Jansche A., Choudhary A.K., Bernthaler T., Schneider G.
Aalen University, DE
Keywords: convolutional neural networks, image inpainting, non-metallic inclusions, partial convolutions, quantitative microstructure analysis
Artefacts like stains or scratches introduced during the metallographic sample preparation process are a problem for manual as well as automatic image analysis. We show the use of image inpainting – normally used for reconstruction of missing parts of an image – for removal of these artefacts from micrographs. Light optical microscopy images of non-metallic inclusions in steels are used to demonstrate the results. We build on top of earlier work, in which the segmentation and classification of oxides, sulphides and artefacts was demonstrated. The masks produced from the segmentation model were re-used in the image inpainting step to mask out only regions which show artefacts. While the approach produced good results in terms of qualitative inspection / human perception the quantitative results suggested that the analysis was biased towards missing some inclusions as they were removed by the inpainting. We conclude by discussing steps to further improve the presented results.
Journal: TechConnect Briefs
Volume: TechConnect Briefs 2021
Published: October 18, 2021
Pages: 118 - 121
Industry sectors: Advanced Materials & Manufacturing | Medical & Biotech
Topics: Advanced Manufacturing, Materials Characterization & Imaging