An Improved Color Coherence Vector Method for
CBIR
Xingfeng Chen1，2，3, Xingfa Gu1，2，3, Hua Xu2，3
1. College of Automation, University of Electronic Science and Technology of China，Chengdu
610054, China
2. State key Laboratory of Remote Sensing Science, Institute of Remote Sensing Applications,
CAS, Beijing 100101, China
3. Demonstration Center of Spaceborn Remote Sensing, National Space Administration, Beijing
100101, China
1
chenxingfeng_001@163.com
Abstract: Content-based image retrieval (CBIR) applications provide great convenience
which help people find what they need from the tremendous amount images. CBIR
retrieves images by features including color, shape and texture. Color is the most popular
one among vision features of image. Color Coherence Vector (CCV) is a classical
method of CBIR. It is similar to the color histogram method, and it also considers some
spatial feature, and is proved to be more effective. This paper describes an improved
color coherence vector method, which has more spatial information, and works more
efficiently.
Keywords: Content-based image retrieval (CBIR); Color Coherence Vector (CCV)
1. Introduction
Database management technology, Computer technology and Network technology are developing
fast nowadays. It is possible to get the needed images from the tremendous images databases, at
the same time, the problem is that images retrieval theory and technology lag seriously. In 1970s,
most retrieval systems always used the text-based method. The traditional text-based retrieval
method was already quite mature. For example, the famous image search engine—Yahoo, Google.
But “Seeing something once better than hearing hundred times”, the text is unable to replace the
direct-viewing visual communication method.
Content Based Image Retrieval (CBIR), in brief, means extracting a range of images which is
relevant with the given image from vision database. CBIR produced to 1990s, it retrieves images
by some visual feature, just like color, shape, texture. Usually, human feel more sensitive to the
color feature than to texture and shape. Also, computer used to describe images by RGB form,
color feature extraction can save much time because of computing more easily. So, retrieval
systems using color is most popular. In the early time, global histogram was used most frequently.
But in fact, histogram can’t describe partial characteristics. That means two images with different
vision may have the same histogram. Also, the histogram feature has too many dimensions.
Considering the disadvantages of histogram, Pass [1] and his team have propounded Color
Coherence Vector (CCV) method. Its kernel idea is divided to color histogram into two parts. A
CCV stores the number of coherent versus incoherent pixels with each color. By separating
coherent pixels from incoherent pixels, CCV contains some spatial information. So it provides
finer distinctions than color histograms.
This paper gives an improved CCV method with more spatial information and without much
added computing work. First, the traditional CCV method will be introduced; then, a particular
description of improved CCV method will be given; following, we describe computation of the
distance between two vectors; at last, we’ll analyze the improved CCV and give our evaluation
and prospect.
2. Color Coherence Vector (CCV) method
Before, many CBIR systems retrieve images by color histogram. But color histogram is not
efficient to small changes for images’ comparing. Often, different images may have the same
histogram. It is not hard to find that color histogram is a globe feature, and contains little spatial
information. In order to contain more spatial information, CCV was propounded. CCV divided
every pixel region of color histogram into two parts. Following words give the detailed definition.
We classify the pixels within a given color bucket as either coherent or incoherent. A
coherent pixel is part of a large group of pixels of the same color, while an incoherent pixel is not.
We determine the pixel groups by computing connected components. If a same color group
contains more coherent pixels than a threshold value defined beforehand, it belongs to coherence
pixels. And the rest pixels are incoherent.
For a given discretized color, some of the pixels with that color will be coherent and some
will be incoherent. Let us call the number of coherent pixel of the j’th discretized color  j and
the number of incoherent pixels  j . Clearly, the total number of pixels with that color is
 j   j , so a color histogram would summarize an image as:
1  1 ,  2   2 ,,  n   n 

Instead, for each color we compute the pair  j ,  j

.
which we will call the coherence pair for
the j’th color. The color coherence vector for the image consists of
1 , 1 ,  2 ,  2 ,,  n ,  n  .
Obviously, CCV contains some spatial information by computing the distribution of coherent
and incoherent pixels.
3. Improved Color Coherence Vector
When CCV was propounded, the author emphasized that it contained spatial information. CCV
counts the coherent and incoherent pixels using a two-dimension vector, which describes the color
distribution. So, it is a promotion of color histogram which is a popular and traditional method for
CBIR.
Now, let’s think something about the computing progress of CCV. Due to CCV pays
attention to the “coherent” pixels, we must check and find the connected pixels in the image
matrix. At last, we get α and β, so as the CCV. When we found the connected pixels, inevitably,
the position information of the connected pixels also appeared at the same time. For the j’th
discretized color, we define  j to be the mean of coordinates of the pixels which is in the
maximum connected region in the coherent pixels.  is different from α (the number of coherent
pixels), because it represents max-connected pixels’ position information but α represents number
of coherent pixels. So, we describe the improved CCV as:
1 , 1 ,  1 ,  2 ,  2 ,  2 ,,  n ,  n ,  n 
.
Compared with the traditional CCV, the improved CCV added γ, but it was gained with α, so it
didn’t cost more added computation quantity. The efficiency of CBIR system is not diminished
evidently.
