Transformer-based placement and harmonized composition for synthetic data generation in PCB component detection

Accurate component detection during the final assembly of complex integrated circuits (ICs) onto system-level or test Printed Circuit Boards (PCBs) is a critical semiconductor back-end process, essential for ensuring both test validation accuracy and overall product reliability. However, collecting sufficient annotated data for diverse component types and board layouts remains extremely costly and time-consuming, limiting the applicability of modern detectors. To address this challenge, we present a novel approach for synthetic data generation in very few-shot object detection, with a focus on PCB component detection. A transformer-based placement module predicts plausible insertion points for a given component patch on target boards, while a harmonized compositing stage seamlessly integrates it into the scene, yielding visually coherent and label-consistent training pairs. The synthesized pairs support effective fine-tuning of YOLO-style detectors under extreme data scarcity and enable rapid adaptation to newly introduced components with minimal real supervision. We evaluate across few-shot regimes and held-out boards and connector types, comparing against random/heuristic placement and standard copy-paste baselines. The results consistently indicate stronger realism, better label consistency, and improved downstream detection accuracy, highlighting that learned placement coupled with harmonized composition reduces labeling cost and accelerates adaptation in industrial inspection.