • 发表日期：

    2025年2月16日

查看全文
  • Stanley Ebhohimhen Abhadiomhen,
  • George Emeka Okereke,
  • Royransom Chiemela Nzeh,
  • Nnamdi Johnson Ezeora,
  • Abel Onolunosen Abhadionmhen,
  • Caroline Ngozi Asogwa

摘要

Image representation learning techniques aim to extract meaningful features from high-dimensional image data to enhance performance in downstream clustering and classification tasks. Recently, low-rank representation (LRR) methods have shown promise for uncovering the hidden low-dimensional subspace structure embedded in high-dimensional data. However, real-world data often deviates from LRR's idealized assumption that similar samples reside closely in the feature space. Specifically, data corruption can distort the spatial relationships in data, potentially misleading LRR into incorrectly interpreting corrupt samples as similar to samples from different classes if they stay close together, resulting in negative correlation and sub-optimal clustering outcomes. In this paper, we propose a novel method, which uses a low-rank consistency regularization (LCR) to overcome this limitation. LCR is introduced as a dual regularization term into the classical LRR model. The aim is to adaptively find such optimal low-rank representation that significantly minimizes the distance between similar samples in the feature space. Thus, a flexible similarity matrix is introduced simultaneously to adaptively capture an accurate similarity between samples. Unlike existing methods, this similarity matrix is employed directly for clustering by imposing a rank constraint on its Laplacian matrix. Experimental results on multiple benchmark image datasets show that our method is more efficient than state-of-the-art LRR approaches. Additionally, our method exhibits greater robustness to corruption across various experimental conditions.