Scalable Learning: Probabilistic Graphical Models for Image Analysis

Fig. 1: A 3D data volume from the microCT beamline at the ALS (left) undergoes image segmentation (right) as part of an analysis pipeline.

Scientific Achievement

Leveraging advances in mathematics and CS, a new method for graph partitioning scales image segmentation to run on large DOE HPC platforms, and with high accuracy [1].

Significance and Impact

This new method enables accelerating image analysis time-to-solution for DOE science projects challenged by an ever-growing data tsunami.

Research Results

  • The method exploits cliques from a Markov Random Field (MRF) formulation, and its solution yields a highly accurate N-label image segmentation
  • Reduced computational complexity, suitable for use on large problem sizes
  • Parallel parameter learning for MRF uses LAP strategy for graph partitioning
  • Shared- and distributed-memory implementations run on DOE HPC platforms



  1. C. Heinemann, T. Perciano, D. Ushizima, and E. W. Bethel, “Distributed Memory Parallel Markov Random Fields Using Graph Partitioning,” in 2017 IEEE International Conference on Big Data (Big Data), pp. 3332–3341, Dec. 2017.