Scholarlink

<p>With the escalation of production efficiency in the automotive industry, conventional models for defect detection in industrial production lines are confronted with challenges, such as high computational resource consumption and low detection efficiency. This is particularly pronounced in the fast-paced and complex environment of automobile body stamping parts workshops, necessitating real-time defect detection. We introduce a lightweight end-to-end detection model tailored for automobile body stamping parts. Our model leverages EfficientViT, incorporating sandwich layout blocks and cascaded group attention to curtail computational costs while enhancing complex image processing capabilities. We further integrate GSConv for lightweight convolution and devise a Programmable Aggregate Information Module (PAI-CSP) to optimize feature map fusion across stages. A lightweight dynamic feature fusion encoder, combining GSConv, PAI-CSP, and DySample, reduces computational delay and optimizes model processing efficiency. Our method’s efficacy is substantiated through experiments on a dedicated automotive body stamping part defects dataset, demonstrating a 20.6 increase in FPS compared to the baseline model and a 61.83% reduction in GFLOPs while meeting factory detection precision requirements. These results underscore the model’s potential for real-time defect detection in automotive body stamping parts.</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

Multiscale deformable transformer meets lightweight design: enabling real-time quality detection for automobile body stamping parts under production-line constraints

  • 期刊论文

  • 发表日期：

    2025年6月6日

查看全文
  • Siyi Ding,
  • Zihao Wang,
  • Yefan Yang,
  • Wei Dai,
  • Jie Luo

摘要

With the escalation of production efficiency in the automotive industry, conventional models for defect detection in industrial production lines are confronted with challenges, such as high computational resource consumption and low detection efficiency. This is particularly pronounced in the fast-paced and complex environment of automobile body stamping parts workshops, necessitating real-time defect detection. We introduce a lightweight end-to-end detection model tailored for automobile body stamping parts. Our model leverages EfficientViT, incorporating sandwich layout blocks and cascaded group attention to curtail computational costs while enhancing complex image processing capabilities. We further integrate GSConv for lightweight convolution and devise a Programmable Aggregate Information Module (PAI-CSP) to optimize feature map fusion across stages. A lightweight dynamic feature fusion encoder, combining GSConv, PAI-CSP, and DySample, reduces computational delay and optimizes model processing efficiency. Our method’s efficacy is substantiated through experiments on a dedicated automotive body stamping part defects dataset, demonstrating a 20.6 increase in FPS compared to the baseline model and a 61.83% reduction in GFLOPs while meeting factory detection precision requirements. These results underscore the model’s potential for real-time defect detection in automotive body stamping parts.