Generation of Synthetic Data for DL-based Defect Detection in the Automotive Context

In the automotive context, the quality of products or assemblies is inspected regularly at quality gates, constituting a laborious and error-prone process if performed manually. Machine Vision (MV) systems can partially automate the inspection processes, relieving workers from such tasks and thus increasing the overall productivity. Although MV systems offer high potential in theory, their performance depends on large volumes of annotated data specific to the considered use cases. Yet, acquiring and annotating such data in industrial settings poses significant challenges due to accessibility constraints, confidentiality issues, high manual effort, and costs. As an alternative to the acquisition of real-world data, this work investigates the utilization of synthetic data for the training of deep learning (DL) models for defect detection, providing high-quality annotations by design and reducing manual effort significantly. The authors present a methodology for image generation and demonstrate that a DL model trained on synthetic data achieves comparable accuracy in object detection to one trained on real-world data. Moreover, the presented approach offers the capability to generate explicit error cases, providing valuable insights for error mitigation strategies in the context of automotive production and assembly processes.