NAME

     bethe2d - Markov Random Field texture segmentation

SYNOPSIS

     bethe2d [ -N file_in ] [ -O file_out ] [ -lambda lambda_file
     ]  [ -coarse coarse ] [ -halfwin halfwin ] [ -par par ] [ -j
     j ] [ -tree_depth tree_depth ] [ [-st|-ch] ] [ [-s|-p]  ]  [
     -coh coh_file ] [ -? ]

DESCRIPTION

     bethe2d uses the bethe tree approximation of a Markov Random
     Field to perform texture segmentation.  For more information
     on the bethe tree approximation technique refer to  (Wu  and
     Doerschuk).  The estimator used is the maximizer of the pos-
     terior marginals (MPM), see (Marroquin, et al.).  The choice
     of  penalty functions can be tailored to the data set and to
     the textures of interest.  For large,  blocky  textures  the
     penalty  functions  TooSmall  and  TooThin  work  well.  The
     penalty       functions       ver_channel_follower       and
     hor_channel_follower are designed to look for channels.

     bethe2d gets all its parameters from command line arguments.
     These arguments specify the input, output, etc.

COMMAND LINE ARGUMENTS

     -N file_in
          The input file containing the disparity values  between
          all neighboring label sites.  This file is generated by
          the  program  disparity   or   the   program   combine.
          (Default:  stdin)

     -O file_out
          The  output  file  containing   the   segmented   data.
          (Default:  stdout)

GENERAL OPTIONS

     -fmat fifn
          the input file containing the forward  index  used  for
          locating  random  neighbors,  generated  by the program
          randomgraph

     -bmat bifn
          the input file containing the backward index  used  for
          locating  random  neighbors,  generated  by the program
          randomgraph

     -lambda lbdfn
          the input file containing the lambda schedule.  The  i-
          th  lambda value is the weight on the penalty functions
          during the i-th iteration (see Wu & Doerschuk)


     [-s | -p]
          the type of update scheme. Serial updating is used when
          the  -s flag is set. Parallel updating is used when the
          -p flag is set (see Wu & Doerschuk).  Only one of these
          flags  can be specified on the command line.  (Default:
          -s)

     -coarse coarseness
          inverse of the resolution at which the segmentation  is
          performed.   That  is,  the  ratio  of  the  number  of
          samples/traces in the data lattice  to  the  number  of
          samples/traces  in the label lattice.  The data lattice
          is the input to the program texture, and the label lat-
          tice  is  the  output  from any segmentation optimizer,
          such  as  bethe2d.   For  highest  resolution  set  the
          coarseness to 1.  (Default: 9)

     [-st | -ch]
          the type of penalty function to use.  If the  -st  flag
          is  set  the penalty functions TooSmall and TooThin are
          used.  If the -ch flag is  set  the  penalty  functions
          ver_channel_follower and hor_channel_follower are used.
          Only one of these flags can be specified on the command
          line.  (Default:  -st)

     -halfwin half_win_size
          half window width used to determine the  width  of  the
          texture  analysis  window over which the cdf's are com-
          puted in the program disparity.  The width of the  tex-
          ture analysis window is given by 2 * half_win_size + 1.
          (Default: 7)

     -par num_par
          the number of partitions to segment the data set  into.
          After  segmentation,  labels with the same value belong
          to a partition.  Each partition represents  a  specific
          texture class.  (Default:  2)

     -j j the weight on the cost function (see Wu  &  Doerschuk).
          (Default:  1)

     -tree_depth td
          the depth at which  the  bethe  approximation  is  ter-
          minated (see Wu & Doerschuk).  (Default:  1)

     -coh coh_file
          when this option is specified on the command  line  the
          coherence  measures  in  the  file coh_file are used as
          weights on the near neighbor terms of  the  cost  func-
          tion,  i.e.  the coherence between two sites, which are
          near neighbors,  determines  the  contribution  of  the
          corresponding  near neighbor term in the cost function.
          The file coh_file is produced by  the  program  cogeom.
          When  this  option is not specified on the command line
          weighting is not applied to the near neighbor terms.

     -h, -?, -help
          Help.

See Also

     texture2d, cogeom2d, combine2d, disparity2d,  randomgraph2d,
     texstat2d, brightsizing

REFERENCES

     Crawford,  Kelly,  and  Marfurt,  Kurt,  1997.   2D  Texture
     Analysis:   A  User's  Guide, Amoco Geoscience Research Bul-
     letin F97-G-14.

     Hoelting, Cory and Kelly, Ken, 1996.   Texture  Analysis  of
     Spectral  Decomposition Data Using a Segmentation Algorithm,
     Amoco Geoscience Research Bulletin F96-G-21.

     Matheney, Mike and Kelly, Ken, 1995.  Texture-Based  Segmen-
     tation  of  3-D Seismic Data, Amoco Geoscience Research Bul-
     letin F95-G-29.

     Wu, C. and Doerschuk, P., 'Texture-Based Segmentation  Using
     Markov  Random Field Models and Approximate Bayesian Estima-
     tors Based on Trees,' Jour. of  Math.  Imaging  and  Vision,
     vol. 5, (1995), pp. 277-286.

     Marroquin, J., Mitter, S., and  Poggio,  T.,  'Probabilistic
     Solution  of  Ill-posed  Problems in Computer Vision,' Jour.
     of the American Statistical Assoc., March 1987, vol. 82, no.
     397, pp. 76-89.

CONTRACT AGREEMENT

     This product is brought to you by  Research  Agreement  #548
     (The Seismic Coherency Cube). Thank you for your support!

AUTHOR

     Cory Hoelting and Ken Kelly (E&PTG, Tulsa, OK, USA).

COPYRIGHT

     copyright 2001, Amoco Production Company
               All Rights Reserved
          an affiliate of BP America Inc.