Scholarlink

<p>Woven fabric textures exhibit periodic warp–weft interlacing, strong directionality, and repetitive primitives. In imaging and industrial applications, defects such as dead pixels, contamination, and warp/weft floats may cause random or through-going structured loss, making texture completion an ill-posed inverse problem. Under a matrix completion framework, this study proposes an interpretable joint-prior model that combines low-rank regularization with two-sided orthonormal transform coefficient sparsity (OTCS). OTCS characterizes warp–weft periodic content, while the low-rank prior enforces global structural consistency. A nuclear-norm-based model, Nuc + OTCS, is first developed. To reduce the over-shrinkage of dominant singular values, truncated nuclear norm regularization (TNNR) is further introduced with a two-level optimization strategy of outer principal-subspace extraction and inner trace-compensated reconstruction, yielding TNNR + OTCS. Experiments on 16 woven fabric texture classes show that OTCS outperforms total variation, especially under stripe-shaped through-going loss. With dominant low-rank energy preserved, TNNR + OTCS achieves the best or tied-best results across loss ratios and mask types, and shows stronger robustness in high-loss and structured-loss settings.</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

Low-Rank and Two-Sided Orthonormal Transform Coefficient Sparsity: A Joint Prior Model for Woven Fabric Texture Completion

  • 期刊论文

  • 发表日期：

    2026年6月20日

查看全文
  • Lei Zhang,
  • Qiyu Wei,
  • Zhu Zhan,
  • Jun Wang

摘要

Woven fabric textures exhibit periodic warp–weft interlacing, strong directionality, and repetitive primitives. In imaging and industrial applications, defects such as dead pixels, contamination, and warp/weft floats may cause random or through-going structured loss, making texture completion an ill-posed inverse problem. Under a matrix completion framework, this study proposes an interpretable joint-prior model that combines low-rank regularization with two-sided orthonormal transform coefficient sparsity (OTCS). OTCS characterizes warp–weft periodic content, while the low-rank prior enforces global structural consistency. A nuclear-norm-based model, Nuc + OTCS, is first developed. To reduce the over-shrinkage of dominant singular values, truncated nuclear norm regularization (TNNR) is further introduced with a two-level optimization strategy of outer principal-subspace extraction and inner trace-compensated reconstruction, yielding TNNR + OTCS. Experiments on 16 woven fabric texture classes show that OTCS outperforms total variation, especially under stripe-shaped through-going loss. With dominant low-rank energy preserved, TNNR + OTCS achieves the best or tied-best results across loss ratios and mask types, and shows stronger robustness in high-loss and structured-loss settings.