This is the homepage for the paper

"Statistical Region Merging"

by R. Nock and F. Nielsen, appearing in IEEE Transactions on Pattern Analysis and Machine Intelligence, 2004.
[ Draft ] [ Source C ] [ Linux binaries (old) ] [ Images and experiments ] [ BSDB Experiments ] [ Windows material ] [ Feedback ]

 

Draft

The final paper is available here.pdf gzipped file in the IEEE draft format.


 

Source C

The zipped source C is available: click hereSource C (<100 Ko) .



 

Linux binaries (old)

(these binaries are a bit old now; you may directly use the Source code above)

The gzipped binary (Linux) is available: click hereLinux binaries (40 Ko) .
Unzip the file (gzip -d srm.gz), and then enter a command line.

A slight help is available by entering
Below is this help you get:
./srm: option requires an argument -- h
Usage :
./srm [option]* -s input.tiff output.tiff
 Welcome on srm (Statistical Region Merging), By R. Nock and F. Nielsen
 Important notice on this software:
    * it is provided AS IS, without ANY guarantee
    * it is provided for RESEARCH PURPOSE ONLY

 Options:
  h : displays this help
  S : in {0,1,2,3}, type of output (0=bordered regions with original pixels,
                                    1=black-bordered regions with white pixels,
                                    2=bordered regions with average-color pixels,
                                    3=randomized colored regions)
  B : 0-2, border color
      0=black (valid for S=0,2)
      1=white (valid for S=0,2)
      2=no border (valid for S=0,2)
  P : 0-1, saves (when P=1) the (at most) 15 largest regions of the segmentation
           in files named [output.tiff]_RegionX, for X=0, 1, ...
  T : in {0,1,2,3,4}, type of noise (0=no noise,
                                     1=transmission noise,
                                     2=salt and pepper noise,
                                     3=Gaussian (Cf Xu et al., PRL (19) pp 1213--1224, 1998),
                                     4=artificial occlusions (s random occlusions) )
  N : 0-100, % of pixels having noise (transmission, salt and pepper),
      >0   : sigma (Gaussian noise)
      >0   : s (random occlusions)
  U : 0-1, type of neighborhood for sorting couples:
      0  : original (Delta=0)
      1  : for noise handling (Delta=10)
  G : 0-1, convolution filter for sorting couples:
      0  : no
      1  : Sobel-type of arbitrary size
  F : >1, parameter for the size of the convolution filter when G=1 (actual size=2*F+1):
  O : 0-1, occlusions merging (0 is No, 1 is Yes)
  Q : >=1, complexity parameter of the statistical model (Cf paper)

 Default settings: S=B=P=T=N=U=G=O=0, Q=32

 Notices:
   * noisified images are also saved; the Filename is [output.tiff]-original-noisified
   * when P=1 and images are noisified, the regions saved are those of the -original image-



 

Images and experiments

The following table displays a sample of SRM's results. Please note the following:


Househouse (Fig. 9)

Command lines :
./srm -S 0 -B 1 -G 1 -F 2 -Q 1 -s house.tif house-result-1.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 2 -s house.tif house-result-2.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 4 -s house.tif house-result-4.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 8 -s house.tif house-result-8.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 16 -s house.tif house-result-16.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 32 -s house.tif house-result-32.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 64 -s house.tif house-result-64.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 128 -s house.tif house-result-128.tif
./srm -S 0 -B 1 -G 1 -F 2 -Q 256 -s house.tif house-result-256.tif

Original image
-Q 1 -Q 2 -Q 4 -Q 8 -Q 16 -Q 32 -Q 64 -Q 128 -Q 256




 

BSDB experiments

Additional experiments of SRM (and other image processing algorithms as well) are available on The Berkeley Segmentation Dataset and Benchmark. A webpage is devoted to this repository, and you will find it hereBSDB exps..



 

Windows material

To get the Windows version of SRM, please click hereF.N. homepage.



 

Feedback

Your feedback is much appreciated. Your can reach us at rnock@martinique.univ-ag.fr, or at Nielsen@csl.sony.co.jp.