Answers are grounded in this paper and cite their sources.
Printed circuit board inspection combines geometric alignment, displacement measurement and final localization. Using eight Gerber designs and photograph–layout pairs, we show that these outputs favor different registration systems. At the designated synthetic condition, dense inverse search achieves a 0.307-pixel base-map error and a 3.979-pixel residual-difference root mean square error, while shared centroid correction yields similar final positions across the shared-initializer backends. Paired-vector accounting resolves initialization drift, refinement response and directional/orthogonal components, explaining how drift cancellation and displacement absorption coexist. Target-excluded readouts expose an accuracy–support trade-off. Image screening links error rejection to group-dependent coverage, and recovered coordinate references support evaluation on 36/60 original-image pairs. The resulting evaluation protocol jointly reports geometric accuracy, displacement retention, final localization, measurable support and failure coverage.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: