Image Processing Reference
In-Depth Information
Area-Based vs. Edge-based Algorithms
The numerous algorithms for image segmentation can be separated into two
basic groups of algorithms: 'area-based' and 'edge-based' algorithms. Edge
detection makes sense if an image scene is dominated by linear elements
such as geological faults. As we are interested in larger and contiguous
geographical features being better represented by area objects, this chapter
subsequently focuses on area-based segmentation producing polygonal
image objects. Three principal approaches exist for area-based segmentation:
(1) histogram thresholding, (2) region-based techniques (including region-
growing and split-and-merge), and (3) 'blobs' or scale space analysis (Pinz
1994
). Histogram or global threshold techniques seek out 'image events' (i.e.
significant changes in contrast), but are only applicable on very high contrast
images. These techniques have lost more and more significance with recently
available multi-spectral high resolution images of high variance but lower
contrast. 'Blob' detection techniques do not primarily seek the delineation of
a consistent set of objects but rather they are focused on the dynamic nature
of scale-dependent objects (Hay et al.
2003
). This is achieved by a stepwise
smoothing of the image, usually through Gaussian filtering. Finally, region-
based algorithms aggregate single pixels into homogenous groups. A crucial
component of a region-based segmentation is the criterion of homogeneity
used in the grouping of pixels. Usually homogeneity is considered in terms
of spectral or textural similarity, but it can also be defined by geometric
properties of newly created objects.
a set of exhaustive 'image objects'. Adjacent pixels with
similar reflectance values are grouped into 'segments' by
constraints of size and shape. The entire image information is
aggregated onto a higher level of detail. Koch et al. (
2003
)
point out that any kind of segmentation involves a certain
level of generalization. Segmentation can be performed over
several nested levels, i.e. lower level image objects are
grouped again and again helping to represent image informa-
tion on different levels of homogeneity and thus scales. This
very point arguably makes image objects correspond to our
conceptual understanding of the hierarchical human cogni-
tion process of scenes, thus better matching human perception of the real world.
The Fractal Net Evolution Approach (FNEA) as documented by Baatz and
Schäpe (
1999
) serves as a framework for operational image segmentation imple-
mented in eCognition software from Definiens AG. The core of the eCognition
software is built around the extraction of homogeneous image objects with a specified
level of resolution. This multi-resolution segmentation technique fits into the category
of bottom-up region-growing techniques. Image objects are generated according to
Image
segmentation
seeks to
delineate
spectrally
homogenous
regions known
as 'image
objects'
