Scholarlink

<p>Sparse-view computed tomography (CT) can reduce acquisition times, supporting inline industrial inspection in suitable settings. In practice, scan time may also be shortened by lowering exposure per view, using faster detectors or motion systems, or leveraging partial/parallel acquisition; here we focus on reducing the number of projections. Fewer projections, however, can introduce streak artifacts that cause measurement deviations during metrological evaluations. This paper presents two self-supervised deep learning approaches using implicit neural representations (INR) to mitigate sparse-view artifacts and enhance measurement accuracy. Both methods represent the 3D object volume using a multi-layer perceptron (MLP) optimized individually for each scan through an incremental forward-backward strategy. The first approach, Neural Representation with Sparse-View Volume-based Loss (NR-SVOL), employs volume-domain training using an initial filtered back-projection (FBP) volume, enabling rapid artifact reduction with limited computational overhead. The second, Neural Representation with Sparse-View Projection-based Loss (NR-SPRO), directly optimizes the INR to match measured sparse projections, analogous to Neural Radiance Fields (NeRF), yielding superior artifact compensation at the expense of increased computation. Comprehensive evaluations were conducted on three industrial objects, a gear, a cylinder head, and a connector, at varying sparse-view configurations (32–256 projections). Both NR-SVOL and NR-SPRO demonstrated substantial artifact reduction, decreasing surface deviations by up to an order of magnitude in standard deviation. NR-SVOL achieved results within approximately five minutes, suggesting compatibility with some inline cycle times for our tested parts, while NR-SPRO delivered even higher accuracy when allowed more computation. This study highlights a practical trade-off between speed and precision, showcasing the potential of these methods for sparse-view inline industrial CT for improved metrological quality.</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

Compensating Streak Artifacts in Sparse-View Inline Industrial CT for Accurate Metrology using Self-Supervised Optimization of Implicit Neural Volume Representations

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月28日

查看全文
  • Faizan Ahmad,
  • Guangpu Yang,
  • Manuel Buchfink,
  • Ammar Alsaffar,
  • Ahmed Baraka,
  • Xingyu Liu,
  • Sven Simon

摘要

Sparse-view computed tomography (CT) can reduce acquisition times, supporting inline industrial inspection in suitable settings. In practice, scan time may also be shortened by lowering exposure per view, using faster detectors or motion systems, or leveraging partial/parallel acquisition; here we focus on reducing the number of projections. Fewer projections, however, can introduce streak artifacts that cause measurement deviations during metrological evaluations. This paper presents two self-supervised deep learning approaches using implicit neural representations (INR) to mitigate sparse-view artifacts and enhance measurement accuracy. Both methods represent the 3D object volume using a multi-layer perceptron (MLP) optimized individually for each scan through an incremental forward-backward strategy. The first approach, Neural Representation with Sparse-View Volume-based Loss (NR-SVOL), employs volume-domain training using an initial filtered back-projection (FBP) volume, enabling rapid artifact reduction with limited computational overhead. The second, Neural Representation with Sparse-View Projection-based Loss (NR-SPRO), directly optimizes the INR to match measured sparse projections, analogous to Neural Radiance Fields (NeRF), yielding superior artifact compensation at the expense of increased computation. Comprehensive evaluations were conducted on three industrial objects, a gear, a cylinder head, and a connector, at varying sparse-view configurations (32–256 projections). Both NR-SVOL and NR-SPRO demonstrated substantial artifact reduction, decreasing surface deviations by up to an order of magnitude in standard deviation. NR-SVOL achieved results within approximately five minutes, suggesting compatibility with some inline cycle times for our tested parts, while NR-SPRO delivered even higher accuracy when allowed more computation. This study highlights a practical trade-off between speed and precision, showcasing the potential of these methods for sparse-view inline industrial CT for improved metrological quality.