Authors:
Christian Schorr
1
;
2
;
Sebastian Hocke
1
;
Tobias Masiak
3
and
Patrick Trampert
3
Affiliations:
1
German Research Centre Artificial Intelligence, Saarland Informatics Campus D 3 2, 66123 Saarbrücken, Germany
;
2
University of Applied Sciences Kaiserslautern, Amerikastraße 1, 66482 Zweibrücken, Germany
;
3
Artificial Intelligence Lab, ZF Friedrichshafen AG, Scheer Tower II, Uni-Campus Nord, D5 2, 66123 Saarbrücken, Germany
Keyword(s):
Synthetic Data, Digital Reality, Digital Twin, Defect Detection, Deep Learning, Quality Control, Computer Vision, Visual Inspection, Simulation, Rendering.
Abstract:
In recent years, the industry’s interest in improving its production efficiency with AI algorithms has grown rapidly. Especially advancement in computer vision seem promising for visual quality inspection. However. the proper classification or detection of defects in manufactured parts based on images or recordings requires large amounts of annotated data, ideally containing every possible occurring defect of the manufactured part. Since some defects only appear rarely in production, sufficient data collection takes a lot of time and may lead to a waste of resources. In this work we introduce a configurable, reusable, and scalable 3D rendering pipeline based on a digital reality concept for generating highly realistic annotated image data. We combine various modelling concepts and rendering techniques and evaluate their use and practicability for industrial purposes by customizing our pipeline for a real-world industrial use case. The generated synthetic data is used in different com
binations with real images to train a deep learning model for defect prediction. The results show that using synthetic data is a promising approach for AI-based automated quality control.
(More)