The second search-level, SL-2 of Fig.1, includes the feature
extraction of objects based on Gabor filters. These features are
compared with those of the query pattern for taking final
decision.
2.1. Processing at Search Level 1 (SL-1)
As mentioned above, aim of this search is to reduce the number
of regions in which a detailed search is to be followed. This is
accomplished by means of histogram matching of query and
test objects. It is important to note that the query and test
objects are acquired in different times of imaging, thus leading
to variations in contrast and brightness between them. The
histograms are therefore to be first normalized, and is taken
care by the following equations: If. (44,04) and (W,,0,) are their
means and standard deviations of query and test objects
respectively, then the new gray value for test patch will be
G""tm, n) 2 G""(m, n) x slope + offset. (1)
Where G represents the gray value at (m, n) pixel location, the
factor slope is given by (04/0) and the offset given by (uu, —
slope X u, ). This amounts to bring the histogram of the test
object close to that of the query object. It is now left to compare
these histograms to realize quick search of possible regions
where the objects of interest lie. The method adapted here for
comparing them is a normalized histogram intersection
approach given by [7].
hag (tq) = > min (heil, ha iD /A [1]. (2)
Where h,[i] and h,[i] denote test and query histograms. The
probable patches thus obtained in this step are passed to SL-2
processing (Fig.1) and, is described below.
2.2. Processing at Search-level-2 (SL-2)
The first step in the SL-2 module is to estimate the Gabor filter
coefficients of the test object patch, and compare them with
those of the query object. This is achieved by decomposing the
given image patch /(x,y) into the a set of Gabor wavelets. For
the sake of completeness, basic Gabor function and wavelets
are described briefly here. For full details, the reader is referred
to Manjunath and Ma [4].
The Gabor wavelet transform of a given image /(x,y) is defined
as
Wmmn(u,v) = J 1(x,y) Gmn*(u-x,v-y)dxdy. (3)
Where * denotes complex conjugate, and the intervals of
integration extend from - æ to + æ, The functions G are a set of
non-orthogonal wavelets obtained from its mother wavelet :
íi 2 2
di 1 PHASE pem
210,0, 2) 10 re
(4)
Where 4 represents frequency of sinusoidal plane wave, and
(0, 0,) the space constants respectively in x and y directions.
From these a class of self-similar wavelets are generated that
would represent features of the object by dilating (scaling) and
rotating G(x,y) of Eqn. (4) as Gym(X,y) = a”” G(x’,y’). Here, the
scale factor a is greater than 1, (m,n) are integers and the
transformed co-ordinates (x’,y’) are given by — [a"(x cosO +
ysinO) and a^ (-x sinO — y cos0)]. The angle 6 is given by
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring", Hyderabad, India,2002
(nz/K), with K representing the total number of orientations.
Once the wavelets are computed, their mean and standard
deviations can be used as measures of feature components for
the object. These measures are given by
mn = J Won(x,)] 2 dx dy
and On” 2 J/|W,ux;y)-Mas| ^ dxdy. (5)
Thus, the feature vector of an object is represented by the
components of Eqn. (5). In this work, we have used 24 pairs of
these components (corresponding to four scales and six
orientations of Gabor wavelets), much similar to Manjunath
and Ma.
3. RESULTS AND DISCUSSIONS
To validate the above method, experiments were carried out on
the IKONOS PAN image of 1m. resolution (size: 512 x512)
and IRS-1D PAN image of 5.8 m. resolution (300 x 300 )
presented in Fig. 2. From these images, an aircraft and a
stadium were selected as query objects. Details of the
experimentation are given in Table 1. Firstly, the performance
of the proposed method for different distortions was evaluated
through simulation. Secondly, tolerance of multi-date data for
same and/or alike objects was also evaluated.
Table 1. Details of experimentation.
Data Object experimentation
IKONOS Aircraft Training/evaluation
IRS-1D Stadium Training done one data
/evaluation with another
data set — same object
IKONOS Aircraft Training with one data-
evaluation with another
data set- alike object
As shown in Fig. 1, when each search window is treated
independently for the object search, and since search window is
shifted by half of its width each, it is, highly possible that the
objects are hit in more than one window. These overlapped
windows are merged in to a single patch. This is achieved by
first estimating the centroid of each such window and by
minimizing the energy difference between this and the query
object. Image entropy has been used as a measure of energy for
this purpose.
y
Th
stt
the
tes
pl:
Ef
Si
an
efi
sh:
Ef