prientations.
nd standard
iponents for
ated by the
| 24 pairs of
es and six
Manjunath
rried out on
45812: x512)
00 x 300 )
craft and a
ails of the
performance
as evaluated
late data for
ntation
aluation
e one data
ith another
me object
| one data-
th another
ke object
v is treated
h window is
ible that the
overlapped
achieved by
low and by
id the query
f energy for
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002
Fig. 2. (a). IKONOS and (b).IRS-1D imagery selected for the
present study. (c) and (d) show objects chosen for recognition.
3.1. Performance evaluation
The test patches were simulated with specific distortions to
study the robustness of the proposed approach. Figure 3 gives
the effect of rotation on the normalized distance between the
test and query patterns. As can be seen, the curve peaks at
places where the angles of test pattern are the same as the
query.
1.00
0.90
0.80 4
0.70
0 100 200 300
dnorm
angle of rotation
Fig.3a.
Effect of rotation on image identification.
Similar analyses were carried out to study the effect of scale
and gaussian noise in the test pattern. Results of these are
depicted in Figs. 3b and c. As can be seen from Fig. 3b, the
effect of the scale is asymmetric and the recognition falls quite
sharply when the magnification factor is less than 0.5.
| 1.00
| oso E >
0.40 :
norm
magnification
Fig:3b.
Effect of scale variation on object recognition.
1.00
0.95
0.90
? 0.85
0.80
norm
0 10 20 30 40
std. deviation of noise
Fig. 3c. Effect on noise on image recognition.
In all the cases, it can be seen that it is quite necessary to set an
acceptable threshold for the normalized distance. In our study,
this threshold was set to a value of 0.75.
3.2. Application in Content based browsing
The use of the proposed method for the CBB is shown in Fig. 4.
In Fig.4a, the test and training objects are taken from the same
IKONOS image (see Table 1). The left column of Figure
represents images obtained after the SL-1 stage. There are
"about 23 windows were hit after the histogram normalization,
while this number is reduced to just 4 after the SL-2 and
overlap-window merging. The performance results are
summarized in Table 4.
Similar exercise is carried out with the IRS-1D imagery with
stadium as the object. Here the testing was carried out over the
image of date of pass other than that of the training data. This
is mainly to test the robustness of the method for real-world
variations in the data in both radiometric and geometric sense.
To demonstrate the tolerance further for detecting objects
outside the training data, the proposed method was applied to
an image of another scene. As can be seen from Figs. 3a and 4c,
the query and test objects are quite different with some
common features. It can also be seen that while the two
aircrafts in Fig. 4c are detected, the leftmost aircraft could not
be identified because of its scale variation falling outside the
scale limits specified in Table-2.