Scholarlink

<p>In this paper, we propose a method for detecting door-closing defects using physics-based simulations. Quantitative inspection of industrial products is essential to suppress human error and variability in evaluation. Door-closing inspections relying on human sensory evaluations are one of the prime targets for quantification and automation. Developing deep learning-based visual inspection models typically requires time-consuming and labor-intensive data collection using specialized measurement equipment. To eliminate this need for data collection, the proposed method uses synthetic data generated with physics-based simulations that capture the physical dynamics of door-closing. A multi-task learning framework is designed to simultaneously perform binary classification (distinguishing normal from defective doors) and regression (estimating the closing energy), with shared parameters. This method enables the model to acquire detailed physical knowledge in advance, thereby achieving high inspection accuracy with only a small amount of real-world image data, and without dependence on specialized measurement systems. Experimental results demonstrate that our method outperforms existing approaches and the method that uses ground-truth door-closing energy data collected with specialized measuring devices.</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

Detecting Door-Closing Defects Via Pre-Training with Physics-Based Simulations

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月16日

查看全文
  • Yota Yamamoto,
  • Ryoga Takahashi,
  • Ryosuke Furuta,
  • Yukinobu Taniguchi

摘要

In this paper, we propose a method for detecting door-closing defects using physics-based simulations. Quantitative inspection of industrial products is essential to suppress human error and variability in evaluation. Door-closing inspections relying on human sensory evaluations are one of the prime targets for quantification and automation. Developing deep learning-based visual inspection models typically requires time-consuming and labor-intensive data collection using specialized measurement equipment. To eliminate this need for data collection, the proposed method uses synthetic data generated with physics-based simulations that capture the physical dynamics of door-closing. A multi-task learning framework is designed to simultaneously perform binary classification (distinguishing normal from defective doors) and regression (estimating the closing energy), with shared parameters. This method enables the model to acquire detailed physical knowledge in advance, thereby achieving high inspection accuracy with only a small amount of real-world image data, and without dependence on specialized measurement systems. Experimental results demonstrate that our method outperforms existing approaches and the method that uses ground-truth door-closing energy data collected with specialized measuring devices.