Based on  Trainable Weka Segmentation and the MorphoLibJ library, we developed an ad-hoc workflow to
  • segment multi-phase 3D X-ray Computed Tomography images
  • separate touching objects
  • characterize the morphological features of the objects.
​The novelty here is the combination of morphological operations and machine learning algorithms to extract image features for segmenting multiphase images. Note that image segmentation becomes extremely difficult when defects are present and adjacent to particle-void boundaries. Because of the similar attenuation coefficients of voids and defects (and the partial-volume effect, the particles encompassing defects tend to be identified as "porous" objects and thus overly segmented into small fragments, by the watershed algorithm. The machine learning based classifier utilizes the highlighted morphological features, such as plate- and blob-like structures, as training data. Once trained, the classifier is able to handle the rest of the similar images.