Monash University Computer Science and Software Engineering 2003
Image Segmentation
Hybrid Method Combining Clustering and Region Merging
By Timothy Liao
Supervisor: Dr Sid Ray
Presentation Outline
Image Segmentation
Clustering Method
Merging Method
Conclusion
Further Research
Image Segmentation
 What is it?
 First step in image analysis
 The partitioning of an image into non-overlapping regions.
 What is it used for?
 Detection of cancerous cells in medical images
 Detection of roads from satellite images
 Many other computer vision based applications
Image Segmentation Methods
Most Image Segmentation techniques
can be placed in the following main
categories.
1.
2.
3.
Characteristic feature thresholding or clustering
(Feature Domain)
Boundary detection (Spatial Domain)
Region growing (Spatial Domain)
Limitations of Image Segmentation
Methods
 Characteristic Thresholding or Clustering
Ineffective by itself in image segmentation because it does not take spatial
information into consideration
 Boundary/Edge Detection
Boundary/Edge detection on noisy, complex images will of the produce
missing edges or extra edges.
 Region Growing
Its not clear what at what point the region growing process should be
terminated, resulting in under or over image segmentation
Motivation of Research
 All methods in each of the three categories have their own limitations.
 By applying two methods hierarchically, it is hoped that the hybrid method
will improve on segmentation.
 In this Project we investigated on combining Clustering and Region
Merging.
Clustering and Region Based
Segmentation
 Clustering
K-means Clustering
ISODATA
Fuzzy K-Means
 Region Based
Region Growing
Split and Merge
K-Means Clustering
 K-means Clustering is Most common method
used in unsupervised clustering.
 Prior knowledge of K is needed.
Algorithm
Select K different grey level values from pixels in an image.
While K mean values != previous k mean value
do
assign each pixel that has the closest grey level value to the k mean value.
work out the new mean values for each k.
end
Clustering of Grey Level Images
Original Image
K=3
K=2
Depending on the application the K value is selected. An
automatic selection of K is desired.
Automatic Determination of K
 The ratio between intra and inter cluster distances was
used as the basis of a cluster validity criterion described
by Ray and Turi to automatically determine the value of
K in colour images.
 A Multiplier function which favours high cluster numbers
within a colour image was introduced.
 The proposed validity criterion is given in the form of:
validity  y 
intra cluster distance
inter cluster distance
Where,
y  c  N (2,1)  1
Liu and Yang’s Validity Criterion
 Research by Liu and Yang had also applied a
multiplier function to their validity criterion to
favour cluster numbers. However they
favoured on low cluster numbers as opposed
to Ray and Trui.
 The validity function:
K
F (I )  K  
i 1
e2
Ni
Intra and Inter Cluster Distances
 Basic
intra
inter
cluster validity criterion.
The criterion was applied to grey level synthetic images, to obtain the
minimum intra
ratio. High cluster number were being detected.
inter
 A favour on low cluster numbers is needed.
 Ray and Turi’s cluster validity criterion
favouring high cluster numbers was
investigated.
 The Liu and Yang’s cluster validity criterion,
favouring low cluster numbers was
investigated.
New Cluster Validity for Grey Level
Images
 A new cluster validity criterion for clustering
 The Criterion:
intra
Validity  ( K  (c  N (2,1)  1)) 
inter
2
 The K 2 is used to favour low number of clusters
using principles from Liu and Yang
 The c  N (2,1)  1 is used to favour high cluster
numbers
Synthetic Test Images
Uniform random noise added
Results
Results using Euclidean Distance
Validity Criterion
Actual Clusters
intra/inter
sqrt(k)*intra/inter
k2 (intra/inter)
Ray and Turi
Proposed
2
15
15
15
2
2
3
11
11
11
3
3
4
20
20
20
4
4
5
20
20
20
5
5
6
20
20
20
6
6
16
20
20
20
16
16
Results using Absolute Distance
Validity Criterion
Actual Clusters
intra/inter
sqrt(k)*intra/inter
k2 (intra/inter)
Ray and Turi
Proposed
2
10
6
10
2
2
3
20
12
20
3
3
4
20
20
20
4
4
5
20
15
20
5
5
6
20
20
20
6
6
16
20
20
20
2
16
Region Merging
 Find Seed values within the image
Seeds
 Merge pixels with similar grey level values
together
Region Merging
Noise Removal
noise
Looking at spatial information to decide
whether to merge a noise with the current
region.
Merging Method
Majority Rule Method
K no. of Counters for each cluster mean value determined from clustering.
Iterate 20 times
For each pixel in the image
Examine its neighbouring pixels with a 3 by 3 mask.
If the neighbouring pixels in teh 3 by 3 mask does not
have the same grey level value as the chosen pixel
Then find the grey level value that exists as the majority of the neighbouring pixels and
assign the chosen pixels to the same value.
Else
Initialise all counters to zero
Locate all neighbouring pixels in the 3 by 3 mask that has the same value as the chosen
pixel
For each neighboring pixel that has the same value as the chosen pixel
apply a 3 by 3 mask to pixel
For each pixel in the mask
Increase the Kth counter by 1 if pixel has the same value as the Kth cluster
mean
Assign the chosen pixel to have the same grey level value as the mean value that has its
counter as the maximum value.
