See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/3717370

Region growing and region merging image segmentation

Conference Paper · August 1997

DOI: 10.1109/ICDSP.1997.628077 · Source: IEEE Xplore

CITATIONS READS

38 1,618

4 authors, including:

Konstantinos Plataniotis Michalis Zervakis

University of Toronto Technical University of Crete
597 PUBLICATIONS   14,284 CITATIONS    349 PUBLICATIONS   4,304 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

2019 IEEE International Conference on Imaging Systems and Techniques - School of Imaging View project

Deep learning-based Radiomics View project

All content following this page was uploaded by Konstantinos Plataniotis on 06 April 2016.

The user has requested enhancement of the downloaded file.

 REGION GROWING AND REGION MERGING IMAGE SEGMENTATION
N.Ikonomakis, K.N. Plataniotis, M. Zervakis, A.N. Venetsanopoulos

Department of Electrical & Computer Engineering

University of Toronto, Toronto, M5S 3G4, ON, Canada
e-mail: [email protected]

Abstract: Image segmentation is an important first task of any image analysis process. This paper presents a seeded
region growing and merging algorithm that was created to segment grey scale and colour images. The approach starts
with a set of seed pixels and from these grows regions by appending to each seed pixel those neighbouring pixels that
satisfy a certain prcdicate. Small regions of far away values were merged to neighbouring regions while regions of
similar value were also merged. Homogeneity functions are introduced for both grey scale and colour images.

1. INTRODUCTION aggregation. In this approach we start with a set of

Image segmentation refers to partitioning an “seed” pixels and from these grows regions by
image into different regions that are homogeneous or appending to each seed pixel tliosc neighbouring pixels
“similar” in some image characteristics. It is usually that have similar properties, such as grey levcl, texture,
the first task of any image analysis process module and or colour. The process starts by assigning tlie first pixcl
thus, subsequent tasks rely strongly on the quality of of the image under consideration as the first sced pixcl.
segmentation. The quality of segmentation determines This seed pixel would be compared to its 8-connected
the eventual success or failure of the analysis. For this neighbours: eight neighbours of the seed pixel. Any 01
reason, care is taken to improve the probability of a the neighbouring pixels that satisfy a homogeneity
successful segmentation. function would be assigned to the first region and its
There are currently a large number of image pixel value would change to the sced pixel value. This
segmentation techniques that are available. These can neighbour comparison step would be repeated for every
be categorized into three general groups: edge-based new pixel assigned to the first region until the region is
techniques[ 11[2][3][4][ 5 ] , clustering-based completely bounded by tlie edge of the image or by
segmentatioii[4][51, and regidfiebased pixcls that do not satisfy the homogeneity function. The
approach[2][4][5]. These techniques are either based on next seed pixel for the second region would bc
concepts of similarity (edgc-based) or on discontinuity determined by choosing the first unassigned (to the
(clustering-based and region-based) of pixel values. previously grown region) pixel while moving through
Region-based segmentation algorithms include region the image in a right-to-left and bottom-to-top fashion.
growing, merging and splitting/merging techniques. The above mentioned steps for growing a region would
The focus of this work will be on region-based once again be applied until the second region becomes
segmentation. In particular, we will examine a region complete. This process would be repeated until every
growing and two region merging techniques. These pixel i n the image would belong to a region. When grey
region-based segmentation techniques will be examined scale images are considered the homogeneity function
on both grey scale and colour images. I t will be shown used to incorporate a given pixel into a region is the
that these interactive techniques have the advantage of absolute difference between the grey level of that pixel
controlling the degree of segmentation which is related and the grey level of the seed pixcl. In particular, i f the
to the number of regions (i.e. the number of grey levels difference between the grey lcvel of a given pixel and
or colours) that are in an image. This is a useful the sced pixel is less than a threshold T then the pixel
property in every image process module. Some is included into the region specified by tlie seed pixel.
relevant papers include Knapman and Dickson [6], and This can be written as:
Sivewright and Elliot [71 who use interactive I I
G - G , ST,
techniques for tlie segmentation of medical images where G is the grey level of the pixel being tested ancl
G, is the grey level of the seed pixel. The optimum
2. REGION GROWING AND REGION value o f T is problem depcndent and would vary among
MERGING images. This will be further discussed i n the next
The objective of segmentation is to partition section.
an image into regions. This is directly accomplished via Despite colour images being easily represented
region-oriented techniques. Segincntation will bc by thc basic RGB (red, grccn, bluc) colour space it does
attained through a region growing step and two region not match the human visual system very well and thus
merging steps. Region growing is the process of leads to poor segmentation results. The HSI (hue,
grouping neighbouring pixels or a collection of pixels saturation, intensity) colour space is closely related to
of similar properties into larger regions. The most the way in which people describe the perception of
common approach to achieve this is through pixel colour and is thus ideal for image segmentation [2]. Foi-

