Synthetic Data for Defect Segmentation on Complex Metal Surfaces
Juraj Fulir, Lovro Bosnar, Hans Hagen, Petra Gospodnetić
Abstract
Metal defect segmentation poses a great challenge for automated inspection systems due to the complex light reflection from the surface and lack of training data. In this work we introduce a real and synthetic defect segmentation dataset pair for multi-view inspection of a metal clutch part to overcome data shortage. Model pre-training on our synthetic dataset was compared to similar inspection datasets in the literature. Two techniques are presented to increase model training efficiency and prediction coverage in darker areas of the image. Results were collected over three popular segmentation architectures to confirm superior effectiveness of synthetic data and unveil various challenges of multi-view inspection.