Mohammad Saadat Seresht 
  
Minimum size of region (wide and length) 
Area and perimeter of region 
Elongation scale of region 
Number of holes in region (Euler number) 
Shape analysis of region (e.g. based on Fourier descriptors of edges of region) 
Image space filtering includes: 
The number and length of long straight lines 
Histogram of edge elements orientations 
Texture analyze 
Color analyze 
4 FINE RECOGNITION OF BUILDING 
After coarse recognition of building areas and assurance to existence of only one building in each region, 
the more complex step i.e. fine recognition is stared. Unlike to previous step, this step specially adverts to 
pictorial data but height information of DSM can be used to preparation of approximated information, 
improvement of reliability or testing of results. In semi-automatic methods, recognition of building area is 
manually but recognition of building details is automatically. 
Basic tools for fine recognition are image features that are including of key-points, edges, lines and regions. 
Selection of Type and level of feature extraction has right relation with accuracy and method of 
reconstruction. Also straight lines are usually basic characteristic of conventional building with regular 
edges. Fine recognition may be done with other methods such as corner analysis, using of line relation 
graph and/or color and shadow analysis. 
4.1 Key-Point Extraction 
Point detection operators must be applied as soon as after preprocessing. Due to size of operator's window, 
each key point is represented to cluster that the best answer is local maxima. These local maxima key- 
points extract by applying a suitable threshold after deletion of their remained neighbors. Of course in order 
to more improvement in key point detection it is divided to some sub-segment and in each one is calculated 
and applied related threshold. 
Different operators for detection of key points are introduced such as Deriche & Giraudon method for point 
detection, Asada & Brady or Medioni & Yasumoto methods for detection of corner points and Heitger, 
Robbins & Owens, Forstner or Moravec that are suitable detectors for each types of intersection, corner and 
centroid key points. But Forstner and Moravec operators have been come of the best key-point detection 
operators for feature based matching of aerial images. Whiles Forstner Operator has some improvement 
related to Moravec operator because operation of Forstner operator is independent than relative orientation 
of images. This is not a considerable benefit in aerial images because exterior orientation of images ina 
strip is very similar. Also this operator enable to estimate subpixel accuracy in the best way. In Forstner 
method each key-point is described by two attributes that are suitable for feature based matching of images. 
4.2 Line Extraction 
Straight-line extraction is a sensitive step in fine recognition of regular edge buildings. After cleaning of 
extra line segments and connecting of their corners, straight lines of building edges are created and so Line 
extraction processing is done in each region. The steps of processing are flowing: 
Edge detection 
Edge thinning 
Edge linking 
Line fitting to edges 
Unsuccessful line segments cleaning 
Corners connection 
  
794 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
  
Tota 
The 
oper 
stati 
dista 
estin 
can 
The 
incl 
The 
In ec 
proci 
extra 
grap! 
usua. 
À tec 
stren 
gradi 
is ed, 
to ob 
fittin 
The r 
After 
and | 
corne 
43 3 
3D st 
such 
regioi 
extra 
at las