Geografic-Information-System/imagery/i.segment/i.segment.html

<h2>DESCRIPTION</h2>
Image segmentation or object recognition is the process of grouping 
similar pixels into unique segments, also refered to as objects. 
Boundary and region based algorithms are described in the literature, 
currently a region growing and merging algorithm is implemented. Each 
object found during the segmentation process is given a unique ID and 
is a collection of contiguous pixels meeting some criteria. Note the 
contrast with image classification where all pixels similar to each 
other are assigned to the same class and do not need to be contiguous. 
The image segmentation results can be useful on their own, or used as a 
preprocessing step for image classification. The segmentation 
preprocessing step can reduce noise and speed up the classification.

<h2>NOTES</h2>

<h3>Region Growing and Merging</h3>

This segmentation algorithm sequentially examines all current segments
in the raster map. The similarity between the current segment and each
of its neighbors is calculated according to the given distance
formula. Segments will be merged if they meet a number of criteria,
including:

<ol>
  <li>The pair is mutually most similar to each other (the similarity
distance will be smaller than to any other neighbor), and</li>
  <li>The similarity must be lower than the input threshold. The
process is repeated until no merges are made during a complete pass.</li>
</ol>

<h3>Similarity and Threshold</h3>
The similarity between segments and unmerged objects is used to 
determine which objects are merged. Smaller distance values indicate a 
closer match, with a similarity score of zero for identical pixels.
<p>
During normal processing, merges are only allowed when the 
similarity between two segments is lower than the given 
threshold value. During the final pass, however, if a minimum 
segment size of 2 or larger is given with the <b>minsize</b> 
parameter, segments with a smaller pixel count will be merged with 
their most similar neighbor even if the similarity is greater than 
the threshold.
<p>
The <b>threshold</b> must be larger than 0.0 and smaller than 1.0. A threshold 
of 0 would allow only identical valued pixels to be merged, while a 
threshold of 1 would allow everything to be merged. Initial empirical 
tests indicate threshold values of 0.01 to 0.05 are reasonable values 
to start. It is recommended to start with a low value, e.g. 0.01, and 
then perform hierachical segmentation by using the output of the last 
run as <b>seeds</b> for the next run.

<h4>Calculation Formulas</h4>
Both Euclidean and Manhattan distances use the normal definition, 
considering each raster in the image group as a dimension.

In future, the distance calculation will also take into account the
shape characteristics of the segments. The normal distances are then
multiplied by the input radiometric weight. Next an additional
contribution is added: <tt>(1-radioweight) * {smoothness * smoothness
weight + compactness * (1-smoothness weight)}</tt>,
where <tt>compactness = Perimeter Length / sqrt( Area )</tt>
and <tt>smoothness = Perimeter Length / Bounding Box</tt>. The
perimeter length is estimated as the number of pixel sides the segment
has.

<h3>Seeds</h3>
The seeds map can be used to provide either seed pixels (random or 
selected points from which to start the segmentation process) or 
seed segments. If the seeds are the results of a previous segmentation 
with lower threshold, hierarchical segmentation can be performed. The 
different approaches are automatically detected by the program: any 
pixels that have identical seed values and are contiguous will be 
assigned a unique segment ID.
<p>
It is expected that the <b>minsize</b> will be set to 1 if a seed 
map is used, but the program will allow other values to be used. If 
both options are used, the final iteration that ignores the 
threshold will also ignore the seed map and force merges for all 
pixels (not just segments that have grown/merged from the seeds).

<h3>Maximum number of starting segments</h3>

For the region growing algorithm without starting seeds, each pixel is
sequentially numbered.  The current limit with CELL storage is 2
billion starting segment IDs.  If the initial map has a larger number
of non-null pixels, there are two workarounds:
<ol>
  <li>Use starting seed pixels. (Maximum 2 billion pixels can be 
labeled with positive integers.)</li>
  <li>Use starting seed segments. (By initial classification or other 
methods.)</li>
</ol>

<h3>Boundary Constraints</h3>
Boundary constraints limit the adjacency of pixels and segments. 
Each unique value present in the <b>bounds</b> raster are 
considered as a MASK. Thus no segments in the final segmentated map 
will cross a boundary, even if their spectral data is very similar.

<h3>Minimum Segment Size</h3>
To reduce the salt and pepper affect, a <b>minsize</b> greater 
than 1 will add one additional pass to the processing. During the 
final pass, the threshold is ignored for any segments smaller then 
the set size, thus forcing very small segments to merge with their 
most similar neighbor.

