s. But, at this
lid objects is
package in the
ole in the HIPS
digital number
0. The map is
HIPS system to
ngs to achieve
lid objects. The
cts like lines or
ng the drawing
ing solid object
m for further
clutter from the
his can be done
nique DN value
is section and
f objects to be
cts of
h area
adient
An algorithm is used which scans the map from top left to
bottom right and while scanning it assigns value DN 1 to
the first object it encounters and DN 2 to second object and
so on to all objects in the map. Figure 1(c) shows each solid
building object, separate roads and each letter of street or
place names as a separate object in the map and all are
assigned an individual unique DN numbers between 0-255.
2.1.3 Elimination of Clutter Objects
At this stage each object in the map has a unique DN value.
An algorithm is used with a threshold size (number of
pixels) value and below this value all objects are removed
from the map. The algorithm first looks for an object
having value DN 1 and counts number of pixels in that
object, and if it is greater in size than the threshold size
chosen then the object with DN 1 remains in the map,
otherwise it is removed i.e. a zero DN value is given. In the
same manner the algorithm looks for other DN value
objects in the map, and removes clutter and small size
objects. Figure 1(d) clearly shows a map of the solid
buildings after removal of objects not required .
2.1.4 Map of Building Region Boundaries
To find the boundaries of the solid building regions, an
algorithm is used which works in a very simple way. It scans
the map of solid building regions from top left to bottom
right. The first pixel of a solid object it encounters is
considered as the first boundary pixel of that object. From
that boundary pixel, the algorithm starts looking for such
neighbouring pixels which have the same DN value of that
object as well as lying adjacent to the background DN value
i.e. DN 0, and it then traces the boundary of that object.
After this solid building region boundary is traced, the
algorithm looks for the next solid object, and in the same
manner traces the boundaries of the object and the
subsequent solid objects in the map [see Figure 1(e)].
2.1.5 Gradient Direction of the Map Boundaries
For determining the best match between the map and the
image, edge pixel direction in the image and boundary pixel
direction in the map are used. An algorithm is used to
determine the directional component of each map boundary
pixel. A two frame sequence input is used to apply this
algorithm. The first frame contains the map boundaries and
the second frame contains the solid regions from which the
boundaries were defined. The output also results in two
frames, first map boundaries, and second map direction as
shown in Figure 1(f). The map with these two frames,
boundaries and its directions, are ready at this stage to be
used as input for matching .
2.2 Preparing the Image for Matching
The Farnborough subscene image is shown in Figure 2(a)
which consists of 230 x 180 pixels. The aim here is to
extract edges that define edges of the building regions and at
the same time to suppress edges that do not represent
building regions. The pre matching steps for preparing
image for matching are described below:
2.2.1 Edge Preserve Smoothing
An edge preserving filter is applied prior to edge detection
to strengthen the grey level discontinuties between different
land cover types, and to reduce the detection of edges in
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
areas of texture that are internal to regions. An algorithm
used is an adaption of that outlined by Matsuyama et al,
(1980) and Tomita et al, (1977). A window with nine masks
(filters) is passed over the image and the variance is
measured in 9 orientations (masks) around the central pixel.
The orientation of minimum pixel variance is determined,
and the mean of this is given to the central pixel. The
selected orientation is never accross an edge. This is
performed for each pixel in the image. The algorithm is
iteratively applied to achieve maximum homogeneity of
each region in the image as shown in Figure 2(b).
2.2.2 Region Segmentation
Segmentation is the splitting up of an image into regions
which hold properties distinct from their neighbours, and it
is generally approached from two points of view: by
detection of edges that separate regions, or by the extraction
of regions directly. Using a histogram derived from the edge
preserving smooth image and thresholding it at value DN 38
resulted in direct extraction of building regions including
some clutter as shown in Figure 2(c).
2.2.3 Edge Enhancement
Edge enhancement determines, for each pixel in the image,
its edge strength and the direction of the gradient of the edge
at that point. This is obtained by image differentation,
which is itself achieved by the convolution of various
kernels with the image. The Sobel Operator is used which
consists of two kernels (X and Y) and are passed accross the
region segmentated Farnborough image. The strength of the
edge at the central pixel of the kernel and its gradient is
determined for each pixel in the image as:
Strength = V[(Result of convolving kernel X)? +
(Result of convolving kernel Y)2] (1)
Direction = tan'![(Result of convolving kernel X) /
(Result of convolving kernel Y)] (2)
The result of edge enhancement is shown in Figure 2(d).
2.2.4 Non Maximal Suppression
The result of edge enhancement shows edges of two or more
pixels thick. Non Maximal suppression seeks to remove
those edgels (pixels) that are not local maxima, thus
sharping the representation of edges effectively to thin the
edge to a single pixel width. An algorithm is used which
considers the edge strength and gradient direction
information. The edge pixel is passed if the two
neighbouring pixels along the gradient direction are less
than or equal to it in strength, if not the pixel is set to zero.
The result of Non Maximal Suppression algorithm is shown
in Figure 2(e).
2.2.5 Alter Directions of Edge Pixel Gradient
The edge gradient direction is a useful element in the
matching procedure, where edge pixel gradient directions are
compared to map boundary pixel gradient directions to
obtain good matches. However, there is always some
rotational difference between the image and the map spaces.
To compensate this, an adjustment is required to the gradient
directions calculated for the edge pixels. Four control point
pairs are selected in well distributed manner manually from
the map and the image, and are used to define a similarity