Collecting sufficient image data is one of the limiting factors in visual quality inspection with artificial intelligence (AI). For each defect type, the manufacturing company must provide at least 30 sample images so that an inspection can be performed with high accuracy. Especially in view of the ongoing individualization of the production with many thousands of product variants, it is often uneconomical to collect enough examples to cover all possible eventualities. One way to counteract this problem is learning from photorealistic synthetic data, which has been increasingly investigated in AI research in recent years. This involves using computers to automatically generate large sets of artificial images based on only a small amount of input data.
However, current approaches to using synthetic data involve a lot of manual work (especially graphics design) to adapt to individual use cases and are often limited in their variability to the position and orientation of objects, as well as the type of background. For application in visual quality inspection, this approach needs to be extended to work at the sub-object and material level as well, making the technology applicable to new products in an automated way.