Comparison of deep learning-based image segmentation methods for the detection of voids in X-ray images of microelectronic components

This work applies two state-of-the-art approaches for semantic and instance segmentation of solder voids in X-ray images. Void segmentation is both: an important task in quality and failure analysis of microelectronic components and a challenge to modern computer vision methods, e.g. convolutional neural networks (CNN). We use a CNN named U-Net to distinguish void pixels from the background by semantic segmentation. For instance segmentation, we evaluate another CNN, namely Mask-RCNN, which allows the identification of distinct voids instead of a simple binary mask. This approach allows to identify, separate, and evaluate overlapping voids or even voids that lie on top of each other. For the examined dataset, the U-Net outperforms the Mask-RCNN. Nevertheless, the result suggests a trade-off: Once the dataset contains more than 20% of overlapping voids area, the Mask-RCNN becomes technically favorable.