Reconfigurable manufacturing systems are becoming the only viable option to respond to changing product volumes and product specification, which are currently major challenges for the manufacturing industry. Part of this adaptation requires vision systems to be quickly updated to handle new unseen products. For deep learning-based vision systems, this means re-training on images that might not be available. Although there is some existing work on synthetic image generation in manufacturing contexts using a variety of domain randomisation techniques, there is a lack of understanding of which domains are critical in the effectiveness of the resulting trained model. There are currently no open tools to systematically conduct such ablation studies. This paper presents a tool based on Blender and CAD models to enable the study of domain randomisation in the generation of synthetic-only datasets that can yield accurate object recognition models. Preliminary results to validate the implemented domain randomisation techniques and the ability to generate the synthetic images are presented. Once generated, synthetic data sets are used to train a YOLOv8 model for object detection as a second tool validation step. Future work will look at performing ablation studies and expanding the range of domain randomisation methods to further study the capabilities of synthetic images.
A Tool for Generating and Labelling Domain Randomised Synthetic Images for Object Recognition in Manufacturing
Lect.Notes Mechanical Engineering
European Symposium on Artificial Intelligence in Manufacturing ; 2024 ; Athens, Greece October 16, 2024 - October 16, 2024
22.03.2025
8 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Hybrid object labelling in digital images
British Library Online Contents | 2007
|Online Contents | 2003
Online Contents | 2002
Object Recognition by Combining Paraperspective Images
British Library Online Contents | 1996
|