Open Journal of Applied Sciences, 2013, 3, 47-52
Published Online March 2013 (http://www.scirp.org/journal/ojapps)
Automatic Classification for Various Images Collections
Using Two Stages Clustering Method
Wan Hyun Cho1, In Seop Na2, Jun Yong Choi1 , Tae Hoon Lee1
1Department of Statistics, Chonnam National University, Gwagnju, Korea
2School of School of Electronic & Computer Engineering, Chonnam National University, Gwangju, Korea
Received 2012
ABSTRACT
In this paper, we propose an automatic classification for various images collections using two stage clustering method.
Here, we have used global and local image features. First, we review about various types of feature vector that is suit-
able to represent local and global properties of images, and similarity measures that can be represented an affinity be-
tween these images. Second, we consider a clustering method for image collection. Here, we first build a coarser clus-
tering by partitioning various images into several clusters using the flexible Mean shift algorithm and K-mean cluster-
ing algorithm. Second, we construct dense clustering of images collection by optimizing a Gaussian Dirichlet process
mixture model taking initial clusters as given coarser clustering. Finally, we have conducted the comparative experi-
ments between our method and existing methods on various images datasets. Our approach has significant advantage
over existing techniques. Besides integrating temporal and image content information, our approach can cluster auto-
matically photographs without some assumption about number of clusters or requiring a priori information about initial
clusters and it can also generalize better to different image collections.
Keywords: Automatic Classification; Images Collections; Clustering; Mean Shift; Gaussian Dirichlet Processing
Mixture Model
1. Introduction
The general goal in the image clustering is to classify the
different image objects or patterns in such a way that
samples of the same cluster are more similar to one an-
other samples belonging to different clusters. However,
clustering task is a difficult problem due to many as-
sumptions, different contexts and the variety of input
data [1]. In the last few decades, it has been growing in-
terests in developing effective and fast methods for clas-
sify an input image into different clusters. These methods
are mainly divided in two types of clustering algorithms,
such as supervised and unsupervised methods. In the
supervised image clustering algorithms, the researchers
incorporate a priori knowledge, such as the number of
image clusters. Huang et al. [2] propose a hierarchical
classification tree that is generated via supervised learn-
ing, using a training set of images with known class la-
bels. The tree is next used to categorize new images en-
tered into the database. Carson et al. [3] used a naive
Bayes algorithm to learn image categories in a super-
vised learning scheme. The images are represented by a
set of homogeneous regions in color and texture feature
space, based on the “Blob-world” image representation.
Yang et al. [4] propose a new clustering algorithm, re-
ferred to local discriminate models and global integration
(LDMGI), which utilizes both manifold information and
discriminate information for data clustering. They theo-
retically prove that K-means and DisKmeans are both
special cases of LDMGI. They also show that LDMGI is
a type of special clustering algorithm. Thus, they provide
a new perspective to discover and understand the rela-
tionships between K-means (or DisKmeans) and other
spectral clustering algorithms. Sleit et al. [5] propose
Content Based Image retrieval (CBIR) scheme that ex-
tracts color, texture, and shape feature of images. Then,
they group similar images together using K-mean clus-
tering. They use the color histogram, Gabor filters, and
Fourier descriptors for color, texture, and shape features
respectively. The main restriction in supervised image
clustering is that human intervention is required.
On the other hand, unsupervised methods aim at provid-
ing the correct number of image clusters without any a
priori information. Goldberger et al. [6] combine dis-
crete and continuous image models with informa-
tion-theoretic based criteria for unsupervised hierarchical
image-set clustering. The continuous image modeling is
based on mixture of Gaussian densities. The unsuper-
vised image-set clustering is conduct by the information

