Synthetic Data-Driven Deep Learning for Industrial Surface Inspection: Advancing Beyond Classical Computer Vision
IEEE International Conference on Emerging Technologies and Factory Automation (ETFA), pp. 1–8
Abstract
Industrial visual inspection tasks with rare defects often suffer from severe data scarcity, forcing many applications to rely on analytical image-processing pipelines instead of deep learning. While such methods require only a few reference samples, they are typically sensitive to illumination changes and parameter tuning. This paper investigates whether deep learning can become feasible under extreme low-data conditions using exclusively synthetic training data, while relying on only a handful of real images for validation and evaluation. A synthetic-data-driven pipeline is developed for detecting transverse scratches on reflective metallic components and evaluated on real inspection images against an analytical benchmark. Results indicate improved tolerance to varying acquisition conditions and more complete defect detection under similarly constrained data conditions. The findings indicate that targeted synthetic data can enable deep-learning-based inspection in scenarios traditionally dominated by handcrafted vision systems.
Authors 6
-
Josefine Monnet Aachen
Affiliation as printed
RWTH Aachen University
-
Ludwig Tonn Aachen
Affiliation as printed
RWTH Aachen University
-
Elen Ayanyan Aachen
Affiliation as printed
RWTH Aachen University
-
Oliver Petrović Aachen
Affiliation as printed
RWTH Aachen University
-
Werner Herfs Aachen
Affiliation as printed
RWTH Aachen University
-
Christian Brecher Aachen
Affiliation as printed
RWTH Aachen University