egion, 
erts to 
ration, 
rea is 
'gions. 
iod of 
egular 
lation 
ndow, 
a key- 
1 order 
ulated 
r point 
eitger, 
ler and 
fection 
ement 
ntation 
es in à 
orstner 
nages. 
ning of 
o Line 
  
Mohammad Saadat Seresht 
Totally, we can use five groups of edge detection methods include: 
Gradient-based methods 
Zero-crossing methods 
Moment-based methods 
Surface fitting methods 
Model Matching methods 
The simplest and most widespread methods are gradient-based methods that detect edges by differentiation 
operators. Some of these operators are includes Roberts, Prewitt, Sobel, isotropical, diametrical or 
statistical operators. Sobel operator is probably the best operator because its sensitivity is proportional to 
distance and orientation of neighbors pixels and also has improvement for smoothness and differentiate 
estimation. Sobel operator has been ingratiated because first gradient is much less sensitivity to noise and it 
can control influence of the noise by varying size of operator window that control smoothness. 
The edges have to be thinned in order to positioning accurately. Some methods of edge thinning are 
includes: 
Skeleton line extraction 
Morphologic operators 
Canny methods 
The output of this step is edges with one pixel wide. 
In edge linking step, a list of sequential edge elements for each edge is acquired that is used for next 
processing. In some of edge detection methods, edge are linked simultaneously e.g. LoG method that 
extract closed curve edges or graph-theory techniques that the low-cost path is derived on edges based on 
graph theory and cost function definition. The techniques of edge detection that are based on threshold, 
usually have some gaps in edges that must be modified with post processing of edge linking. 
A technique of edge linking specially in edge detection techniques based on gradient is using of orientation, 
strength and location of edge elements. Orientation of edge elements can be derived from orientation of 
gradient vector. If Difference of orientation of two sequential pixels is less than a suitable threshold, then it 
is edge element. In order to better and more accurate representation, straight lines have to be fitted on edges 
to object boundaries are completed and geometric processing can be applied easily. The techniques of line 
fitting are includes: 
Polyline composition algorithm 
Polyline decomposition algorithm 
Tolerance based algorithm 
Hop-Along algorithm 
The next step is final cleaning process that a number of these processing are includes: 
Linking of line segments that are coincide on a straight line 
Deletion of irrational edge segments 
Accurate identification of corners and linking of edges 
Grouping of linked edges 
Composition of edges with 3D points 
After cleaning and linking of edges that are based on characteristics of edge such as its orientation, straight 
and location, corners must be identified accurately and over/under-shoot of sides must be cleaned on these 
corners, 
43 3D Lines Creation of Building Edges 
3D straight lines have been attended, because they are an important characteristic of man-made objects 
such as conventional buildings. In order to extraction of these 3D lines, 3D points and 2D lines of each 
region are extracted. 3D-coordinates of key-points are derived by image matching process and 2D lines are 
&Xtracted by feature extraction techniques. After that, 2D line are coincided on two or more 3D points and 
atlast 3D lines are made and then are used for building modeling in reconstruction step. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 795