W. H. CHO ET AL.
bottleneck principle. Krinidis et al. [7] present a unsu-
pervised image clustering approach based on the image
histogram, which is processed by the empirical mode
decomposition (EMD). The Ensemble Empirical Mode
Decomposition (EEMD), which provides noise resistance
and assistance to data analysis, decomposes the image
histogram into a number of Intrinsic Mode Functions
(IMFs). The local maxima of the IMFs summation pro-
vide the desire number of image clusters and a combina-
tion of them is used as a criterion for image clustering.
In this paper, we present an unsupervised clustering me-
thod for large images dataset using two statistical clus-
tering methods based on local and global invariant fea-
tures. First, we think about a various types of feature
vector that is suitable to represent local and global prop-
erties of images, and similarity measures that can be rep-
resented an affinity between these images. Next, we con-
sider a clustering method for image collection. Here, we
first build a coarser clustering by partitioning various
images into several clusters using the Mean shift cluster-
ing and K-means clustering algorithms. Second, we con-
struct dense clustering of images collection by optimiz-
ing a Gaussian Dirichlet process mixture model taking
initial clusters as the derived coarser clustering. Finally,
we have conducted the comparative experiments between
our method and existing methods on various images da-
tasets.
2. Extraction of Global and Local Feature
2.1. Global Image Feature
First, we consider global properties of color image. Here,
we will use three kinds of feature information for clus-
tering of given images. These are color feature, texture
feature and shape feature. Here, we first consider color
feature as color histograms [8]. It indicates the frequency
of occurrences of every color in an image, and can be
defined as a mass function. Our work is based on the
HSV color histogram feature extraction. Second, we will
consider texture feature as Discrete Wavelet Transform
(DWT) [9]. Texture refers to visual patterns with proper-
ties of homogeneity that do not result from the presence
of only a single color such as clouds and water. Texture
features typically consist of contrast, uniformity, coarse-
ness, and density. There are two basic classes of texture
descriptors, namely, statistical model-based and trans-
form-based. The former one explores the gray-level spa-
tial dependence of textures and then extractes some sta-
tistical feature as texture representation. The latter ap-
proach is based on some transform such as DWT. Third;
we consider a shape feature vector as moment invariants
of image. Moment invariants have been frequently used
as features for shape representation of object. They are
computed based on the information provided by both the
shape boundary and its interior region.
2.2. Similarity Estimation
In order to verify the ability of distinguish between glob-
al feature vectors we compute the similarity or dissimi-
larity measures based on exponential the cosine distance
for these two vectors. This measure between two feature
vector i
f and
j
f is defined as

i
C
ij
,
Si,j exp(1)



j
ff
ff  , 1i, .jN
Figure 1. Similarity matrix between global feature vectors
extracted for 110 images.
Figure 1 shows the similarity matrix between color fea-
ture vectors for 10 group images consisting of 11 images
with same colors. In Figure 1, the main diagonal areas
represent the similarities of images with the same color,
but the non-main diagonal areas represent the similarities
of images with different colors. Therefore, the images
with same colors represent by pure black color, otherwise
they represent by white color.
2.3. Local Image Feature
Here, we introduce image representation using the bag of
visual words models based on local features. We have
first various detectors and descriptors describing the im-
age characters that are locally invariant with image rota-
tion, scale transformation and illumination changes. We
then describe a local features histogram that is made
from bag of visual words using numerous local descrip-
tors. A salient region in an image is a connected part of
an image showing a significant and interesting image
property. It is usually determined by the application of a
region of interest detector to the image. If a region de-
48

