Zusammenfassung:
Porosity defects in metals can severely impact material properties and performance. Standard porosity detection techniques, such as dye penetrating testing, are time-consuming and often lack precision. In this study, we report on a deep-learning multi-class segmentation model developed to automatically identify porosities on laser-deposited bronze surfaces over steel, after lathe machining. Using samples with four distinct roughness levels, we constructed a dataset of 216 images (512 x 512 pixels each), labeled across three classes: background, uncovered porosity, and bronze-covered porosity. The model achieves F1-scores of 90% and 75% for the last classes, demonstrating its effectiveness in distinguishing porosity types. This model enables quantitative analysis of porosity geometry, revealing that, regardless of surface roughness, the average major and minor axes are approximately 0.2 mm, with a mean shape factor of 1.4 and an average area of 0.03 mm2. Notably, as roughness decreases, covered porosity levels rise non-linearly, likely due to material spreading during low-feed-rate facing operations.