3.1 Image processing 
Original images by default present an excessive amount of 
information where sometimes, like now, this image 
representation is not desirable. In order to transform this 
information to a useful one, an edge detector (Pratt, 1991; 
Gonzalez and Woods, 1992), for example the gradient operator, 
is used. If the edge is defined as a change in image intensity, a 
gradient operator is the choice (Mikhail et al., 2001). Roberts, 
Prewitt and Sobel edge detector kernels are commonly included 
in the gradient operators. 
Another edge detector, which is very popular in image 
processing, is the Canny operator (Canny, 1986). Susan 
operator (Smith and Brady, 1995) is another powerful edge 
detector. Both they can be used to produce edge images. The 
fate is to get an edge image presenting the edges in an enhance 
mode, which is an effective way to detect edges and moreover 
to find the breaklines. 
  
Figure 3. Edge image 
Additionally, image processing is applied in one more stage of 
the framework; in the original images in order to perform a 
radiometric balance between images that are used in image 
matching. Images’ photometric characteristics are a vital 
condition to have the optimum balance so as image matching 
has good chances to succeed. Differences in contrast and 
brightness between images define bad conditions for image 
matching. 
3.2 2D image model reconstruction 
In close range problems, especially in architectural and 
industrial applications, lines are considered to be almost 
horizontal or almost vertical. This definition was assumed in the 
current approach. 
In order to proceed to the reconstruction of 3D model, first is 
requisite to calculate the 2D model in image space. Hough 
Transform (HT) was selected as the tool for the 2D image 
model reconstruction. HT is used to extract geometric shapes 
from an image and still remains powerful tool for detecting 
predefined shapes like lines and ellipses or circles. HT has been 
used for more than three decades in the research field of image 
processing, pattern recognition and computer vision (Duda and 
Hart, 1972; Ballard, 1981; Ballard and Brown, 1982; 
Illingworth and Kittler, 1988). However, in digital close range 
Photogrammetry HT has only rarely been used (Adamos and 
Faig, 1992; Stylianidis and Patias, 2000). 
In the current approach, HT is a line-searching tool for specific 
values of lines' direction. Actually, HT is working in the range 
of angle values near to 0° and near to 90°, i.e. is searching for 
almost horizontal or vertical lines. In real conditions, HT is 
searching every black pixel in the edge image that satisfies 
angle restrictions. Additional constrains, like line-length can be 
set up as well, in order to avoid small line segment tracking 
where commonly are not used in 3D reconstruction. 
The result is like the one presented in Figure 4. In Figure 4a it is 
shown how hough lines are extracted in the image space. The 
2D image model is constructed by the intersection of lines, 
which is clearly presented in Figure 4b. 
À | | l 
  
  
  
  
  
  
  
  
(b) 
Figure 4. (a) Hough lines in image space (b) 2D image model 
3.3 3D approximate model 
Using inverse collinearity equations, XYZ ground coordinates 
are calculated for each corner point of the created 2D image 
model. Therefore, an approximate 3D object model can be 
generated. Besides, using the direct collinearity equations, the 
produced approximate 3D model can be projected to any other 
image. 
In Figure 5 it is shown with the line-frame (4 points) how and 
where the 3D approximate model is located after its projection 
in other image. The projected model in any other image is not 
adjusted in the correct position due to the fact that the Z 
coordinate taken into account is not the real, but an approximate 
one. Moreover, in the same figure, the 4 individual points 
present the correct position of the corresponding approximate 
model points. 
The Z value (depth), which is the only unknown parameter in 
collinearity equations, is assumed to be in the range of a 
minimum (Z,) and a maximum (Z;) value in the object space. 
This condition defines the two extreme pairs of (x,y) values, 
(Xmin:Ymin) 4nd (Xmax-Ymax), i.e. indicates the searching area for 
—218— 
  
023 
Fh ^ e pne N