ol. XXXVIII, Part 7B
In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
MAGES BY
JTHM
ction, Weight-based
ation techniques fall
;h extracts structures
Conventional image
ts. In this study, we
detection and multi -
Features) method to
apology rules, which
tgical map-matching
and present a robust
; and image rotation
ity and robustness to
ts that are covered to
ter accuracy than the
dded into three main
", 2. Description of
between images. The
its repeatability. The
tinctive but robust to
trie and radiometric
tching technique has
:tion scheme has to
for automatic image
RTop) for temporal
me and space-borne)
1STRATION
ist four stages as any
e significant features
Jp Robust Features)
ges and converted to
relationship of these
id converts them to
ion model between
1 rules for a weight-
im (tMM) manage,
sed image according
it number of CPs the
vith test point error
2.1 Significant Features
The extraction of significant features performed by SURF
based on sum of Haar wavelet responses in two directions
(Brown, and Lowe, 2002). The Harris comer detection is the
most widely used detector up to day due to its excellent results
(Lindeberg, 2004). Fast-Hessian Detector is based on integral
image and approximation. Integral image represents the sum of
all pixels in the input image within a rectangular region formed
by origin. Approximated Hessian detector uses box filter.
While, scale variations are detected by different sizes of box
filter. The next stage is to make descriptor of local gray level
geometry feature. The local feature representing vector is made
by combination of Haar wavelet response. The values of
dominant directions are defined in relation to the selected
region orientation.
The final stage of features extraction based on Canny edge
detector (Canny, 1986), which is popular edge operator that
widely used in digital image processing including remote
sensing, and lines are extracted using Hough Transform (Duda,
and Hart, 1975). Since SURF image of magnitude and direction
is the base layer for feature detection, we propose to adjust
several stages of Canny operator. In our work the four stages of
Canny operator modified into three stages. Firstly, the image is
smoothed by Gaussian convolution. We could skip the first
derivative operator as it provides in the SURF image. Secondly,
the process of non-maximal suppression (NMS) is imposed on
the smoothed SURF image. Finally, the edge tracking process
exhibits hysteresis controlled by two predefined thresholds.
Traditional Canny operator carries on the edge tracking
controlled by two thresholds, namely a high threshold and a low
threshold. The tracking of one edge begin at a pixel whose
gradient is larger than the high threshold, and tracking
continues in both directions out from that pixel until no more
pixel whose gradient is larger than the low threshold. The
process is called hysteresis. It is usually difficult to set the two
thresholds properly, especially for remote sensing image. The
illumination and contrast of different portions of remote sensing
image are often non-uniform.
The suggested process extracts long edges related to roads
features with Hough Transform prior Canny operator. Thus roof
detection could be implemented without predefined thresholds.
As now the long edges related to roofs features and the edge
tracking is carried out by inside edges. AS, it is difficult to
detect continuous and stable edges solely from the images the
morphological closing operation is employed. It's produced by
the combination of dilation and erosion operations. During the
process, the edges detected areas are integrated into the
individual roof features. Finally, all the extracted features (roads
and roofs) were converted from raster to vector format and
saved as GIS project. While roofs converted to polygons, roads
have been converted to polylines that cross-along the central
line of detected (long edged) features.
2.2 Topological Method
Topological matching is usually used to reduce the search
range or check the results of geometric matching, since it is
seldom used alone. Topological methods can spread the
matching into the whole network, but this requires high
topological similarity of two data sets. Topological transfer
method (Tomaselli, 1994) is representative of this type. If the
polygons are matched, then according to the relationship of
polygons and polylines, polylines to polylines matching can be
deduced.
Data preprocessing stage standardizes the input data sets,
ensures the conflation data sets have a same data format, the
same north direction, and have overlapped spatial coverage. It
also ensures the data sets have maximum similarity which is the
basis of common objects matching.
A Reeb graph is a topological and skeletal structure for an
object of arbitrary dimensions (Berg and Kreveld, 1997). In
Topology Matching, the Reeb graph is used as a search key that
represents shapes of the features. A node of the Reeb graph
represents a connected component in a particular region, and
adjacent nodes are linked by an edge if the corresponding
connected components of the object contact each other. The
Reeb graph is constructed by re-partitioning each region. The
Multiresolutional Reeb Graph (MRG) begins with the
construction of a Reeb graph having the finest resolution
desired. Second, position of an inserted vertex is calculated by
interpolating the positions of the relevant two vertices in the
same proportion. Thirdly, the T-sets (connected components of
triangles) are calculated. Fourthly, if two T-sets between
adjacent ranges are connected, corresponding R-nodes are
connected by an R-edge. The complete notification as follows:
1. R-node: A node in an MRG, 2. R-edge: An edge connecting
R-nodes in an MRG, 3. T-set: A connected component in a
region, 4. pn-range: A range of the function pn concerning an
R-node or a T-set.
2.3 Weight-based Topological Map-matching
This subsection gives an overview of how similarity is
calculated using MRGs. A weighting approach in selecting the
correct feature from the candidates improves the accuracy of
correct pair identification (Greenfeld, 2002). The suggested
algorithm assigns weights for all candidates using similarity in
linear network and transfer matching and selects the pair with
highest weight score as the potentially corrected CPs. The
mathematic representation of Root Mean Square (RMS) error
value of map-matching process estimates accuracy of algorithm
by predefined threshold. If the algorithm fails to identify the
correct CPs pair among the candidate pairs and RMS error
oversize threshold then the algorithm regenerates another pair
with lower weight by an optional loop stage.
2.4 Test Point Error (TPE)
The cubic B-Spline convolution supports image
transformation and resampling, as it is computed raw-by-raw
and column-by-column (Unser et al., 1993). The results of
proposed convolution tested with Test Point Error (TPE). The
test points are CPs that were deliberately excluded from the
calculation of the mapping parameters. The concept of this
method can be extended such that the distance between
corresponding ‘test’ lines or surfaces is measured (Nelson et al.,
1997). TPE cannot be set to zero by overfitting. This method
can be used only if a sufficient number of the CP’s is available.
Otherwise, the exclusion of several CP’s may result in
inaccurate estimation of mapping parameters. In our algorithm
10% of all CPs are excluded for TPE evaluation. Once again, if
the algorithm fails to transform and resample sensed image and
TPE error oversize threshold then the algorithm regenerates
another pair with lower weight in stage 2 (Weight-based
Topological Map-matching) by an optional loop stage.
3. RESULTS
The following section presents both simulated and real-world
results. First, we evaluate the effect of multi-temporal
el Aviv 69978, Israel,
99