Synthetic data generation for flank wear segmentation on cutting tool images by neural networks

R. Roth¹, S. Widmer¹, R. Ryser², A. Scholz², J. Sturm³, L. Meier⁴

The use of neural networks for qualitative or quantitative analysis of images is becoming more and more widespread in industry. Industrial applications such as quality control and defect detection are typical examples where solutions are available on the market. With an appropriate training of commercially available or open-source models, reliable results can be generated. The cost-effectiveness of the technology largely depends on the training effort compared to the quality and economic benefit of the working models. The use of synthetic generated data can be of great advantage for many applications, which may need frequent trainings to quickly adapt to new conditions. In this work the training of tool wear detecting segmentation models with synthetic images is systematically investigated. The influence of different parameter such as randomization of noise, blur, position, or the introduction of image postprocessing to equalize the synthetic images with real images are tested. Results show unexpected interaction among different parameters, which one by one may have different impact than combined. The best synthetic trained model reached a meanIoU of 80.3% while the presented benchmark trained with real data reached 93.4% meanIoU.
Keywords: neural networks, semantic segmentation, artificial intelligence, flank tool wear, synthetic data, computer vision, meanIoU

¹RhySearch, Forschung und Innovationzentrum Rheintal, Werdenbergstrasse 4, 9471 Buchs, Switzerland
²OST Ostschweizer Fachhochschule, Werdenbergstrasse 4, 9471 Buchs, Switzerland
³Manthano (Synthetic Future), Eichfeldstrasse 33, 8645 Jona, Switzerland
⁴Blaser Swisslube AG, Winterseistrasse 22, 3415 Hasle-Rüegsau, Switzerland