The increasing miniaturization of printed circuit boards (PCBs) poses significant challenges to vision-based sensing and measurement systems for automated optical inspection (AOI), including extreme multi-scale defect variations, intrinsically low signal-to-noise ratios (SNR) for subtle anomalies, and rapidly escalating computational costs at higher inspection resolutions. To address these challenges, we propose a lightweight RT-DETR-R18 framework that systematically enhances both feature extraction and cross-scale measurement fusion. Specifically, we introduce an Adaptive Multi-Scale Large-Kernel Dilated Convolution with shallow feature injection (AMDC-FI) to expand the receptive field while preserving high-frequency spatial details essential for accurate boundary measurement, and an improved Cross-Scale Continuous Feature Fusion (CCFF) neck with Parametric Wavelet Downsampling (PWD) and Adaptive Weighted Feature Fusion (AWFF) for frequency-preserving multi-scale aggregation. On a six-class PCB defect dataset, our method achieves 51.58\% mAP@0.5:0.95, 90.67\% mAP@0.5, and 36.03\% small-object AP with only 15.82M parameters and 25.20G FLOPs---outperforming the RT-DETR-R18 baseline by 8.58\%, 7.04\%, and 12.64\%, respectively, while reducing parameters and FLOPs by 21.3\% and 14.9\%. These results confirm that combining enlarged receptive fields with frequency-preserving downsampling effectively enhances measurement accuracy and computational efficiency, offering a practical solution for real-time PCB analytics in intelligent manufacturing environments.

Vision-based sensing,PCB measurement,Multi-scale feature fusion,RT-DETR,Nondestructive evaluation