TUHH Open Research
Help
New user? Click here to register.Have you forgotten your password?
Communities & Collections
Publications
Research Data
People
Institutions
Projects
  1. Home
  2. Research Information
  3. Funding
  4. SynthDefekt: A Gen-AI-Assisted Pipeline for Generating Synthetic Data for Industrial Visual Inspection
Projekt Titel
SynthDefekt: A Gen-AI-Assisted Pipeline for Generating Synthetic Data for Industrial Visual Inspection
Förderkennzeichen
SCHU 2883/9-1
Funding code
945.03-1135
Startdatum
November 1, 2026
Enddatum
October 31, 2029
Gepris ID
585736575
Loading...
Thumbnail Image
Funder
Deutsche Forschungsgemeinschaft (DFG)
Funding Program
Sachbeihilfen
Institut
Flugzeug-Produktionstechnik M-23
Principal Investigator
Schüppstuhl, Thorsten
Supervised learning methods are the de facto standard in industrial machine vision, despite challenging boundary conditions (varying/uncontrolled illumination, moving objects, unknown viewpoints). Their performance, however, depends critically on training data that are of sufficient quality, adequately diverse, and correctly annotated. In visual inspection tasks, defect images are often not available in sufficient quantity and variety: rare cases are underrepresented. Machine vision systems must reliably detect these rare cases as well. Synthetic, computer-generated data are a promising alternative: rare defect classes can be generated in a targeted manner, annotations are available “by design,” application-relevant variations can be covered systematically, generation is scalable and IP-compliant, and it enables controlled class balancing at high variance. A common approach is image synthesis via computer graphics. In rule-based, 3D scene–based synthesis, however, extensive manual modeling effort is required (scene geometry/composition, material and texture models, defect models), and thus expertise spanning computer graphics, computer vision, and domain knowledge. In prior work, we were able to show that domain-specifically generated datasets enable specialist models to perform robustly in real-world visual inspection. SynthDefekt addresses the high manual effort through a semi-automated, coupled pipeline: procedural scene composition, object modification, defect generation, and material modeling are linked with post-processing (image enhancement, augmentation, domain adaptation), and the suitability of the generated dataset is quantified. The end-to-end process chain subsequently allows optimization. In an iterative workflow, besides rendering time, evaluation effort as well as time and compute resources are critical. The gold standard is to train ML models on the generated image data and validate them on real data in order to quantify suitability directly. To accelerate the evaluation step and enable more optimization cycles, we investigate to what extent suitability can be assessed indirectly with sufficient accuracy—for example by fusing or sequentially applying similarity metrics (FID, KID), domain-adaptation metrics (contrastive domain loss), semantic consistency measures (CLIP embedding distances), and proxy tasks. The development of this evaluation module constitutes the second project objective. In addition, GenAI assistance is being explored for manual sub-steps, such as pipeline configuration. Depending on technological maturity, diffusion models (e.g., for generating material models) or LLMs (e.g., for script synthesis for scene composition) are to be employed, among other methods.
TUHH
Weiterführende Links
  • Contact
  • Send Feedback
  • Cookie settings
  • Privacy policy
  • Impress
DSpace Software

Built with DSpace-CRIS software - Extension maintained and optimized by 4Science
Design by effective webwork GmbH

  • Deutsche NationalbibliothekDeutsche Nationalbibliothek
  • ORCiD Member OrganizationORCiD Member Organization
  • DataCiteDataCite
  • Re3DataRe3Data
  • OpenDOAROpenDOAR
  • OpenAireOpenAire
  • BASE Bielefeld Academic Search EngineBASE Bielefeld Academic Search Engine
Feedback