TEXTURE IMAGES FOR TESTS

Texture analysis research has been suffering from the lack of common test images
on which the algorithm performances are illustrated. Here we attempt to decrease
this difficulty of comparison by providing a set of test pictures. The idea is
to have few but sufficiently difficult real textures with known
texture boundaries as test images. Automatic segmentation algorithms can then
provide their results on these images which can be compared
with published results (on the same images) without that each author needs to
implement other authors algorithms (and run them).

The texture patches are made by Josef Bigun and  J. M. Hans du Buf, 1988-1992,
when working at the Swiss Federal Institute of Technology, Lausanne. In the particular
choice of the texture patches 3 independent human experts were constulted.

The images are in 4x4 patches. Only 7 different textures exist though.
They are repeated in such a way that every texture is a neighbor of
another at least once. Every patch is photometrically normalized
by the function  a*f+b,  where f is the pixel (gray) value, so that
all patches have the same mean and variance. The vertical
boundaries consist of  three random walks and
the x-coordinates of the vertical boundaries are given in the file
border.txt




p1.gif: Patches  from real aerial images.
They represent fields, forests, and
a residential area


p2.gif: Patches from real aerial images.
They represent fields and forest.


p3.gif: Patches originally photographed
by Brodatz, (see reference [5] below)
but arranged and normalized as in P1 and P2.


answer.gif:  The answer image displaying
the patch identities with unique gray levels.
All patch images are arranged according to this
file.
The same gray levels represent the same textures.
For humans, it is easiest to see this image
with a false color map. It can be seen that
the vertical borders are random walks whereas
the horizontal borders are straight lines.

border.tex: Vertical boundary locations, that are realizations of three separate random walks 
(ascii file).  In this file there are exactly 256 rows with three  numbers in each. The first row 
of these triplets represents the x coordinates of the three borders separating the four texture 
patches in the first row of the pixels of
the image (see answer.gif above). The second row of the triplets is the x coordinates of the 
three separating borders in the second row ..etc.


All images are 256x256 with 1 unsigned byte per pixel (gif format).

Published automatic texture segmentation results that use these images,
include the references [1]-[4]:

COPYRIGHT

All rights of P1 and P2 images above belong to their authors.
However, in an unaltered form, these images can be used and published
in scientific contexts e.g journal articles, conference articles ...etc
freely provided that the origin of the images is indicated at the caption (where
these images appear). Citing   Refererence [1] below is an acceptable way of
doing this e.g. by including something like "The patch image is constructed by
Bigun & du Buf, see [Reference]" in the caption. Even commercial products
can use them provided that the origin is indicated. An acceptable way of doing
this is putting these images in the same directory as with the README file.


The use of P3 in scientific publications, as far as the above authors'
added value is concerned, is granted freely, provided that the origin
is cited in the same way as the P1 and the P2 images.

REFERENCES

[1]
@Article{bigun94pami2,
  author =       "J. Bigun and J. M. H.\ du Buf",
  title =        {N-folded symmetries by
 complex moments in {G}abor space},
  journal =      "IEEE-PAMI",
  pages =        "80--87",
  volume =       "16",
  number =       "1",
  year =         "1994",
  OPTannote=    { %old key bigun91pami2
%old key bigun93pami2
},
}


[2]
@article{Bigun92sgproc,
  Author =       {J. Bigun},
  Title =        {Frequency And
Orientation Sensitive Texture Measures Using Linear Symmetry},
  Journal =      {  Signal Processing},
  Year =         {1992},
  Volume =       {29},
  Pages =        {1-16},
  Optannote =    { %Old Key Bigun90sgproc}
}

[3]
@Article{bigun91pami,
  author =       "J. Bigun and G. H. Granlund and J. Wiklund",
  title =        "Multidimensional Orientation Estimation with
                 Applications to Texture Analysis and Optical Flow",
  journal =      "IEEE-PAMI",
  pages =        "775--790",
  volume =       "13",
  number =       "8",
  month =        aug,
  year =         "1991",
  annote=   { Report LiTH-ISY-I-0828 1986 and Report LiTH-ISY-I-1148 1990, both at Computer 
Vision Laboratory, Linköping University, Sweden}
                  }

[4]
@Article{schroeter95pr,
  author =       {P. Schroeter and J. Bigun},
  title =        {Hierarchical image segmentation by multi-dimensional
clustering and orientation adaptive boundary refinement},
  journal =      {   Pattern
Recognition},
  year =         {1995},
  volume =       {28},
  number =       {5},
  pages =        {695-709},
  OPTannote =    {}
}


[5]
@Book{brodatz,
  author =       {P. Brodatz},
  title =        {Textures},
  publisher =    {Dover
publications, Inc.},
  year =         {1966},
  address =      {New York},
  OPTannote =    {}
}



















































.