W. H. CHO ET AL.
tector returns only an exact position within the image, we
also refer to it as interest point detector. The most im-
portant information that ideal region detectors give to us
is the location of features, but other characteristics such
as shape (scale) and orientation of a region of interest
have to deliver additionally. Next, we have also dis-
cussed the interest feature descriptors and their charac-
teristics. A descriptor is a process that takes information
of features and image to produce descriptive information
i.e. features’ description, which are usually presented in
form of features vectors. The descriptions then are used
to match a feature to one in another image. And two im-
portant aspects that a descriptor has to satisfy are dis-
criminative and invariant.
2.4. Local histogram of visual words
For image classification, we use the Bag-of –Visual
Words approach, where images are represented as a his-
togram of visual words. The visual words denote local
features extracted from the images and the vocabulary is
learnt task-specifically from a training database [10-11].
The construction procedure of histogram features of vis-
ual words given from images goes as follows. First, we
extract local feature descriptors form image patches
around feature detectors which are invariant with scale or
rotation change, and apply PCA transformation with
these descriptors to reduce their dimensionality. Second,
to create efficient codebook of visual words, we have
portioned the local descriptor space into several informa-
tion regions using various clustering methods such as
K-means clustering or GMM clustering model. Here, we
can create a bag of visual words by choosing the center
of each cluster as the visual word. Third, a bag of visual
words is used as the codebook to build an image histo-
gram of local features
3. Automatic Images Clustering
We automatically categories cluster images collections
using two-stage clustering method.
Figure 2. Block diagram of automatic images clustering
The first step builds initial coarser clustering by consid-
ering the contents of color images using Mean shift clus-
tering and K-means clustering algorithm. The second
step constructs accurate dense clustering by considering
simultaneously the global and local features of color im-
ages using a Gaussian Dirichlet process mixture model.
Fig. 2 shows the block diagram of two-stage clustering
method using global and local features.
3.1. Coarser Clustering
First, we briefly review the Mean Shift clustering and
K-means clustering algorithms. And then we built initial
coarser clustering initial coarser clustering by consider-
ing the contents of color images using Mean shift clus-
tering and K-means clustering algorithm. First, we brief-
ly review the traditional Mean shift clustering procedure
[12]. This is guaranteed to converge to a point where the
gradient of density function is zero. Here, the key point
of Mean shift clustering procedure is how to take a size
of bandwidth. We have used the flexible bandwidth size
obtained through multiple iterations of the implementa-
tion. Second, we briefly review the K-means clustering
algorithm. The most common algorithm uses an iterative
refinement technique. The algorithm is deemed to have
converged when the assignments no longer change.
Here, we construct initial coarser clustering for experi-
mental image collections using the Mean shift clustering
algorithm with the flexible bandwidth size, and K-means
clustering algorithm based on global and local feature
vectors respectively. Here, global feature consist of color,
texture and shape features. They are the histogram of
RGB color space, histogram of coefficients generated by
2-step discrete wavelet transformation and invariant
moments. Moreover, local feature is used by the histo-
gram of visual words occurrences derived from applying
K-means with SIFT descriptors.
3.2. Dense Clustering
Here, we first introduce a statistical theory of the Gaus-
sian Dirichlet process mixture model. And then we con-
struct accurate dense clustering by considering simulta-
neously the global and local features of color images
using a Gaussian Dirichlet process mixture model. The
Dirichlet process, denoted as , is a random measure on
measures and is parameterized by the innovation pa-
rameter α and a base distribution 0
G [13]. One of the
most important applications of the Dirichlet processes is
as a nonparametric prior distribution of a mixture model.
We want to model this data by mans of nonparametric
Bayesian formulation of the Gaussian Dirichlet process
mixture model. For this purpose, since the number of
mixture component is unknown, we have to consider the
mixture model with countable infinite components.
49

W. H. CHO ET AL.
Therefore, we will use a Diriclet process mixture model
as the prior distribution over the number of components
generating the data, and we also assume the probability
distribution of observations as the multivariate Gaussian
distribution. Here, we assume a coarser clustering model
of given image-sets obtained by Mean shift clustering or
K-means algorithm as an initial clustering model for
Gaussian Dirichlet process mixture model. If we apply
Variational Bayesian inference principle for the Gaussain
Dirichlet process mixture, then we can obtain the ap-
proximating likelihoods and posterior distributions


q

for all model parameters and latent clustering variables.
They are represented with the following formulas. First,
for the posterior distributions over the DP parameters, we
have

kk,1k,2
qv Beta(β,β)
where
N
k,1 n
n1
β1q(zk)

 

KN
k,2 n
lk1 n1
βq(z l) 
 
