XXIX-B4, 2012 
> a local adaptive 
window to separate 
. by a sequential line 
feature vectorization. 
1g point and initial 
r, and a predefined 
window) is added on 
n, thinning, and line 
moving the window 
he above operations 
until an endpoint or 
broken or a tracking 
of the line, manual 
tion or move back to 
jential line. tracking 
orized. 
~ Input next point > 
A 
  
re vectorization 
on sliding window 
>parating a specified 
| proposed approach 
g window, and is 
nversion, k-means 
there are numerous 
ve first convert the 
56 grey-scale image 
m. This is due to the 
confusion can be 
cale. Secondly, it is 
e grey-scale image 
en foreground and 
t of its advantage of 
ur data. In the YIQ 
three components: 
[he first component 
le the last two 
By converting an 
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 
RGB image into YIQ format, the grey-scale information can be 
extracted without loss (Chen et al., 2000). 
The conversion from RGB to YIQ (Ford et al., 1998) can be 
expressed as 
y 0:299 0.587 0.114 R 
I|= 0596 -—0.274. —0.322 
Q 0:231] —0.523 0.312 
(I) 
w Q 
where Y component is the grey-scale value converted from 
RGB value. Figure 3a and 3b shows the colour image and the 
converted grey-scale image in a 25 X 25 sliding window of 
Figure la. 
  
(a) (b) 
  
Figure 3. The image in a 25 X 25 sliding window. a Colour 
image. b Gray-scale image. c Segmentation by thresholding 
3.1.2 Adaptive image segmentation: The commonly used 
segmentation approaches to  grey-scale images include 
thresholding, feature space clustering, region growing, etc. 
Segmentation by thresholding is particularly useful for images 
containing objects resting upon contrasting background. But for 
a colour map image with many colour confusions, the converted 
grey-scale image may not have obvious valley in the histogram. 
So it is difficult to separate objects from background by using 
the theoretical optimal threshold, especially for the image with 
low contrast and low Signal-to-Noise Ratio (SNR). Figure 3c is 
the experimental result using Otsu thresholding method (Fu, 
2001) for image segmentation. It can be seen that the lines in 
the window are bordered with noises and adhesions. 
In order to achieve satisfying segmentation effect, this paper 
presents an adaptive image segmentation algorithm. The 
algorithm combines k-means clustering with directional region 
growing to segment the specified linear feature in the sliding 
window. First, k-means clustering is applied to a small 
neighbourhood (5 X 5) in the centre of the sliding window, and 
the pixels within the neighbourhood are divided into two 
regions of object and background. Then, the object region is 
expanded to the whole window by using directional region 
growing.. The reason why 5X 5 neighbourhood is selected is 
that there are just only a single line in that small region, 
considering the line width and the distance between two lines in 
a topographic map image with a resolution of 300dpi. The 
neighbourhood should be correspondingly adjusted with the 
variation of linear feature width and the resolution of map 
image. K-means clustering is applied here because it is one of 
the simplest unsupervised learning algorithms for solving 
clustering problem (Wagstaff et al., 2001). K-means clustering 
and directional region growing are performed automatically in 
the sliding window no matter how the brightness changes, 
therefore it is an adaptive segmentation algorithm. 
K-means clustering 
The process of k-means clustering is as follows. 
Step 1: Choose the pixel with minimum grey-scale as a seed 
within a 5 X 5 neighbourhood in the centre of the sliding 
window. 
Step 2: Find the maximum and minimum grey-scale values in 
the 5 X 5 neighbourhood of the seed pixel, and let them be the 
initial clustering centre c, and c, of the object and 
background region respectively. 
Step 3: For each pixel with grey-scale gj in the neighbourhood, 
calculate q, — ra = e| and d, = lg; = Ca) ‚Id < d,; hen 
the pixel belongs to the object region, otherwise the background 
region. 
Step 4: Calculate the average grey-scales m, and m, of the 
object and background region respectively. If they converge to 
c, and c, , the clustering ends; otherwise, let 7m, and m, be 
the new clustering centre c, and c, , go to Step 3. 
Step 5: Set all the grey-scale values of the pixels belonging to 
the target region within the 5 X 5 neighbourhood to be 1. 
For the grey-scale matrix of Figure 3b, k-means clustering is 
performed in the 5 X 5 neighbourhood of the seed pixel (Its 
grey-scale value is 105). Figure 4 shows the object and 
background region after k-means clustering. 
tar Lis 129 141] 166 
138 110 120[167 187 
156|115 05] 160 190 
145 133 119 114 143 
156 159|i42 110 112 
  
  
  
  
Figure 4. K-means clustering within a 5 X 5 neighbourhood 
Directional region growing 
Based on the k-means clustering result, directional region 
growing is performed in the whole sliding window. Before 
presenting the algorithm, the initial direction given by the 
operator is transformed into eight discrete directions di, d5, ..., 
ds (as shown in Figure Sa), and the four sides of a 5 X 5 
neighbourhood centred at a seed pixel are defined as Li, L5, Ls 
and L; (as shown in Figure 5b). 
  
  
  
  
   
(a) (b) 
Figure 5. a Eight discrete directions. b The four sides of a 5X 5 
neighbourhood 
The directional region growing is described as follows. 
Step 1: Initialization. 
101