SDP-RTDETR: a real-time end-to-end detector for small pcb surface defects
Machine Vision and Applications · 2026-11 · 37 (6)
DOI: 10.1007/s00138-026-01922-w ↗Kurz gefasst
Fachartikel in der Zeitschrift Machine Vision and Applications, online erschienen am 11.09.2026. Verfasst von Miao Yi und 3 weiteren Autor:innen. Die Arbeit nennt 57 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Natural Science Foundation of Sichuan Province.
Abstract (Originalsprache)
Automated optical inspection of printed circuit board (PCB) surface defects remains challenging because defects are often small, slender, low-contrast, and visually similar to dense traces, pad arrays, and other repetitive background patterns, leading to missed detections, false alarms, and inaccurate localization. To address these issues, a real-time end-to-end detector, termed Small Defect Perception Real-Time Detection Transformer (SDP-RTDETR), is developed based on RT-DETR through three coordinated designs. First, a lightweight backbone, named Lightweight HGstem–RGCSPELAN backbone (LHG-RGC), is designed to suppress redundant texture responses while preserving shallow edge and local geometric details. Second, a Small Defect Perception Path Aggregation Network (SDP-PAN) is developed to enhance cross-scale propagation of defect-sensitive features through selective bidirectional fusion and shallow-detail enhancement. Third, the existing minimum point distance intersection over union (MPDIoU) and Gaussian Combined Distance (GCD) metrics are combined to construct a task-oriented regression loss for improving localization stability under low-overlap or non-overlapping conditions. Experimental results on a PCB defect dataset show that SDP-RTDETR achieves 98.7% precision, 99.2% recall, 99.3% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mA{P}_{50}$$\end{document}, 85.5% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mA{P}_{75}$$\end{document}, and 69.7% \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mA{P}_{50:95}$$\end{document}. Compared with the baseline RT-DETR, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mA{P}_{75}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:mA{P}_{50:95}$$\end{document} increase by 8.1% and 4.6% points, while the parameter count and computational cost decrease from 19.9 M to 57.0 GFLOPs to 13.6 M and 48.7 GFLOPs, respectively. The proposed model achieves an inference speed of 200.1 FPS on an NVIDIA GeForce RTX 4090D. Evaluations on additional industrial defect datasets further show that SDP-RTDETR consistently outperforms the RT-DETR baseline across different defect-detection scenarios.
Förderung
- National Natural Science Foundation of China
- Natural Science Foundation of Sichuan Province
Zitieren
APA
Yi, M., Xia, C., Lin, Q., & An, W. (2026). SDP-RTDETR: a real-time end-to-end detector for small pcb surface defects. Machine Vision and Applications, 37(6). https://doi.org/10.1007/s00138-026-01922-w
Vancouver
Yi M, Xia C, Lin Q, An W. SDP-RTDETR: a real-time end-to-end detector for small pcb surface defects. Machine Vision and Applications. 2026;37(6). doi:10.1007/s00138-026-01922-w
BibTeX
@article{Yi2026,
author = {Yi, Miao and Xia, Changjiu and Lin, Qiang and An, Weisheng},
title = {SDP-RTDETR: a real-time end-to-end detector for small pcb surface defects},
journal = {Machine Vision and Applications},
year = {2026},
volume = {37},
number = {6},
doi = {10.1007/s00138-026-01922-w}
}
EUVIMED