The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
386
Figure 1. Left: A cropland object in Weiterstadt (Germany) in
an orthorectified RGB IKONOS image with a
resolution of 1 m (acquired 24/06/2003). Right: the
edge image as a first step of the verification algorithm.
It is the goal of this paper to present such a segmentation
algorithm and first examples for how it can be used to improve
the overall verification process. We start with a description of
the segmentation algorithm. After that, the way the
segmentation algorithm can be embedded into the verification
process will be presented. This is followed by preliminary
results achieved for images of different resolution and from
different locations, which will be the basis for a discussion of
the possibilities and the limitations of the segmentation
algorithm for this specific application. The paper concludes
with a summary and an outlook.
(2007) used mean colour difference, edge strength of the shared
borders and colour standard deviation to merge segments of
road objects in an iterative way after generating an over
segmented image using the Normalized Cut algorithm. Their
algorithm requires a priori knowledge given by a GIS and the
setting of several thresholds.
2.1 General Segmentation Approach
Let a multispectral image of N bands be represented by the grey
level vectors g(x, y) = [g,(x, y), g 2 (x, y), .... g N (x, y)] T at
position (x, y). It is the goal of region-based segmentation to
partition that image into disjunct regions /?, of homogeneous
grey level vectors and to determine the closed boundary
polygons of these regions. Whereas in theory the boundaries
separating these regions are infinitely thin, the reality of the
imaging process will blur these boundaries, so that they have
actually a certain extent in image space. Typically the region
boundaries correspond to edges in the image that can be
approximated by polygons. Forstner (1994) represents an image
as the union of segment regions R h line regions L„ and point
regions P h based on a classification of each pixel of the images
as being either homogeneous, linear, or point-like. In Fuchs
(1998), the symbolic representation was expanded by the
neighbourhood relations of these regions to define a Feature
Adjacency Graph (FAG). In order to distinguish homogeneous
pixels from other pixels, a measure for homogeneity H can be
used that is based on an analysis of the first derivatives of the
grey values in a local neighbourhood (Forstner, 1994):
2. REGION-BASED SEGMENTATION
The segmentation of objects provides the basis for the
interpretation of images for humans as well as for the fields of
Image Analysis and Computer Vision. Compared to the human
ability to segment objects directly from an image without great
effort, the automatic extraction of objects in the field of image
analyzing is difficult due to problems such as variable lighting
conditions, poor contrast and the presence of noise. Whereas
many segmentation approaches have been presented in the past
(e.g. Gonzalez and Woods, 2002; Forstner, 1994), there is no
generally accepted optimal approach for segmentation,
especially if homogeneous regions are to be extracted. One the
one hand, the extracted segments should represent the digital
image as precisely as possible, even showing relatively small
detectable features; on the other hand, a certain generalisation is
required in order to reduce the impact of noise on the
segmentation results. Furthermore, segmentation should only be
based on a small number of control parameters that should be
easily interpretable.
The algorithm presented in this section starts with a Watershed
segmentation (Gonzalez and Woods, 2002) that achieves a
strong over-segmentation of the image. After that, neighbouring
segments are merged on the basis of a statistical analysis of the
properties of the initial segments and their shared boundary.
The merging process should only require the setting of few
control parameters and no training. In that regard it differs from
existing grouping algorithms. For instance, Luo and Guo (2003)
introduced a general grouping algorithm based on Markov
random fields, using single segment properties such as area,
convexity and colour variances, and pair-wise properties such
as colour differences and edge strength along the shared
boundary. The algorithm requires a training phase. Grote et al.
?
In Equation 1, g ix and g iy are the first derivatives of the
grey levels g, of band i by x and y, respectively. G is a
Gaussian smoothing filter with scale parameter , and 'J is
the variance of the smoothed grey level differences G * g^
and G * g iy , which can be derived from an estimate of the
noise variance of band i (Briigelmann and Forstner, 1992)
by error propagation. The sum is to be taken over the N bands
of the digital image. The scale parameter defines the size of
the local neighbourhood that is taken into account. By
normalising the smoothed grey level differences by their
standard deviations, the selection of a threshold H max for H to
distinguish homogeneous pixels from others can be reduced to
the selection of a significance level for a statistical test
(Forstner, 1994).
The image regions /?, could be determined as connected
components of homogeneous pixels, thus of pixels whose
homogeneity measure H is smaller than H max . However, small
gaps within extracted line regions that occur due to poor local
contrast often cause a spilling effect, i.e., the erroneous merging
of regions that represent different object parts. Furthermore, it is
not straightforward to obtain meaningful closed boundary
polygons of the homogeneous segments, gives a typical result.
The main edges of the image are represented well, although the
edges information appears to be captured incompletely. The
segments in the label image do not represent the image structure
well due to small gaps in their boundaries. A considerable
portion of the image is not assigned to any label.
