Printed circuit board (PCB) defect inspection is a key task in industrial automatic optical inspection (AOI). Its deployment in low-label scenarios remains difficult because PCB defects are often small, weakly contrasted, and unevenly distributed across categories and scales. Semi-supervised object detection can reduce annotation demand by using unlabelled images, but its performance depends strongly on the reliability of pseudo labels. In PCB inspection, confidence scores are shaped not only by defect category but also by object scale, which makes fixed or class-only filtering insufficient for small-defect supervision. This paper proposes dual-granularity pseudo-label calibration (DGPC) for semi-supervised PCB surface defect detection. DGPC first fuses weak-view and strongview teacher predictions into a unified candidate pool. It then estimates pseudo-label thresholds from both classlevel and class–scale statistics, and applies a scale-dependent lower-bound relaxation to retain informative smalldefect candidates. The method is applied only during training and leaves the detector architecture and inference procedure unchanged. Experiments on Deep PCB at 5%, 10%, and 20%labeled ratios show that DGPC improves pseudo-label-based training under low-label settings. The gains are most consistent when the teacher provides a sufficiently informative candidate pool, especially at 10% and 20% labelled ratios. Ablation, stability, crossdetector, pseudo-label statistics, scale-wise, qualitative, and sensitivity analyses show that class–scale-aware calibration improves pseudo-label selection for annotation-efficient PCB inspection.

Abstract
1. Introduction
2.Related Work
    2.1 Deep Learning-Based PCB Defect Detection
    2.2 Semi-Supervised Object Detection
    2.3 Pseudo-Label Calibration for Small-Scale Defect Detection
3.Methodology
    3.1 Overview of the Proposed Framework
    3.2 Teacher-Based Pseudo-Label Generation
    3.3 Cross-View Pseudo-Label Fusion
    3.4 Dual-Granularity Pseudo-Label Calibration
    3.5 Student Training
4.Experiments
    4.1 Lower-Bound Relaxation
    4.2 Main Results Under Different Labelled Ratio
    4.3 Stability Analysis Under the 10% Labelled Ratio
    4.4 Cross-Detector Generalization
    4.5 Ablation Study
    4.6 Pseudo-Label Statistics Analysis
    4.7 Scale-Wise Detection Analysis
    4.8 Qualitative Results
    4.9 Sensitivity Analysis of the Percentile Parameter
5.Conclusion
Data availability
References

• Jingkai Bi(School of Naval Architecture and Port Engineering, Shandong Jiaotong University, Weihai 264209, China)
• Mingzhi Chen(School of Naval Architecture and Port Engineering, Shandong Jiaotong University, Weihai 264209, China)
• Zhuoyi Yang(School of Naval Architecture and Port Engineering, Shandong Jiaotong University, Weihai 264209, China) Corresponding author
• Kexin Zhang(School of Naval Architecture and Port Engineering, Shandong Jiaotong University, Weihai 264209, China)