More than pretty images – Towards confidence bounds on segmentation thresholds

Peter Moonen (1, 2) , Pascale Sénéchal (1)

1 DMEX - Développement de méthodologies expérimentales
2 LFCR - Laboratoire des Fluides Complexes et leurs Réservoirs

We present an approach to assess the uncertainty associated to classifying voxels into different material phases. The approach consists in a spectral deconvolution of the grey-level histogram using a Gaussian mixture approach, followed by an iterative classification procedure based on the occurrence frequency. As phase attributions become increasingly more uncertain as iterations proceed, confidence bounds on the final segmentation are naturally obtained.

Mots clés: Gaussian mixture model; segmentation; classification; uncertainty

Domaines: Traitement du signal et de l'image [eess.SP]

Peter Moonen, Pascale Sénéchal. More than pretty images – Towards confidence bounds on segmentation thresholds. 4th International Conference on Tomography of Materials and Structures (ICTMS 2019), Jul 2019, Cairns, Australia. ⟨hal-02175825⟩