 

and

12
ˆˆ qαGamma( α|η,η)
where
11
ˆ
ηηK1

K1
22k,2k,1 k,2
k1
ˆ ηη [ψβ ψ(ββ)]


 

and ψ() denotes the digamma function.
Second, regarding the posteriors over the likelihood pa-
rameters, we have





kk k
kkkk
qΘ(, |λ,,ω,)μΛ mΨ
where

kkk
λλ
N



kkk kk
k
1(λN)
λ
mmy

kkk
ωω
N
,

kkkk NΨΨ S

T
kk
nknk
kk
λN
λN

yyyy.
and we use the notation defined as
N
kn
n1
N
q(z k)



N
knn
n1
k
1q(z k)
N


yy

T N
nnknk
n1
k
1q(z k)
N


k
Syyyy
Finally, the posteriors over latent clustering variables
generating the clustering model is given as


nknk
ˆ ˆ qz kπp( |Θ) vy
where
 
k1
klk
l1
ˆ πexp{log1vlog v}



 

v
and
nk
d1
ˆ p( |Θ)exp{ log2πlog
22


k
yΛ

T
nk nk
1()}
2
 
k
yμΛyμ
As a last step, after conducting the updates of all poste-
rior distributions and likelihood parameters at each itera-
tion of the variational Bayesian inference algorithm for
the Gaussain Dirichlet process mixture, the estimates Z
of the latent clustering variables must be also update.
Then, we have obtained the cluster membership of each
image by maximization of posterior distribution
over k. Hence, if each image is assigned to the cluster
with cluster membership
n1kKn
ˆ
zargmax q(zk)


 ,
then we obtained final clustering result for given image
datasets.
4. Experimental results
4.1. Synthetic Image Data
(a) (b)
(c) (d)
Figure 3. Coarser clustering results using means-shift and
K-means based on color histogram with 8 bins and 4 bins
for each channel: (a) Mean shift(8 bins), (b) K-means(8
bins), (c) Mean shift(4 bins), (d) K-means(4 bins)
50

W. H. CHO ET AL.
(a) (b)
(c) (d)
Figure 4. Dense clustering results using means-shift +
VDPGMM and K-means+VDPGMM based on color histo-
gram with 8 bins for each channel: (a) Mean shift +
VDPGMM(8 bins), (b) K-means + VDPGMM(8 bins), (c)
Mean shift+VDPGMM(4 bins), (d) K-means + VDPGMM
4.2. Real image data
We conducted experiments using Matlab on an image
database that is collected from internet public data set
CBIR. Our image database consist of 10 different groups,
namely, Festival, Bus, Horse, Rose, Beach, Dinosaurs,
Food, Scene, Elephant, The Parthenon temple as shown
in Fig. 5. Each group contains 10 similar images. The
spatial resolution of each image is size of 128 × 128 pix-
els. We first test our clustering algorithm using global
features such as color, texture, shape features, K-means
and GDPM clustering method. In the coarser clustering
step, we build initial clusters by considering the global
features of images and K-means clustering method. And
then in the dense clustering step, we construct fine clus-
ters using a Gaussian Dirichlet process mixture model
taking an initial clustering model as K-means clustering
results. Our experimental results show that we can cluster
properly an example image collection into 10 groups.
Our clustering result is shown in Fig 5. Second, we test
our clustering algorithm using local features generated by
SIFT, visual word and local image histogram, K-means
and GDPM clustering method. In the coarser clustering
step, we build initial clusters by considering the local
features of images and K-means clustering method. And
then in the dense clustering step, we construct fine clus-
ters using a Gaussian Dirichlet process mixture model
taking an initial clustering model as K-means clustering
results. Our experimental results show that we can cluster
properly an example image collection into 10 groups.
Our clustering result is shown in Fig 5.
Table 1. Comparative results for global and local
feature.
Feature Precision Recall F-measure
Global 64.0 54.0 55.6
Local 50.1 49.0 43.8
(a) DWT
(b) Hue+Saturation
(c) Hue+Saturation+DWT
51

