International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
In this paper, ROI of the input image are extracted by wavelet
transform and morphological processing to reduce the object
search region, after that, the ROI is divided into several
windows to facilitate the feature extraction. From template
image, a hierarchical template database is built for hierarchical
feature template correlation to recognize the interested object.
the process of the object recognition begin from the correlating
with the most common feature( i.e. first level template), and
possible object candidates are acquired. Then after the further
processing by the detail template (second-level and to last), we
can get the more accuracy object that we want to find.
3. ROI EXTRACTION BASED ON WAVELET
TRANSFORMS
It is a difficult thing to find a small object in a large remote
sensing image, and it is also a tedious work to search the object
by matching the object template image to the input image, Even
we know there is a object in the image, it is still difficult to find
this object for computer itself. To avoid consuming long time to
search large image. It is necessary to find where are the most
possible parts in which the object exists. We can find that the
regions in which have man-made objects has abrupt variety of
intensity because man-made object has outstanding edge
feature. Transform methods convert raw data to transform
coefficients in order to obtain a more efficient representation of
the data for feature extraction Wavelet transform just provide a
tool for detect this kind abrupt variety feature. Wavelet methods
provide several advantages over the Fourier-based methods.
One of the most important advantages of the wavelets is that
wavelet bases have local support in the space and frequency,
other advantages of the wavelet methods are the availability of
fast algorithms. Daubechies wavelet is used to transform the
input image to detail parts which reflects the gray values variety
of the images. Binarizing the detail image we can acquire an
image that reflects the abrupt gray variety regions which are our
interested region, but may be in the image there are several
discrete edges, the morphological processing is used to deal
with this problem, by dilation, removing spur pixel and so on,
the morphological processing transform the detail image to a
region, we can get the parts in the image that possibly object
exists In.
4. ESTABLISHING THE HIERARCHICAL FEATURE
TEMPLATE OF OBJECT
The object template is a representative of object that is built
based on analysis of the properties of object. There are two
kinds of object template used in image processing. One is
geometric template of object based on the real gray values of
object image, such as the object image cut from the large
example image, and searching the whole input image based on
this geometric template. This is the simplest class of object
recognition methods (Brown, 1992), a survey of matching
approaches is given in (Brown, 1992). The template and input
image can get a good match when there are no distinct change
in view point, view angle and climate condition during the time
image acquired. But we can't ensure we can get two images
with little change at two different time, especially in remote
sensing image, and to match the two image, the computation is
also time consuming. Those drawbacks (huge Time-consuming
and unrobust to change) limit the application range of this
method.
Another more complex method for building the template for
object recognition is a abstract feature template based on te
description of object feature properties. There are several kind
of feature that can be used to describe the object. Figure 2 lists
the feature we can acquire in remote sensing imageries.
888
Spectrum vector
PCA
Object Properties Location
Edge
Spectrum feature | | Shape
Geometric feature |_ s -
Relative location
| Neighbor relation
Shadow
Context feature
Radiometric feature ||
— Grey values
Texture
Color
Figure 2. The features for object properties in remote sensing
The features for representing the object are variety because
there are so many different objects. Recognizing different
object from remote sensing image can't get a good result if we
just use one kind of feature. But we can get a good distinction
just by threshold when the ratio of gray intensity between
objects and background, we can think it is a vehicle with low
false probability when there is a oblong shape on the highway
in the remote sensing image. So, we should choose the most
obvious feature in the image that can reflect the properties of
object to build the object feature template.
From the process that a people recognizes another person who
walks near from far (figure 3), we can know that recognition is
a process from fuzzy to accuracy, from coarse to detail. And
this process is a hierarchical recognizing process.
[Find a people I«— Profile and pose of a person
Distinguish between : |
5 Hair and clothes
man and woman |
Distinguish between
Stranger and familiar |“ |
Particular feature of the familiar
Such as complexion statue, weight
JE3U 0] J£] UIOJ
Further Identification
Change of feature
Figure 3. The process of recognizing a person from far to near,
the shadow parts are the feature used at different recognition
stage
In the process above, we use different features at different
recognition stage. Emulate the process what we do in
recognizing a person. We should build a hierarchical feature
template database to adapt to the different stage of recognition.
The first level is the basic object feature template that reflects
the common features of the same class object. This level feature
should satisfy three assumption: 1) the selected feature can be
acquired in remote sensing image, 2) they are obvious features
can be relatively extracted, 3) those feature represent the most
common properties of this object class not a certain one. Other
levels feature templates represent gradually object from coarse
feature to detail one. The principle of determining the levels is
Intern
accor
comp
featut
in the
image
establ
groun
image
featur
The |
vehic
air. T
inforr
objeci
templ
Figur
remot
We us
width
is qui
with
be de:
we fo
profil
most «
What
object
recogi
sectio
Hausc
repres
demai
chang
assum
match
(DPC
profili
5.1 TI
When
locati
that tl
the nu
compi
