YOLOv12-DynaFocus: A Parameter-Budgeted Dynamic-Attention Detector for Thin-Defect-Aware PCB Inspection Under Compute and Latency Constraints
Abstract
Effective inline printed circuit board (PCB) automated optical inspection requires accurate, calibrated, latency-aware detection of low-contrast thin defects. Rather than proposing a new learning rule, this study develops YOLOv12-DynaFocus, a task-driven architectural co-design targeting approximately 9.14 M parameters, and evaluates forward-only and prediction-pipeline latency under separate A100 and Jetson protocols. Built on a compressed YOLOv12m backbone, it combines DF+ECA, TG-DF, and pre-head UGF to improve fixed-operating-point reliability for geometry-defined limited-spatial-support defects and regulate confidence. Because aggregate mAP can mask micro-defect failures, we report Thin-Recall and calibration metrics under the definition. Here, “thin-defect-aware” denotes this design and evaluation focus; PKU-PCB results are interpreted through recall and calibration, not an internal thin-versus-non-thin contrast. On the fixed PKU-PCB test split, A0(F) improves Thin-Recall from <inline-formula> <tex-math notation="LaTeX">$81.02~\pm ~1.75$ </tex-math></inline-formula>% to <inline-formula> <tex-math notation="LaTeX">$86.81~\pm ~1.20$ </tex-math></inline-formula>% (+ 5.79 pp). Under V1 validation, its mAP50 and mAP<inline-formula> <tex-math notation="LaTeX">${}_{\mathrm {50-95}}$ </tex-math></inline-formula> remain comparable to those of YOLOv12m. With approximately 9.14 M parameters, A0(F) uses 73% fewer parameters than the 34 M-parameter YOLOv12m baseline. Under the separate D1 FP16 batch-1 A100 prediction-pipeline protocol at <inline-formula> <tex-math notation="LaTeX">$640\times 640$ </tex-math></inline-formula>, it runs at <inline-formula> <tex-math notation="LaTeX">$6.39~\pm ~0.08$ </tex-math></inline-formula> ms per image (<inline-formula> <tex-math notation="LaTeX">$\approx 157$ </tex-math></inline-formula> FPS); the Jetson Orin NX benchmark characterizes edge-side throughput, memory footprint, and board-level power. Thus, A0(F) is interpreted as a parameter-count-efficient, Thin-Subset-oriented accuracy–complexity trade-off rather than a FLOP-minimal or latency-minimal detector at the equal-parameter YOLOv12s scale. Under severe PKU<inline-formula> <tex-math notation="LaTeX">$\to $ </tex-math></inline-formula>MIXED-clean domain shift, zero-shot mAP50 remains very low, motivating C3 class-wise-<inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-capped fine-tuning for line-specific use.