0-7803-41 37-6/97/$10.0001977 IEEE DSP 97 - 299

 colour images, the homogeneity function used to region be below the threshold T. Since the original
incorporate a given pixel into a region is the Square image has 256 different grey levels, threshold values
Euclidean Distance, varying from 1 to 255 were attempted. Experimental
1 H - H,/ 2+1 S - Sd ’+I I - I,I , (1) analysis revealed that the degree of segmentation
where H,, S,, and I, are the HSI colour values of the depended on the threshold value. A low T would resull
seed pixel and H, S, and I are the HSI values of the in an image that had a large number of regions and a
pixel being tested. To include a pixel into a region this high T would result in the opposite. For the ‘airplane’
distance function would have to be less than a threshold image a T of 50 resulted in 682 regions and a T of 100
T. The HSI representation of a pixel, with the RGB resulted in 136 regions.
’ Region merging by means of noisy region
colour values known, is obtained with the use of the
following equations [2]: removal was attempted with the number of pixels N i n
1 a region being set to 3. When the noisy regions were
I = - (R + G + B) removed in the ‘airplane’ image (T set to 50) the
3 number of regions decreased from 682 to 206. This
3
S=l- [min(R + G + B)] SE [0,1] significant decrease in regions helps increase the
(R+G+B) degree of segmentation. Merging adjacent regions also
-1 1 /2[(R - G ) + ( R - B ) ] decreased the number of regions i n the ‘airplane’ image
H =cos { 112 1 (2) from 206 to 135 with the merging threshold M set to
[(R - G ) 2 + ( R - B)(G - B)] 55. This image is shown in Figure 2.
H=360”-H if (B/ I) > (C/ I)
H=H/360” HE [0,ll Figure 3 displays a monochrome (or intensity)
version of the colour image ‘Claire’ used for
The RGB values of the pixels are normalized in the segmentation. This image is representative of the kinds
range [0, I]. Through experimental analysis, it was of images encountered in important applications, such
found that the powers i n (1) should be variable powers. as video conferencing. The segmentation of such
This will be addressed in the next section. images is important because it reduces the transmission
After segmenting the image into regions, overhead by reducing the number of regions in the
merging adjacent regions with similar properties is the image. Experimental analysis showed that the when
next task. Merging regions allows for further control of equation (1) in Section 11 was used to grow the regions
the number of regions the image will contain. This can in the image poor results were found. In fact, a variety
be accomplished with two procedures. Regions that are of threshold values ranging from 0.01 to 0.5 were used
N pixels or less in size are considered to be regions of and they all produced unsatisfactory images. An
noise. That is, noisy regions are regions of abnormal attempt to understand why an optimal segmented image
grey level or RGB values lar away from the values of was not achieved led to a new algorithm. This new
the surrounding regions. These noisy regions can be algorithm segmented a colour image as if i t were a
removed by joining them to nearby regions. The pixel collection of grey scale images by looking at the hue.
values of these regions can be changed to those of saturation, and intensity of the image individually. That
neighbouring regions. The next merging procedure is. a region growing algorithm was created to segment
involves merging regions that are neighbours and a colour image by looking at the hue, saturation, and
which satisfy the homogeneity function mentioned intensity separately. The results from this algorithm
above with a new merging threshold M. This threshold would determine which of the three HSI variables
value M would be slightly greater than the threshold segments the image better. Through experimental
value T that i s used to grow the regions. Region analysis it was found that the ‘claire’ image was better
merging is used so that regions that are neighbours and segmented when the saturation of the image was
which could not be joined during the region growing considered separately. This analysis is summarized i n
process (due to a slightly low value of T) ‘are merged to Table 1. With this knowledge, (1) was modified to
become one region. In order for two regions to be reflect a greater dependence on saturation:
considered neighbours, at least J pixels from each 2 1 H - H,I ’+I S - SA “+I I - 1J I T, (3)
region’s boundary edge must be 8-connected. A low where a is a variable power. Increasing a fi-om 2
value of J would result in a low degree of segmentation makes the segmentation process more dependent on
and more regions merging, while a high value of J saturation. With increasing a, the saturation difference
would result in the opposite. I S - Sd is nonlinearly decreased making the sum in (3)
smaller and more likely to be less than the threshold
3. EXPERIMENTAL RESULTS value T. Experimental analysis was conducted on
The grey scale image ‘airplane’ used in this varying values of T and a. A summary of the results
experiment i s shown i n Figure 1 , It is a practical image found are shown in Table 2. A relationship, for optimal
encountered in aerial photography. Referring to Section segmentation, was determined for the two variables and
11, the homogeneity function used to include a pixel was found to be:
into a region was that the absolute difference between T = 0.21 CL.
the grey level of the pixel and the seed pixel of the

