International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
(2) Automatic line tracking often fails when meeting other 
cartographic features with similar colour. Linear features can 
hardly be tracked in those maps with forest tints and relief 
shadings. 
In order to overcome the above shortcomings, this paper 
presents a new technique for linear feature vectorization 
directly in colour map images. It is accomplished by using 
adaptive image segmentation and sequential line tracking based 
on sliding window. 
2. THE ANALYSIS OF COLOR TOPOGRAPHIC MAP 
IMAGES 
Topographic maps typically use only a few distinct colours to 
represent different cartographic feature layers, for example, 
black cultural features, brown geomorphologic features, blue 
water areas, green vegetation, and so on. Influenced by the 
RGB misalignment in the scanning process and the quality of 
the original map, large numbers of scattered colours and noises 
are generated in a scanned colour map. For instance, when a 
brown colour patch is scanned into a computer, many scattered 
colour pixels such as light brown, yellowish brown, dark brown 
are generated, which do not exist in the original map. In 
addition, cartographic features in different colours are more 
likely to overlap and intersect one another. These factors cause 
the phenomenon that features in the same layer do not have the 
same colour and similar colours do not represent the same 
feature layer, and therefore introduce great difficulty for colour 
segmentation based on colour information. Figure 1 shows a 
part of a colour topographic map with relief shadings and the 
result of segmented contour line layer. It shows that shading 
areas adhere together and a large number of broken lines occur 
in colour segmented layer. Therefore, automatic vectorization 
can not be performed at all in such low quality image. 
  
Figure 1. A part of a colour topographic map with relief 
shadings. a Original image. b The colour segmentation result of 
contour line layer 
3. THE PROPOSED APPROACH 
The objective of our work is to find a way of vectorizing linear 
features directly in original colour map images without colour 
segmentation. How to distinguish linear features adaptively 
from the complicated background and how to track linear 
features rapidly are the key problems to be solved. In a colour 
topographic map image, different regions usually show marked 
differences in colour, brightness and contrast, especially those 
regions with forest tints and relief shadings. So it is difficult to 
distinguish linear features from the background using a global 
method. Considering this, we propose a local adaptive 
segmentation method based on sliding window to separate 
linear features from background, followed by a sequential line 
tracking to vectorize linear features. 
Figure 2 shows the procedure of linear feature vectorization. 
For a specified linear feature, a starting point and initial 
direction are first input by the operator, and a predefined 
rectangle window (which is called sliding window) is added on 
the line. Then, adaptive image segmentation, thinning, and line 
tracking are performed in the window. By moving the window 
continuously along the line and doing the above operations 
iteratively, the line is tracked sequentially until an endpoint or 
an intersection is met. If the tracking is broken or a tracking 
error occurs before arriving at the end of the line, manual 
operation is necessary to cross the intersection or move back to 
the correct position. After that, the sequential line tracking 
continues until the whole line has been vectorized. 
    
Input a starting point and direction 
>» ] 
^ 
Sliding window creation 
Adaptive image segmentation 
| Get current point and direction | Y Input next point / 
A Line tracking A 
Meet endpoint or intersection? 
Yes 
  
  
   
  
  
   
       
  
  
     
    
  
N 
Figure 2. The procedure of linear feature vectorization 
3.1 Adaptive image segmentation based on sliding window 
In this study, image segmentation aims at separating a specified 
lincar feature from colour map image. The proposed approach 
is applied to the aforementioned sliding window, and is 
performed by using colour space conversion, k-means 
clustering and directional region growing. 
3.1.1 Colour space conversion: Since there are numerous 
colours in a scanned colour map image, we first convert the 
colour image in the sliding window into a 256 grey-scale image 
so as to reduce the complexity of the problem. This is due to the 
following considerations: Firstly, colour confusion can be 
improved in the image with limited grey-scale. Secondly, it is 
relatively easy to separate objects from the grey-scale image 
because there is a marked contrast between foreground and 
background. 
YIQ colour space is adopted here on account of its advantage of 
separating grey-scale information from colour data. In the YIQ 
colour space, image data consists of three components: 
luminance (Y), hue (I), and saturation (Q). The first component 
represents grey-scale information, while the last two 
components represent colour information. By converting an 
100 
RG 
ext 
Th 
exp 
wh 
RG 
cor 
Fig