Knowledge based object extraction technique
Experimental Results
Several experiments with real data are implemented. These experiments are divided into two main
categories; the first category deals with scenes consisting of rectangular shapes only, while the second
x.category deals with scenes consisting of rectangular shapes in addition to other shapes. For numerical
optimization, gradient information, steepest descent optimizer, is used at the beginning to move the initial
parameters to a good location, a low cost area, then different simplex iterations deliver the result to each
other. The simplex iterations start work by beginning with population, initial values, of a big step and
ending with population with a small step. The reason for this is to overcome local minima as much as
possible.
1. Scenes with Rectangular Shapes Only
Three different scenes are selected. The first scene (154x172 pixels) consists of two different
objects, the second (114x107) consists of three objects with the same crop type and the third (373x342)
contains six objects with three different crops. The scenes are shown in figure (1). Before going to further
processing it is necessary to obtain information about the width of the edge spread function. Since there
is no information available about it, the edge-spread function is assumed to be a piecewise linear
symmetrical function. A width of three pixels is selected for edge spread function. This value is used in
the real data experiments. It should be noted that experiment 1 and 3 contain an object with a repeated
pattern texture, a glass house, which is a problematic issue in bottom up approaches. Fig (1) shows an
example of the output of our MBIA algorithm. The figure shows an automatic overlay of DXF layer that
contains the extracted object and the input image. In addition to the DXF layer, the algorithm produces
five geometrical parameters for each object.
Figure (1) Result of real data (scenes with only rectangular shapes)
The results of the three experiments show that the method detected and extracted the existing objects.
Some error appears in experiment 1 due do inaccurate modeling of ESF (the given value of ESF width
was too big). In experiment 2, the existence of noise near the corners makes one of the objects to be
extracted with some error. The reason is that the optimizer could not overcome the local minimum due to
this noise. In experiment 3 the overlaying of the extracted objects by the image shows that all objects are
extracted successfully. Only the green house object is too large. The reason is that the similarity between
the glasshouse and the surrounding which makes both of them appear as one object. The similarity is
mainly, between the neighbor object and one of the textures that construct the glasshouse.
2. Irregular Shapes
To consider agriculture fields as only rectangles is hardly realistic. The existence of irregular shapes
beside the rectangular shapes is more probable. In this section a method for detection of only
rectangular shapes from a scene is explained. Considering an irregular shape as shown in figure (2) left,
Noise.until now the proposed method will try to find a rectangle with minimum cost inside this irregular shape.
For example consider the solid rectangle, candidate object, as identified object inside the actual irregular
object. This object should be rejected. The rejection method is based on the detection of any irregular
extension of the candidate rectangle. Studying the two hypothetical thin objects (ob1 and ob2), it is
evident that Ob2 has lower cost value compared with Ob1. The reason is that since an object is
homogeneous, increasing the size of the object reduces the cost function value. This previous remark is
considered the criteria for rejection of an irregular object. To reject a rectangle, in other words not to
extract a candidate rectangle with minimum cost function, several hypothetical thin rectangles inside and
just outside this candidate object are checked. Figure (2) right shows the checks location. Since we are
interested in detecting any small amount of change across the border, the StdM as cost function for
check is more convenient in this case.
Figure 2 Irregular shapes rejection strategy and checks locations
The parameters of those thin rectangles depend on the definition of the rectangular shape. In the current
context the rectangular shape is identified as any candidate rectangle (with minimum cost function) and
satisfies the following properties.
- (Couple of hypothetical thin rectangles (like object1, 2)) |CF object 1 < CF object 2.
- The extensions of these thin objects are 20% of the dimension of the candidate rectangle.
- The width of the thin object is 0.2 of the width of the candidate rectangular in other words we can say
that the algorithm performs the following steps:
- Chooses a number of small strips such that the extracted rectangle is covered.
- Extends each strip with 20% over the boundary of the rectangle. Do this for both sides of the strip.
If in any case, the cost function of ob2 is less than the cost function of ob1 then the rectangle is an
irregular object and is rejected as a rectangular object. Otherwise the candidate object is rectangle.
To work with scenes that contain these irregular objects, the searching strategy is modified a little to
avoid any slowness. The change is based on using another copy of the image in such a way that one
image is fixed (no object is removed from it). The second image is the processed image. In the
processed image all candidate rectangles are removed. The fixed image will be used for checking only
if the candidate rectangle is rectangle or not. By this modification the algorithm avoids searching at the
areas of rejected rectangles.
