Scholarlink

<p>Surface defects on printed circuit boards degrade product quality and impact device performance, stability, and reliability. Accurately identifying these defects amidst a complex background is challenging due to the compact PCB layout. To address this, we propose the Multi-Scale Detection Transformer (MS-DETR), designed to enhance detection accuracy while maintaining efficiency for resource-constrained environments. MS-DETR introduces a Multi-Stage Convolution module to replace traditional convolution layers, enhancing feature extraction for small objects while reducing model complexity. The Slim-Scale Adaptive Fusion architecture, incorporating Ghost Squeeze-and-Excitation Convolution and Scale Adaptive Fusion modules, reduces network computation and enhances semantic information extraction. Additionally, integrating high-frequency and low-frequency (Hilo) information improves fine-grained feature extraction and focuses on critical features, while Inner Generalized Intersection over Union optimizes the loss function to accelerate convergence and enhance detection accuracy. Experimental results on a PCB defect dataset and a self-constructed dataset show that MS-DETR achieves a mean Average Precision of 96.90%, reducing arithmetic operations by 11.7 GFLOPs and parameter count by 5.9&#xa0;M compared to Real-time Detection Transformer, while maintaining a detection speed of 115 frames per second. These attributes make MS-DETR suitable for edge systems or embedded devices, balancing accuracy, computational cost, and speed effectively.</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

MS-DETR: a real-time multi-scale detection transformer for PCB defect detection

  • 期刊论文

  • 发表日期：

    2025年1月17日

查看全文
  • Li Ji,
  • Chaohang Huang,
  • Haiwei Li,
  • Wenjie Han,
  • Leiye Yi

摘要

Surface defects on printed circuit boards degrade product quality and impact device performance, stability, and reliability. Accurately identifying these defects amidst a complex background is challenging due to the compact PCB layout. To address this, we propose the Multi-Scale Detection Transformer (MS-DETR), designed to enhance detection accuracy while maintaining efficiency for resource-constrained environments. MS-DETR introduces a Multi-Stage Convolution module to replace traditional convolution layers, enhancing feature extraction for small objects while reducing model complexity. The Slim-Scale Adaptive Fusion architecture, incorporating Ghost Squeeze-and-Excitation Convolution and Scale Adaptive Fusion modules, reduces network computation and enhances semantic information extraction. Additionally, integrating high-frequency and low-frequency (Hilo) information improves fine-grained feature extraction and focuses on critical features, while Inner Generalized Intersection over Union optimizes the loss function to accelerate convergence and enhance detection accuracy. Experimental results on a PCB defect dataset and a self-constructed dataset show that MS-DETR achieves a mean Average Precision of 96.90%, reducing arithmetic operations by 11.7 GFLOPs and parameter count by 5.9 M compared to Real-time Detection Transformer, while maintaining a detection speed of 115 frames per second. These attributes make MS-DETR suitable for edge systems or embedded devices, balancing accuracy, computational cost, and speed effectively.