Scholarlink

<p>Accurate component detection is a fundamental yet challenging requirement for automated defect inspection in Electric Multiple Unit (EMU) bogies. Existing general object detection models often suffer from insufficient precision due to unavoidable positional misalignment between inspection robot images and predefined standard templates. To address this, we propose a novel dual-branch detection network that explicitly fuses tensor-form prior structural knowledge with visual features. Our method constructs a multi-channel semantic tensor from standard images, encoding Gaussian-based spatial priors: anisotropic Gaussian ellipses for fixed components and isotropic Gaussian circles distributed along conic motion trajectories for rotating/moving parts. This prior is injected via a dedicated input branch and aligned with visual features using a deformable convolution (DCN) module to correct positional variance. Evaluated on a custom dataset of 499 real-world bogie images, our approach, when integrated with a Faster R-CNN FPN baseline, achieves 84.01% mAP@0.5 and 67.00% mAP@0.75, representing a +16.6% point gain in mAP@0.75 over the baseline. The model also attains high precision (0.82) and recall (0.78), demonstrating its effectiveness and robustness for industrial inspection tasks.</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

High-speed train bogie component detection network with dual-branch architecture based on tensor fusion of prior structural knowledge

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Gang Peng,
  • Chaoze Wang,
  • Chaowei Song,
  • Mingjun Cong,
  • Sheng Zhong

摘要

Accurate component detection is a fundamental yet challenging requirement for automated defect inspection in Electric Multiple Unit (EMU) bogies. Existing general object detection models often suffer from insufficient precision due to unavoidable positional misalignment between inspection robot images and predefined standard templates. To address this, we propose a novel dual-branch detection network that explicitly fuses tensor-form prior structural knowledge with visual features. Our method constructs a multi-channel semantic tensor from standard images, encoding Gaussian-based spatial priors: anisotropic Gaussian ellipses for fixed components and isotropic Gaussian circles distributed along conic motion trajectories for rotating/moving parts. This prior is injected via a dedicated input branch and aligned with visual features using a deformable convolution (DCN) module to correct positional variance. Evaluated on a custom dataset of 499 real-world bogie images, our approach, when integrated with a Faster R-CNN FPN baseline, achieves 84.01% mAP@0.5 and 67.00% mAP@0.75, representing a +16.6% point gain in mAP@0.75 over the baseline. The model also attains high precision (0.82) and recall (0.78), demonstrating its effectiveness and robustness for industrial inspection tasks.