Scholarlink

<p>Surface defect segmentation is crucial in industrial manufacturing for ensuring product quality. Traditional methods face challenges such as visual ambiguity and data scarcity. We introduce PRIF, a novel semi-supervised framework leveraging prototypical rectification with interpretable fusion. PRIF dynamically focuses on ambiguous regions through a multi-source uncertainty mechanism, integrating epistemic, structural, and divergence-based uncertainties. Our architecture features a PRIF block with a path aggregation nexus for long-range dependency modeling and a bilateral cross-attention Ffsion module for superior feature integration. With only 3.8M parameters, PRIF is highly efficient and enables real-time deployment in industrial environments. Extensive experiments on five diverse benchmarks demonstrate PRIF’s robustness and efficiency, achieving state-of-the-art performance with 110 FPS on an NVIDIA RTX 4060 GPU. Visualizations confirm enhanced interpretability and well-calibrated uncertainty estimates.</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

Prototypical rectification with interpretable fusion for enhanced industrial defect segmentation

  • 期刊论文

  • 发表日期：

    2026年1月23日

查看全文
  • Chuanning Wang,
  • Zhiyong He,
  • Yi Guo,
  • Yanzhao Zhou,
  • Zhenxiong Gu,
  • Song Lin,
  • Mei Lin

摘要

Surface defect segmentation is crucial in industrial manufacturing for ensuring product quality. Traditional methods face challenges such as visual ambiguity and data scarcity. We introduce PRIF, a novel semi-supervised framework leveraging prototypical rectification with interpretable fusion. PRIF dynamically focuses on ambiguous regions through a multi-source uncertainty mechanism, integrating epistemic, structural, and divergence-based uncertainties. Our architecture features a PRIF block with a path aggregation nexus for long-range dependency modeling and a bilateral cross-attention Ffsion module for superior feature integration. With only 3.8M parameters, PRIF is highly efficient and enables real-time deployment in industrial environments. Extensive experiments on five diverse benchmarks demonstrate PRIF’s robustness and efficiency, achieving state-of-the-art performance with 110 FPS on an NVIDIA RTX 4060 GPU. Visualizations confirm enhanced interpretability and well-calibrated uncertainty estimates.