W. H. CHO ET AL.
(d) KMeans
(e) KMeans+GDPM
Figure 5. Dense and coarser clustering results
5. Conclusion
In this paper, we present methods for clustering images
using Mean Shift algorithm and Gaussian Dirichlet mix-
ture model. Our approach has significant advantage over
existing techniques. Besides integrating temporal and
image content information, our approach can cluster au-
tomatically photographs without some assumption about
number of clusters or requiring a priori information about
initial clusters and it can also generalize better to differ-
ent image collections.
6. Acknowledgements
This work was supported in part by the Korea Research
Foundation Grant by the Korean government (KRF
2012-0002487). And "This research was supported by
the MKE(The Ministry of Knowledge Economy), Korea,
under the 3D and Smart TV Competitiveness Program
support program (NIPA-2012-I2901-12-1031) supervised
by the NIPA(National IT Industry Promotion Agency)
REFERENCES
[1] K. Barnard, P. Duygulu, and D. Forsyth, Clustering art,
Present in Comput. Vis. Pattern Recognition, Dec., 2001,
pp 434-441.
[2] J. Huang, S. R. Kumar, and R. Zabith, An automatic hier-
archical image classification scheme, in ACM Conf. Mul-
timedia, Set., 1998, pp 219-228.
[3] C. Carson, S. Belongie, H. Greenspan, and J. Malik,
Blobworld: Image segmentation using Expecta-
tion-Maximization and Its Application to Image Querying,
IEEE Tran. On PAMI, Vol. 24, No. 8, pp 1026-1038,
2002.
[4] Y. Yang, D. Xu, F. Nie, S. Yan, and Y. Zhung, Image
Clustering Using Local Discriminant Models and Global
Integration, IEEE Tran. On Image Processing, Vol. 19,
No. 10, pp 2761-2773, 2010.
[5] A. Sleit, A. Abugalhoun, M. Qatawneh, M.Ai-Sharief, R.
Al-Jabaly and Ola Karajeh, Image clustering using Color,
Texture and Shape features, KSII Tran. On Inter. And In-
for. Syst., Vol. 5, No. 1, pp 212-227, 2011.
[6] J. Goldberger, S. Gorden, and H. Greenspan, Unsuper-
vided Image-Set Clustering Using an Information Theo-
retic Framework, IEEE Transaction on Image Processing,
Vol. 15, 2006, pp 449-458.
[7] S. Krinidis, M. Krinidis, and V. Chatzis, An Unsuper-
vised Image Clustering Method Based on EEMD Image
Histogram, Jour. Of Infor. Hiding and Multi. Sig. Proc-
essing, Vol. 3, No. 2, pp 152-163, 2012.
[8] N. Sharma. S, P. Rawat. S and J. Singh. S, Efficient CBIR
using Color Histogram Processing, Signal & Image Proc-
essing : An International Journal, Vol. 2, No.1, pp 94-112,
20011.
[9] K. H. Ghazali, M. F. Mansor, M. M. Mustafa and A.
Hussain, Feature Extraction Technique using Discrete
Wavelet Transform for Image Classification, The 5th
Student conference on Research and Development, 11-12
December 2007, Malaysia.
[10] R. J. Lopez-Sastre, T. Tuytelaars, F. J. Aceve-
do-Rodriguez, and S. Maldonado-Bascon, Towards a
more discriminative and semantic visual vocabulary.
Computer Vsion and Image Understanding, Vol. 115, pp
415-425, 2011.
[11] T. Deselaers, L. Pimenidis, and H. Ney, Bag-of-Visual
Words Models for Adult Image Classification and Filter-
ing, 208 IEEE
[12] C. Xiao and M. Liu, Efficient Mean-shift Clustering
Gaussian KD-Tree, Pacific Graphics, Vol. 9, No. 7, pp ,
2010
[13] D. M. Blei and M. I. Jordan, "Variational Inference for
Dirichlet Process Mixtures," Bayesian Analysis, vol. 1, no.
1 pp 121-144, 2006.
52