Automated image segmentation of worn PVD-coated cutting edges using multimodal convolutional neural network-based approaches

Le résumé fourni par la source

Automated wear analysis of coated milling cutters based on multimodal microscopy data combining scanning electron microscopy (SEM) and energy-dispersive X-ray spectroscopy (EDS) remains challenging due to non-linear correlations between structural and compositional features. In this study, four distinct deep learning and clustering approaches were developed to address these challenges. A fully supervised convolutional neural network for baseline segmentation, a weakly supervised model with iterative pseudo-label refinement, a hybrid approach integrating separate supervised and unsupervised segmentations through density-based fusion, and a hierarchical method refining local tool regions via targeted clustering. The comparative evaluation revealed substantial differences in computational effort and segmentation behavior. The purely supervised model achieved rapid training within approximately one hour but was limited in adaptability to out-of-distribution wear patterns. The weakly supervised approach required the longest runtime yet demonstrated that convolutional networks can autonomously refine coarse labels into detailed segmentations by exploiting intermediate feature representations. Hybrid methods combining CNNs with clustering achieved intermediate runtimes (∼14–15 hours) while producing the most fine‑grained boundaries. Overall, increasing model complexity led to higher computational cost but also improved sensitivity toward subtle wear zones and cross‑modal correlations.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.