Following, an example of improved CCV will be given. We make a threshold value τ=4, and
give a gradation image with just 3 colors. The digital number matrix of the given image is:
2 1 2 2 1 1
2 2 1 2 1 1
2 1 3 2 1 1
2 2 2 1 1 2
2 2 1 1 2 2
2 2 1 1 2 2
First, we find the connected regions. We define A, B, C …… as which DN=1, 2, 3 …… A1
to be the first connected region which DN=1, so, A2 is the second. Also, C means the only
connected region which DN=3. Following the above rules, the image matrix becomes:
B1 A2 B1 B1
A1
A1
B1 B1 A2 B1
B1 A2 C B1
A1
A1
A1
A1
B1 B1 B1
B1 B1 A1
A1 A1 B 2
A1 B 2 B 2
B1 B1
A1 B 2 B 2
A1
It is easy to get the connected regions table:
Label
A1
A2
B1
B2
C
Color
1
1
2
2
3
Size
12
3
15
5
1
Table 1: connected regions of each color
According to the definition of improved CCV, we then find regions which have more than τ=4
pixels. Obviously, A1, B1, B2 are classified as coherent. A2, C1 are classified as incoherent. Now,
we have got α and β. For color=1, A1 is the maximum connected region in coherent pixels, so
γ1=(4,5); for color=2, B1 is max, so γ2=(3,2); for color=3, no coherent pixels, α=0 and no γ=(0,0).
The improved CCV of this image is:
Color
1
2
3
α
12
20
0
β
3
0
3
γ
(4,5)
(3,2)
(0,0)
Table 2: CCV of the image
It also can be written:
<(12,3,(4,5)), (20,0,(3,2)), (0,3,(0,0))>.
γ gives the mid-point of the max-connected coherent region. So it adds spatial information of the
image. γ don’t need extra computation, it can be got when we computing α.
4. Distance computation for Improved CCV
Improved CCV contains four columns in fact. It costs several steps to get the final distance
between two images.
First, we compute the traditional CCV distance with α and β. The distance formula:
n
D1   ( j  a j ' )  ( j   j ' )
j 1
(1)
This is an efficient and simple distance formula.
Second, we compute the γ distance between two images. Formula as follows:
n
D2   ( j (1)   j ' (1)) 2  ( j (2)   j ' (2)) 2
j 1
(2)
Third, D1 and D2 are different features. And their numbers are far, so we must make
them unitary and weight them before computing the final distance. We assume that there are M
images waiting for retrieval. So, the distance between reference image and image database is the
matrix [ D1 , D2 ] which has the size 2×M, then we do a Gauss normalization on both vector
D1 and D2 . Normalized matrix is [ D1 ’, D2 ’].
Final distance vector:
D  D1 'D 2 '
(3)
5. Experimental result
We have 230 different images to experiment. Before comparing them, a mean-filter processed all
the images in order to eliminate small variations between neighboring pixels. We get the feature
database from the image database through extracting the improved CCV. Fig.1 is the program flow
chart.
First，we extracted both CCV and improved CCV of Fig.2. Its size is 256×256, just one layer.
The former work cost 31.875s and the later cost 31.656s. The added time ratio is less than 0.7%. It
just makes a little impact on efficiency of establishing feature database.
Image database
Reference image
CBIR programs
Mean filter
Mean filter
Feature extraction
Feature database
Feature extraction
Similarity computation
Arrangement
Show the result
Fig.1: Program flow chart
Fig.2: Computing-time test image
Second, let’s check the efficiency of CCV and the improved CCV. CBIR hasn’t an exact
criterion of evaluation. In our experiment, CCV and the improved CCV retrieve images with
different distance formulas, CCV uses (1), and the improved CCV uses (3). It is not easy to
compare the two methods with detailed numbers. Always, we evaluate a retrieval method with
three universal criterions: effectiveness, efficiency, flexibility. Effectiveness gets most attention.
Through 53 experiments, the improved CCV performs better than the traditional CCV. The
improved CCV is in the person of detail and color distribution.
6. Conclusion
We have introduced an improved CCV method for CBIR. Color was investigated by many
researchers. Some improvements for histogram can also be used on CCV, just like dividing the
image into several subimages in order to get more spatial information. The improved CCV
contains more spatial information, so it performs better than the traditional CCV as a color feature.
But it is still deficient to describe an image, CBIR systems work with color, texture, shape and so
on. We are expecting more exact description for an image, most important, we hope an
international criterion for CBIR and its evaluation.
References
[1] Greg Pass and Ramin Zabih, Comparing Images Using Color Coherence Vectors, In ACM International
Conference on Multimedia. Boston, MA, 65-73, 1996
[2] Zhiyong Zeng, Key techniques research of CBIR, Doctoral dissertation of Xidian University, 2006
[3] Gunther Heidemann, Combining spatial and colour information for content based image
retrieval, Computer Vision and Image Understanding 234–270, 94, 2004
[4] WEI Na and GENG Guo-hua, An Overview of Performance Evaluation in Content-based
Image Retrieval, Journal of Image and Graphics, Vol.9, No.11, Nov. 2004
[5] HUANG Cheng and WANG Guoying, A Method of Image Retrieval Based on Color
Coherence Vector, Computer Engineering, Jan. 2006
[6] Mario A. Nascimento, Effective and efficient region-based image retrieval, Journal of Visual
Languages and Computing, 151–179,14, 2003
Biographies
Xingfeng Chen, male, born in 1984, graduate of College of Automation, University of
Electronic Science and Technology of China. His research interests include remote sensing image
processing, database management, and pattern recognition.
Xingfa Gu, male, born in 1962, Deputy and Executive Director of the Institute of Remote
Sensing Application, Chinese Academy of Sciences. He researches on the quantitative RS and the
radiometric calibration for a long period and has published many research papers.
Hua Xu, male, born in 1978, teacher of Institute of Remote Sensing Application, CAS. His
research refers to software design, remote sensing calibration.