Digital Signal Processing Reference
In-Depth Information
1. Pixel/voxel identification methods—These techniques segment an image by
labeling individual image pixels or voxels based on characteristics such as
intensity, texture, etc. Groups of pixels/voxels with identical labels then form
different segments of the image.
2. Region/volume identification methods—These techniques are based on iden-
tifying groups of pixels with identical characteristics. These approaches are
predominantly based on specific characteristics of regions such as intensity, tex-
ture, etc. and employ different region growing or splitting-and-merging methods
to link together regions with similar values for a given (set) of parameters to
create image segments of interest.
3. Edge/surface identification methods—These techniques typically involve identi-
fying boundaries in images. Edge/surface identification methods are typically
based on techniques used to identify change in the value of a given (set) of
parameters between regions, e.g., change in intensity between two regions.
Segments of interest can either be the edges/surfaces or the regions contained
within or separated by these boundaries.
Region/volume-based methods and edge/surface-based methods can be used to
identify the same image primitives. Edge/surface identification help identify the
surfaces/edges in an image and thus separate the volumes/regions contained within.
Thus, these two approaches differ mainly in terms of methodology rather than
the targeted image primitive. The choice of the approach to solve a particular
segmentation approach depends largely on the characteristics of the images involved
and the ease and accuracy with which either set of primitives are identifiable.
Different image processing techniques can be adopted to segment by identifying the
basic image primitives. These techniques are grouped into different categories based
on whether they utilize a variety of simple image characteristics (such as intensity
or image gradient) and/or more sophisticated mathematical or statistical models. Of
these, the two most common groups are the following:
1. Threshold-based techniques, which classify image segments (globally or locally)
based on predefined criteria (e.g., intensity), where the information is represented
by the intensity histogram.
2. Deformable model-based techniques, which involve (1) detecting and fitting
segments in the image to a priori known template, and (2) modifying the template
to more accurately match the individual image. The templates can be parametric
shapes, e.g., straight lines, circles, ellipses, or special predefined shape templates
such as an atlas of the brain or a model of the left ventricle of the heart.
The next two sections give a brief overview of these two groups of techniques, which
can be used to segment an image based on identification of any of the three basic
image primitives noted above.
