|
|
|
Data Processing, Algorithm and Modelling
|
Knowledge based object extraction technique
Mohamed Abdel-salam
PhD candidate & Research Assistant
Department of Geomatics Engineering, The University of Calgary
2500 University Drive N.W.
Calgary, AB, T2N 1N4, Canada
Email: mabdel@ucalgary.ca
Phone +1 (403)-282-9308
Fax +1 (403)284-1980
Abstract
Since the beginning of remote sensing technology, bottom up methods for interpretation of scenes
monopolize remote sensing activities. For example, in the conventional per-pixel classification, which is a
bottom up approach, the system uses the information hidden in spectral data only. Such techniques
suffer from drawbacks caused by mixed pixels and spectral overlap. Knowledge about objects, sensor
and environment is hardly considered. Therefore a lot of problems appear in the conventional methods
are related to insufficient representation of the available knowledge. The difference between human
interpretation and any computer assiste d image analysis is that humans use spatial information like
texture, shape, shade, size, site, association and topology, while computers are powerful in handling the
gray values in the image. In this paper a new technique for object recognition and identification is
proposed. The proposed technique employs knowledge about objects and edge spread function. These
two types of knowledge are hardly included in the conventional methods. The heart of proposed
technique is the integration of the existing knowledge and numerical optimization techniques. A new
proposed cost function, which is used by the optimizer, and a rectangular class of objects have been
studied in this paper. The high performance of the method is clearly shown in the obtained results.
Varieties of objects have been extracted with the proposed method some of them as will be shown are
almost impossible for bottom up approach.
Introduction
Bottom approaches for remote sensing data interpretation use only spectral knowledge represented in
pixels. Other existing knowledge in the scene is not considered at all. These exiting knowledge can be
categorized into three main branches. The first branch is related to the object: geometric and radiometric.
The second branch is related to the sensor: imaging technique, bluer model, noise model, image
acquisition model and projection model. The third branch is related to the environment: path and
illumination model (Mulder and Fang, 1994), (Van der Heijden, 1994). For most farms, the agriculture
fields enjoy certain shapes and border thickness, and even orientations. Integrating this simple existing
knowledge in the interpretation process can enhance the outcome. The way to integrate different types of
knowledge in known as model based image analysis. It is popular in different areas like industrial
applications, medical applications and computer vision. It tries to model a scene in terms of objects and
parameters and then it estimates these parameters. Therefore, using model based approach for
interpreting imagery can be powerful.
Proposed Technique
The proposed technique starts with the initiation of a hypothetical object model and it checks the
corresponding statistical information of the multispectral information inside this object. Based on this
statistical characteristic, the object parameters are adjusted through a numerical optimizer till a given
satisfactory statistical criterion is satisfied. This criterion acts as a cost function. Then the object will be
stored as identified object. The method will start again to search for another object, avoiding the positions
of the previous objects.
- Object Model
A rectangular shape class is used in this paper. The rectangular-shape class is five parameters primitive.
Its parameters are locations (two parameters), length of the rectangular primitive, width of rectangular
primitive and the orientation of the primitive.
- Edge Spread Function (ESF) Model
Pixels at boundaries are among the major causes of uncertainty of object boundaries that lead to
inaccurate parameter estimation of objects from remotely sensed data. Two issues are related to the
edge-spread function: how to model it and how to include it. It is assumed that the edge-spread function
is a piecewise linear function of a certain width. The edge-spread function is assumed to be symmetry
independent of the orientation of the edge. Compensating the effect of ESF is based on detection of the
optimum size of the Hypothetical object and then the hypothetical object size is increased by half the
width of the edge-spread function in each direction
- The Proposed Objective (Cost) Function
Cost, objective, function expresses the consistency of a hypothesis with the truth. It should reflect big
value if the hypothesis wrong and small value if the hypothesis true. In most model-based techniques,
pixel classification results are the cost function parameters. The dependence of cost function on
classification should be avoided since classification procedure is subjective. In this paper a new cost
function is proposed. The standard deviation of the mean (StdM), the normalized standard deviation
(NStd), and the Std are candidates as objective functions. Therefore, the only prior knowledge taken into
account is that different objects in the image show a contrast. No knowledge about the radiometric
distribution is necessary. The Std tends to make the object as small as possible; therefore, it can’t be a
cost function. Standard deviation of the mean is a well-known quantity in observation and measuring
samples. StdM and NStd depend on the both the population variance and sample size. One reason for
choosing these objective functions is that they tend to enlarge homogeneous objects as much as
possible. The average of the cost function in the three bands is used. In the following the equations of
NStd are given.
where:
x : is the digital number, n is the number of pixels
is the mean of the digital number inside the hypothetical object
The StdM is comparable in its performance to normalized Std. The normalized Std has the same
behaviour of StdM, but NStd is not sensitive toward the noisy parts like the StdM. So the selection of
StdM is suitable for the current problem and this is validated by real data.
- The Optimizer
Model based image analysis needs a huge amount of calculations. The exhaustive search method,
which evaluates all possible combination of the parameter values, is time consuming. It is not realistic for
large problems. To overcome this problem, optimization is used. Optimization starts with an objective
function and a set of initial parameters and the goal is to find the set of parameters that gives the optimal
objective value (Gile,1997). The optimal objective value of the objective function is problem dependent.
Two types of optimizers used in this technique are: the Steepest Descent and Simplex method
(McKeown,1990).
|
|
|
|
|
|
|
