ICMarkingSynthetic: A Synthetic Image Dataset of Defective Integrated Circuit Markings Generated by Physics-Informed Guidelines

Inspection of markings on integrated circuit (IC) packages requires a large dataset of images of defective markings. Such images are scarce and seldom shared publicly. This paper introduces a set of physics-informed guidelines for generating synthetic images of defective IC markings from images of defect-free markings. These guidelines build each defect type from its physical cause, distinguish defects of the marking from defects of the package surface, and produce defect labels. Four common defect types of scratch, contamination, missing text and smeared text are implemented, together with images that combine two or three defect types. In addition, the steps for adding further defect types are provided. The guidelines are then evaluated on a small dataset of real images of marking defects. It is shown that the generated images are close to the real images of defective markings in terms of four quantitative similarity measures. It is also shown that a detector trained on the generated images alone detects most of the real defects. In essence, this paper provides a dataset of 6,000 synthetic IC marking images for six image classes, named ICMarkingSynthetic, to enable IC marking inspection studies, as well as the guidelines for generating these and other defect types.