<h3>Goodness of Fit</h3>
The <b>goodness</b> of fit for each pixel is calculated as 1 - distance 
of the pixel to the object it belongs to. The distance is calculated 
with the selected <b>similarity</b> method. A value of 1 means 
identical values, perfect fit, and a value of 0 means maximum possible 
distance, worst possible fit.

<h2>EXAMPLE</h2>
This example uses the ortho photograph included in the NC Sample 
Dataset. Set up an imagery group:
<div class="code"><pre>
i.group group=ortho_group input=ortho_2001_t792_1m@PERMANENT
</pre></div>

<p>Set the region to a smaller test region (resolution taken from
input ortho photograph).

<div class="code"><pre>
g.region -p raster=ortho_2001_t792_1m n=220446 s=220075 e=639151 w=638592
</pre></div>

Try out a low threshold and check the results.
<div class="code"><pre>
i.segment group=ortho_group output=ortho_segs_l1 threshold=0.02
</pre></div>

<center>
<img src="ortho_segs_l1.jpg">
</center>

<p>
From a visual inspection, it seems this results in too many segments. 
Increasing the threshold, using the previous results as seeds, 
and setting a minimum size of 2:
<div class="code"><pre>
i.segment group=ortho_group output=ortho_segs_l2 threshold=0.05 seeds=ortho_segs_l1 min=2

i.segment group=ortho_group output=ortho_segs_l3 threshold=0.1 seeds=ortho_segs_l2

i.segment group=ortho_group output=ortho_segs_l4 threshold=0.2 seeds=ortho_segs_l3

i.segment group=ortho_group output=ortho_segs_l5 threshold=0.3 seeds=ortho_segs_l4
</pre></div>

<center>
<img src="ortho_segs_l2_l5.jpg">
</center>

<p>
The output <tt>ortho_segs_l4</tt> with <b>threshold</b>=0.2 still has
too many segments, but the output with <b>threshold</b>=0.3 has too few
segments. A threshold value of 0.25 seems to be a good choice. There
is also some noise in the image, lets next force all segments smaller
than 10 pixels to be merged into their most similar neighbor (even if
they are less similar than required by our threshold):

<p>Set the region to match the entire map(s) in the group.
<div class="code"><pre>
g.region -p raster=ortho_2001_t792_1m@PERMANENT
</pre></div>

<p>
Run <em>i.segment</em> on the full map:

<div class="code"><pre>
i.segment group=ortho_group output=ortho_segs_final threshold=0.25 min=10
</pre></div>

<center>
<img src="ortho_segs_final.jpg">
</center>

<p>
Processing the entire ortho image with nearly 10 million pixels took
about 450 times more then for the final run.

<h2>TODO</h2>
<h3>Functionality</h3>
<ul>
<li>Further testing of the shape characteristics (smoothness, 
compactness), if it looks good it should be added.
(<b>in progress</b>)</li>
<li>Malahanobis distance for the similarity calculation.</li>
</ul>
<h3>Use of Segmentation Results</h3>
<ul>
<li>Improve the optional output from this module, or better yet, add a 
module for <em>i.segment.metrics</em>.</li>
<li>Providing updates to <em><a href="i.maxlik.html">i.maxlik</a></em>
to ensure the segmentation output can be used as input for the
existing classification functionality.</li>
<li>Integration/workflow for <em>r.fuzzy</em> (Addon).</li>
</ul>

<h3>Speed</h3>
<ul>
<li>See create_isegs.c</li>
</ul>

<h2>REFERENCES</h2>
This project was first developed during GSoC 2012. Project documentation, 
Image Segmentation references, and other information is at the 
<a href="http://grasswiki.osgeo.org/wiki/GRASS_GSoC_2012_Image_Segmentation">project wiki</a>.
<p>
Information about 
<a href="http://grasswiki.osgeo.org/wiki/Image_classification">classification in GRASS</a> 
is at available on the wiki.

<h2>SEE ALSO</h2>
<em>
<a href="g.gui.iclass.html">g.gui.iclass</a>,
<a href="i.group.html">i.group</a>, 
<a href="i.maxlik.html">i.maxlik</a>, 
<a href="i.smap.html">i.smap</a>, 
<a href="r.kappa.html">r.kappa</a>
</em>

<h2>AUTHORS</h2>
Eric Momsen - North Dakota State University<br>
Markus Metz (GSoC Mentor)

<p>
<i>Last changed: $Date$</i>