Figure 4. The result from map-guided detection of old roads. 
Recognized road segments are shown in white, and non- 
recognized in black. The subsection to the left shows an area 
The line pattern in the binary image is now decomposed to 
elementary line segments, i.e. line segments with only two 
ends. The end points are labeled in the image, and for each 
elementary line segment a computer object containing a 
parametric description of the segment is created. The 
computer objects contain, amongst other information, the 
general orientation of the line ends, the position of the line, 
and its length. 
The system uses a prototype description of a road line to 
determine "the profile" of the recognition process. Amongst 
others, the prototype contains information about how large 
gaps in binary lines can be, how large gaps in feature lines 
can be (i.e. sections of the road where there are no contrast to 
the surrounding environment), and how curved roads can be. 
When the system has generated a description object for each 
line segment, each segment/description object and its 
neighbours are checked to determine if any of the 
neighbouring line segments possibly can be a part of the 
same road. The distance between end points and the relative 
orientation of the line ends are computed. For each relevant 
pair of neighbours a possibility value for a connection line 
between them is calculated. Each possibility value is a 
product of partial possibility values. E.g., one of the partial 
values is relative orientation between the two line segments: 
where Aa is the orientation difference between the line ends, 
max is the maximal orientation angle difference that can be 
accepted, and Ky is a weight constant. 
The neighbouring line pair which has the highest possibility 
value determines the connection line hypothesis. Each 
hypothesis is tested by trying to recognize a connection line 
in the feature image. The profile analysis algorithm 
mentioned in section 3.1 is used. Small gaps in the line, 
sections where no line structure is present, can be accepted 
according to the value set for the "feature line gaps" 
parameter. If a line structure is detected, a corresponding line 
is generated in the binary image, and the two neighbouring 
lines and their connection line are joined to form one line 
object. 
Before making the proposal for new roads in the map, small 
line segments which could not be connected to other line 
segments are removed. These segments are supposed to be 
352 
where larger parts of the road were not detected due to 
shadows from a steep, forest covered hill. 
"noise". The resulting image can be visually compared with 
the old map and the original image as a background and 
manually edited by the user. The line structures accepted as 
new roads are then converted to the internal raster format, 
and the structures are linked to the old road network. The 
new roads in the resulting map are then converted to vector 
format and integrated with the old vector map. 
4.2 Urban Areas 
Two versions of the thresholded feature image (the line 
interference image) have been used as a binary image: First 
the thresholded smoothed feature image, and secondly the 
thresholded original feature image. By smoothing the feature 
image, noise reduction is achieved before recognition of 
urban areas. By using the thresholded original feature image 
more information is present, leaving smoothing as a 
postoperation to the recognition process. 
The postprocessing analysis is built on knowledge about 
urban area features in satellite images, and how urban areas 
in maps are marked is important. Such features are size, 
smoothness of area contours, inner holes, and road density. 
The algorithm consists of the following steps: 
1. The information from the old map is added to the 
binary image pixel: Pixels in the old map representing old 
urban areas which are not candidate urban pixels in the binary 
image, are set to urban pixels. 
2. The candidate urban areas in the binary image is 
traversed in order to mark the areas and simultanously 
generate a data description object for each isolated area 
(connected structure of pixels). Data description objects are 
generated for the following structures: 
- connected structures of pixels 
- outer border for these connected structures 
- inner holes in the connected structures. 
Each connected structure is a candidate urban area. Features, 
such as the number of pixels,extention (maximum and 
minimum (x,y) coordinates), and position are calculated for 
all three kinds of area objects. 
For the connected structures the number of pixels, Nf, in the 
feature image where more than N1 lines have interfered, is 
calculated. The candidate area description objects are linked