DSP 97 - 300
This relationship was found by plotting the points in variable HSI power values can all be adjustcd to control
Table 2 and interpreting the graph, as shown in Figure the number of regions that a segmented image is to
6. Figure 4 displays the ‘Claire’ image segmented with contain. This allows for control of the degree of
T set to 0.03, M set to 0.04, and a set to 6. segmentation.

5. REFERENCES

lThre;t:(T)i

0.03
1 H
; I I 1
Segmentation Rating
O(bad) - S(g0od)
&tuition lnteisity
[I] B. Bhanu and S. Lee, Genetic Learning for
Adaptive Image Segmentation, Boston, Kluwer, 1994.
[2] R. C. Gonzalez and R. E. Woods, Digital Image
Processing, Addison-Wesley Publishing, 1992.
[3] J. Morel and S . Solimini, Variational Methods in
Image Segmentation, Boston, Birkhauser, 1995.
[4] K.S. Fu and J.K. Mui. “A survey on image
segmentation,” Pattern Recognition, 13, 3- 16, 198 I .
[5] R. M. Haralick and L. G. Shapiro. “Image
segmentation techniques,” Computer Vision, Graphics,

;
0
4
4
3
2
2
2
and Image Processing, 29, 100-132, 1985.
[6] J. Knapman and W. Dickson. “Heirarchical
probabilistic image segmentation,” Image and Vision
0.50 0 3 2 Computing, 12(7), 447-457, 1994.
0.60 0 3 1 [7] G.J. Sivewright and P.J. Elliot. “Interactive region
0.70 - 0 2 1 and volumn growing for segmenting volumes in MR
and CT images,” Med. Inform., 19(1), 71-80, 1994.
variables individually [8] D. Crevier, “Hue-Based Segmentation of Color
Images,” CCECEKCGEI ‘93, IEEE, 1250-3, 1993.
[9] G. Armano, C. Regazzoni, S. B. Serpico, and G.
Vernazza, “Region Growing and Merging Techniques
for Accurate Image Segmentation,” VIII IASTED Int.
Symposium Model ling, Identification and Control,
(MIC ‘89), Grindlewald, Switzerland, 264-268, Feb.
1989.

LIST OF FIGURES
Figure 1 Original ‘airplane’ image
Figure 2 ’airplane’ image for T=50 and M=55
~
Figure 3 Original ‘Claire’ colour image
Table 2 Values of T vs. a for optimal segmentation Figure 4 ‘Claire’ image for T=0.03, M=O.04, and a=6
Figure 5 ‘Claire’ image for T=0.1, M=0.11 and HSl
powers of 1,3,2 respectively
Although the segmented images of Table 2 are Figure 6 Graph of T vs. a
satisfactory, the segmented image can be improved
further by changing the powers of all three HSl values
along with the threshold value. A different form of (1)
was once again attempted. This time the powers of the
hue, saturation, and intensity differences in (1) were
changed to I , 3, and 2 respectively. The resulting
segmented image is shown in Figure 5 . Arguably, this
image is one of the best segmented images of ‘Claire’.

4. CONCLUSION
The proposed method of image segmentation
via region growing and region merging was shown to
be very effective when an image needed to be
partitioned into different homogencous regions.
Usually, this is the first task necessary in an image
analysis process and therefore needs to be effective.
The segmentation algorithm is interactive. The
growing threshold value T, the merging threshold value
M, the regions of noise size value N, and the three

DSP 97 - 301
 1 :
,*
Figure 2

Figure 3 Figure 4
Plot o f 0 versus 1

1 2 3 4 5 F 7 R 9 IO
2

Figure 5 Figure 6

DSP97 - 302

View publication stats