ibul 2004
dge map
nbination
5.00, c) -
0.5 and
Ss! Type =
- k) Low
98 and
| are: (7)
function
mega, as
ich is in
Features
ined for
Jezdek
etection
ne area,
ed by
due to
e map
st score
ias for
).4359)
ion of
c)- e)
and
Type =
) Low
and
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
2.4 The HOUGH Transform: Implementation
According to Fitton and Cox (Fitton and Cox, 1998), before the
application of the Hough Transform, several binary edge maps
could be tested, that have been derived using diverse pre-
processing methods. In this work, three different pre-processing
“ methods were tested. The first method (Method A’) is the
method initially proposed by the authors (edge enhancement /
gain-control filtering / thresholding / Zhang-Suen thinning).
Method B’ was based on applying an enhanced Sobel filter,
followed by thresholding and thinning. Finally, Method C" was
based on a 3x3 grayscale morphological dilation, followed by
applying an isotropic filter and thresholding. The input
parameters to the Fitton-Cox algorithm that produced the
highest Pratt value on the satellite image (Table 1) were: (a)
Method C', (b) cutoff h (*6) 730, and, (c) normalization factor
k=1.0. For the DEM, the input parameters to the Fitton-Cox
algorithm that produced the highest Pratt value (Table 1) were:
(a) Method A’, (b) cutoff h (9o) —85, and, (c) normalization
factor k=3.5.
Figure 7: The lineament map extracted by Fitton-Cox
algorithm on the satellite image (left) and the DEM (right).
2.5 Performance evaluation measures of the edge detection
algorithms: Description and implementation
For the quantitative evaluation and assessment of the employed
edge detection algorithms, two evaluation metrics were used:
eThe Rosenfeld evaluation metric (El): This evaluation
scheme is based on the local edge coherence and
measures how well an edge fits to the local
neighborhood of edge pixels but it does not concern
itself with the actual position of the edge, therefore it
is a supplement to Pratt's evaluation metric (Kitchen
and Rosenfeld, 1981; Parker, 1997).
eThe Pratt evaluation metric (E2): This metric is a
formulated function of the distance between correct
and measured edge positions, but it is also indirectly
related to the false positive and false negative edges.
Prat's metric is considered to be a performance
evaluation measure that requires ground truth files.
Therefore, it is directly related to the actual position
of the edge pixels and serves as a more objective
quantitative evaluation measure (Abdou and Pratt,
1979; Parker, 1997).
In the following table the performance evaluation of edge
detection algorithms and the HOUGH Transform (Fitton-Cox
algorithm) is presented for the satellite image and the
corresponding DEM (Table 1):
793
PRATT |
ROSENFELD PRATT!
METRICS ROSENFELD
ALGORITHMS METRICS
ON ON THE
LANDSAT DEM
TM-5
CANNY 0.4680 / 0.6263 0.4332 / 0.6494
ROTHWELL 0.4508 / 0.6529 0.4693 / 0.6318
BLACK 0.4773 / 0.5765 0.4642 / 0.6364
SUSAN 0.3171 / 0.6097 0.3068 / 0.8003
IVERSON- .
ZUCKER 0.4967 / 0.6297 0.2635 / 0.7357
BEZDEK 0.4428 / 0.6595 0.4286 / 0.7442
EDISON 0.4507 / 0.6379 0.4359 / 0.6513
FITTON-COX 0.3017 / 0.7858 0.3716 / 0.7824
Table 1: Performance evaluation metrics (Rosenfeld and Pratt)
for the satellite image and the DEM of Nisyros Island
The Canny and Rothwell algorithms can be found at
ftp://figment.csee.usf.edu/pub/Edge Comparison/source code.
The EDISON algorithm can be found at the address:
http://www.caip.rutgers.edu/riul/. The SUSAN algorithm can be
found at http://www.fmrib.ox.ac.uk/-steve/susan/. The Iverson-
Zucker algorithm can be found at
ftp://ftp.cim.mcgill.ca/pub/people/leei/loglin.tar.gz. The
algorithm by Black and that of Bezdek can be found at
http://marathon.csee.usf.edu/edge/edgecompare main.html.
The modified HOUGH transform suggested by Fitton and Cox
can be obtained from the FTP site of the “Computers and
Geosciences" Journal, Elsevier Publishing at ftp://ftp.iamg.org.
More explanatory details concerning the theoretical background
of the applied algorithms can be found in the corresponding
papers (Canny, 1986; Rothwell, etal, 1994; Iverson and
Zucker, 1995; Black, etal, 1998; Smith and Brady, 1997;
Sutton and Bezdek, 1998; Meer and Georgescu, 2001; Fitton
and Cox, 1998).
2.6 Results and Discussion
A set of optimal edge detectors for a varying combination of
input parameters was applied and provided interesting results as
far as their quantitative assessment is concerned.
Satellite image: For the volcanic field of Nisyros Island, all the
applied algorithms provided a relative high Rosenfeld metric
(local edge coherence) in a small range of values ((0.57-0.66)),
and a Pratt metric in the range of (0.31-0.47). The Rosenfeld
metric stands for the pixel coherence, which appears to be
almost over 6096 for all the applied edge detectors.
The Canny edge detector performs best with the Rothwell
algorithm to follow. It also should be noted that the ground
truth file used in Pratt's evaluation metric contained only the
geologic lineaments, therefore the output values were not high.
If this file had included all lineaments (geologic and non-