Abstract Ball Grid Array (BGA) inspection is critical for ensuring the reliability of semiconductor assemblies. Conventional RGB imaging under a single illumination often yields high classification metrics with pretrained CNNs, yet explainability analyses reveal that models may attend to spurious artifacts rather than solder ball defects.This paper introduces a png4ch multi-illumination dataset by stacking grayscale captures under four illumination intensities into a multi-channel representation.A 1×1 convolutional projection maps the four channels into three, enabling compatibility with pretrained CNN backbones. Comparative experiments with imgShiny (RGB single-illumination) and png4ch show that while RGB yields more consistent high F1 scores, multi-illumination significantly improves attention alignment with defect regions (Grad-CAM), underscoring the importance of dataset design for trustworthy industrial deployment.
Campos et al. (Thu,) studied this question.