DSP seminar: September 11, 1997

  A Binary Markov Random Field Model for the Quantized Wavelet
  Coefficients of Images.

  by Sergio D. Servetto


  Zerotree based algorithms represent the state of the art in wavelet
  based image coding.  At a high level, these algorithms can be described
  as first sending some map of locations of zero coefficients (the set of
  zerotree symbols), and then sending the value of nonzero coefficients.
  However, the decision of what map to send is typically made using some
  simplifying assumption on the structure of the map, motivated by some
  empirically observed property of the data (e.g., that zero coefficients
  are likely to appear in tree structured sets): alternatively, we propose
  to estimate the map of locations of zero coefficients as a hidden
  binary Markov Random Field (MRF) instead.  

  After a brief discussion on wavelet based image coding algorithms, we
  will present a new coder built using the above mentioned MRF model.
  We will start by defining a data model for both the observed wavelet
  coefficients and for a hidden binary map, to then set up an
  optimization problem whose solution will yield an estimate of the
  hidden field.  Then, we will present an algorithm for the the actual
  computation of the hidden field.  And finally, we will present algorithms
  both for encoding the field, and for encoding the data assuming a known
  field estimate.  Simulation results on real data will be presented, which
  verify in practice the validity of our modeling assumptions.

  Our talk will conclude with a discussion of possible improvements to
  the proposed framework.


Joint work with Profs. Joseph M. Rosenblatt and Kannan Ramchandran.