Evaluation Criteria and Implementation Details
The improved integrated
MSER
and Harris-Affine algorithm is
used to extract the initial matched features from all test image
pairs. Then, the proposed
OLSM
method is applied to improve
the positional accuracy of the extorted initial matches. To
compute the positional accuracy, the proposed
OLSM
match-
ing results in the search image are transformed to the refer-
ence image space using known or manually computed trans-
formation model in simulated and real images. The interval
between the transformed features and the extracted features in
the reference images is considered as the positional accuracy
of matching in each point pair. A distance threshold
T
D
= 1
pixel is applied to determine the success of
OLSM
method.
Two criteria, including
SuccessRate,
and
RMSE
are used
for evaluation. The
SuccessRate
is the ratio of the number of
point pairs with successful
OLSM
matching to the number of
all point pairs. Also, the root mean square error (
RMSE
) is com-
puted over all the successful
OLSM
match point pairs.
In the
OLSM
matching method, an affine geometric model is
used as the geometric transformation and a linear model with
two parameters is used as the radiometric transformation.
Two window sizes, including 15 × 15 and 31 × 31 pixels are
used. Two Stopping criteria are used for the least square solu-
tion: shift location vector reached 0.01 pixel after 25 itera-
tions. All mentioned experiments were also performed using
standard
LSM
method for comparison.
Results and Discussion
In this section, the performance of the proposed
OLSM
method
is evaluated and compared with the standard
LSM
method in
terms of positional accuracy (
RMSE
), and
Success Rate
. Four sets
of simulated inter-band images and two sets of real images are
used to evaluate the effectiveness of the proposed
OLSM
method.
Figure 8 shows the comparative positional accuracy results
of the proposed
OLSM
method and the standard
LSM
on the
simulated image pairs for three types of synthetic geometric
distortions, including scale, rotation, and viewpoint differ-
ences. Also, the comparative results on the simulated image
pairs for the success rate term are shown in the Figure 9. All
results are shown for two different window sizes 15 × 15 and
31 × 31. In Figure 8 the
RMSE
of initial corresponding features
is also shown for comparison after accurate matching process
using proposed
OLSM
and standard
LSM
methods.
As shown in Figure 8 and Figure 9, the proposed
OLSM
descriptor outperforms the
LSM
method in all the cases. For
little simulated geometric distortions the performance of the
LSM
and
OLSM
is similar, but the percentage of success rate
and positional accuracy of the
LSM
method decreases rapidly
with the increase of synthetic geometric distortion, especially
for significant rotation differences.
As shown in Figure 8, in all cases the
RMSE
of the extracted
features are significantly improved using
OLSM
method. The
obtained results confirm that the original local affine features
cannot provide high accurate matched feature and a post-
processing process to improve the positional accuracy of the
initial features is necessary.
The scene content of the second input image (Figure 5b)
contains homogeneous regions with distinctive edge bound-
aries. As shown in Figure 8, the success rate and positional
accuracy value computed in this image (Figure 8b), are much
less than the obtained
RMSE
of other images, which have more
textured image content. These results indicate that the suc-
cessfulness and positional accuracy of the
LSM
method con-
siderably depend on the scene type. It can also be noticed that
the
RMSE
and success rate results for larger matching window
size (31 × 31) are invariably better than smaller matching win-
dow size (15 × 15) in this test image. It can be concluded that
for images with structured scene content, a larger window
size is preferred since it leads to a matching window with
more information content.
Except for the second image that contains a structured
scene with homogeneous areas, the 31 × 31 matching window
does not provide better results than the 15 × 15 window for
the proposed
OLSM
method. However, a larger window size
generally offers better results for standard
LSM
method due to
more subscription of the template and matching window.
The
RMSE
and success rate results for the real image pairs
are presented in Table 2. For two real image pairs, the pro-
posed
OLSM
method significantly outperformed the standard
LSM
method. As seen in Table 2, the
RMSE
initial extracted
features for first and second real image pairs are 1.346 and
1.061 pixels, respectively. After the proposed
OLSM
accurate
matching, the
RMSE
values are improved to 0.498 and 0.352
pixels (for a 15 × 15 window size), which indicate significant
capability of the proposed method in high accurate feature
matching. Results obtained for the real images also demon-
strated the capability of the proposed method to provide sig-
nificant improvement of the positional accuracy of the local
affine features.
Because the proposed
OLSM
method appropriately approxi-
mates the matching window shape and size in the search
image, it can reduce the number of iteration, and as a result
the computational complexity of the matching process against
the standard
LSM
method. To investigate this subject, the aver-
age number of iterations in all experiments is computed for
both proposed
OLSM
method and standard
LSM
method. Based
on our experimental results, the average iteration number
for
LSM
is 21, and for the proposed
OLSM
method is 9. This
result demonstrates the effectiveness of the oriented matching
window in the proposed
OLSM
method for improving compu-
tational complexity.
As an example, the extracted matched features of the pro-
posed method on the second real image pair of SPOT4 sensor
are shown in Figure 10.
The proposed
OLSM
method modifies the shape and size
of the matching window in the search image from a constant
square to an oriented rectangle based on the ellipse shape
parameter of an initial local affine invariant feature pair. To
decrease the effect of the image scene content on matching
result, an adaptive window size selection process based on
the information content can be used (Gruen, 1985).
T
able
2. M
atching
R
esults
for
R
eal
I
mage
P
airs
Image
Type
Real
Image
Pair
Number of Extracted
Features
RMSE (Pixel)
Success Rate (%)
Harris-Affine MSER
Extracted
Features
LSM
15×15
OLSM
15×15
LSM
31×31
OLSM
31×31
LSM
15×15
OLSM
15×15
LSM
31×31
OLSM
31×31
Close
Range
Graffiti
1375
499
1.346
1.128 0.498 1.091 0.428
9.20
93.91 14.41 91.97
Graffiti
1045
333
Satellite
SPOT 4
1456
601
1.061
0.930 0.352 0.979 0.337 22.99 94.65 27.80 95.72
SPOT 4
1631
685
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2015
739
