Scholarlink

<p>The interpretability of machine vision algorithms in industrial applications remains an escalating challenge to the reliable deployment of intelligent manufacturing, particularly given the complexity and diversity of traditional vision techniques. This manuscript proposes a novel visual explanation method based on the fusion of saliency maps to address the interpretability deficiency and enhance it accordingly. The proposed method generates robust visual explanations compatible with industrial applications by integrating weighted summation of output scores and data correlation analysis under encoded perturbations. Experiments on real-world industrial samples and the MVTec Industrial Anomaly Detection dataset demonstrate that the approach accurately highlights essential pattern regions and reflects the focus of algorithms, thereby improving the comprehension and trustworthiness of industrial machine vision applications. The proposed method contributes to the intuitive understanding of effectual machine vision mechanisms, mitigates potential misjudgment risks, and facilitates the scientific management of commercial vision integrations.</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

Enhancing industrial machine vision interpretability through saliency maps fusion

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年6月20日

查看全文
  • Xiaoshun Xu,
  • Xuchun Gan,
  • Tao Zhou,
  • Yunfei Zhang,
  • Jinqiu Mo

摘要

The interpretability of machine vision algorithms in industrial applications remains an escalating challenge to the reliable deployment of intelligent manufacturing, particularly given the complexity and diversity of traditional vision techniques. This manuscript proposes a novel visual explanation method based on the fusion of saliency maps to address the interpretability deficiency and enhance it accordingly. The proposed method generates robust visual explanations compatible with industrial applications by integrating weighted summation of output scores and data correlation analysis under encoded perturbations. Experiments on real-world industrial samples and the MVTec Industrial Anomaly Detection dataset demonstrate that the approach accurately highlights essential pattern regions and reflects the focus of algorithms, thereby improving the comprehension and trustworthiness of industrial machine vision applications. The proposed method contributes to the intuitive understanding of effectual machine vision mechanisms, mitigates potential misjudgment risks, and facilitates the scientific management of commercial vision integrations.