Merging Results
Original
Clustered
Merged
Merging Result
Original
Clustered
Merged
Conclusion
Strength and Weakness of Merging Method
 Strengths
 By applying more iterations to the merging method it does not
over merge the image into one region. It preserves the
regions when they are large enough.
 Able to remove noises from Images that were not able to be
removed using clustering
 Weakness
 Lines tend to get merged into surrounding regions if they are not
thick enough
Further Research
A stronger cluster validity criterion to
automatically detect the number of
clusters.
A region merging technique that does not
discriminate thin lines.
References




















A.K. Jain, M. and P.J.Flynn (1999). Data clustering: A review, ACM Computing Surveys
vol.31(3): 264-323.
Anderberg, M. (1973). Cluster Analysis for Applications, New York: Academic Press.
Boulton, D. and Wallace, C. (1973). An information measure for hierarchic classification, The Computer Journal vol.16(3): 254-261.
C.Hampton, T.Persons, C. W. and Y.Zhang (1998). Survey of image segmentation, Technical Survey .
http://isgwww.cs.unimagdeburg.de/ ~ karsten/segmentation.pdf (Available 7/03).
Davis, D. and D.W.Bouldin (1979). A cluster separation measure, IEEE Trans. On Patter Analysis and Machine Intelligence vol.1(4):
224{227.
Dunn, J. (1973). A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters, J. Cybern. vol.3(3):3257.
Forgy, E. (2000). Cluster analysis of multivariate data: efficiency vs. interpretability of classifications, abstract, Biometrics vol.21: 768{769.
J.T.Tou (1979). Dynoc - a dynamic optimal cluster-seeking technique, International Journal of Computer and Information Sciences vol.8:
541-547.
K. Harris, S.N. Efstratiadis, N. M. and Katsaggelos, K. (1998). Hybrid image segmentation using watersheds and fast region merging.,
IEEE Transactions on image processing vol.7(12): 1684-1698.
K.N.Plataniotis, N. and Venetsanopoulos, A. (2000). Unsupervised seed determination for a region-based colour image segmentation
scheme, Proceedings of International Conference on Signal and image Processing, Digital Signal Image Laboratory, Department of
Electrical and Computer Engineering University of Toronto, Toronto, ON, M5S 3G4, Canada.
K.S.Fu, R. and C.S.G.Lee (1987). Robotics: Control, sensing and intelligence, McGraw-Hill, New York.
Lezoray, O. and Cardot, H. (2003). Hybrid color image segmentation using 2d histogram clustering and region merging, ICISP 03 vol.2:
289-396.
Lim, Y. and Lee, S. (1990). On the color image segmentation algorithm based on the thresholding and fuzzy c-means techniques, Pattern
Recognition vol.23(9): 935{952.
MacQueen, J. (1967). Some methods for classification and analysis of multivariate observations, Proceedings of Fifth Berkeley
symposium on Mathematical Statistics and Probability, pp. 281-279.
Nevatia, R. (1986). Image segmentation, Handbook of Pattern Recognition and Image Processing pp. 215-231.
N.R.Pal and S.K.Pal (1993). A review on image segmentation techniques, Pattern Recognition vol.26(9): 1277-1290.
Oliver, J. and D.Hand (1994). Introduction to minimum encoding inference, Technical Report No.94/205 .
Palus, H. and Bereska, D. (1999). Region-based colour image segmentation, Silesian University of Technology, Department of Automatic
Control.
R C. Gonzalez, R. E. W. (1992). Digital Image Processing, Addison-Wesley.
References










R.Bhowmik (2003). Clustering algorithms for spatial data mining, in G. M. N. Debnath and
D. Riesco (eds), Proc International Conference on Computer Science, Software Engineering, Information Technology, E-Business and
Applications (CSITEA03), pp. 295-300.
Siebert, A. (1997). Dynamic Region Growing, Department of Computer Science, University of British Columbia.
T.Gevers and Smeulders, A. (1997). Combining Region Splitting and Edge Detection through Guided Delaunay Image Subdivision,
Faculty of Mathematics Computer Science, University of Amsterdam.
Tou, J. and Gonzalez, R. (1974). Pattern Recognition Principles, Addison-Wesley., Massachusetts, USA.
Trivedi, M. and Bezdek, J. (1986). Low-level segmentation of aerial images with fuzzy clustering, IEEE Trans. Syst., Man, Cybern
vol.16(4): 589-589.
Turi, R. and Ray, S. (1998). K-means clustering for colour image segmentation with automatic detection of k, Proceedings of International
Conference on Signal and image Processing, Las Vegas, Nevada, USA, pp. 345{349.
Wallace, C. (1998). Intrinsic classification of spatially-correlated data, The Computer Journal vol.41(8): 602-611.
Wallace, C. and Boulton, D. (1968). An information measure for classification, The Computer Journal vol.11: 185-194.
Wallace, C. and Dowe, D. (1994). Intrinsic classification by mml-the snob program, In proceedings of the 7th Australian Joint Conference
on Artificial Intelligence, UNE, Armidale, Australia, pp. 37-44.