Scholarlink

<p>Traditional methods for detecting foreign matter in plastic powder rely on magnetic bars to remove magnetic contaminants and human visual inspection for non-magnetic ones. However, magnetic bars cannot capture non-magnetic contaminants. Visual inspection is also limited, as it can only detect contaminants larger than about 100&#xa0;μm. Finer powders and smaller foreign matters therefore require magnifying tools. Recently, deep learning-based vision inspection has begun to replace these traditional methods. Yet industries that produce multiple, visually similar powders must still build separate models for each type. Manual annotation for every model is time-consuming and labor-intensive, which reduces overall adaptability. This study introduces a residual correlation alignment (ResCORAL) model in the field of domain adaptation (DA). ResCORAL learns domain-invariant feature representations, which improves predictive performance across different domains. A key advantage of DA is that only source domain images need annotation. The target domain requires no labeling, allowing the model to transfer directly to new domains. We use superabsorbent polymers (SAP) and polyvinyl chloride (PVC) as experimental examples. The proposed method rapidly builds a cross-powder defect classification mechanism that enhances inspection efficiency and adaptability.</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

Domain adaptation for foreign matter inspection in plastic powders

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Yen-Ting T. Chou,
  • Kai-Wen Lu,
  • Ssu-Han Chen

摘要

Traditional methods for detecting foreign matter in plastic powder rely on magnetic bars to remove magnetic contaminants and human visual inspection for non-magnetic ones. However, magnetic bars cannot capture non-magnetic contaminants. Visual inspection is also limited, as it can only detect contaminants larger than about 100 μm. Finer powders and smaller foreign matters therefore require magnifying tools. Recently, deep learning-based vision inspection has begun to replace these traditional methods. Yet industries that produce multiple, visually similar powders must still build separate models for each type. Manual annotation for every model is time-consuming and labor-intensive, which reduces overall adaptability. This study introduces a residual correlation alignment (ResCORAL) model in the field of domain adaptation (DA). ResCORAL learns domain-invariant feature representations, which improves predictive performance across different domains. A key advantage of DA is that only source domain images need annotation. The target domain requires no labeling, allowing the model to transfer directly to new domains. We use superabsorbent polymers (SAP) and polyvinyl chloride (PVC) as experimental examples. The proposed method rapidly builds a cross-powder defect classification mechanism that enhances inspection efficiency and adaptability.