Você está na página 1de 4

2009 International Conference on Information Management, Innovation Management and Industrial Engineering

Classification for Trademark Image Based on Normal Cloud Model

Ya-Li Qi
The Computer Department
Beijing Institute of Graphic Communication
Beijing, China, 102600
e-mail: qyl@bigc.edu.cn

Abstract—Classification is an important research topic in complexity reduction, identifying interacting attributes, etc.
the field of image data mining. There have been many data In [9] application cloud model to puts forward the theory
classification methods studied, including decision-tree of cloud detecting which uses multi-detectors to evaluate
method, statistical methods, neural networks, rough sets, the uncertain antigen, and will decrease the false positive
etc. This paper proposed a method to classify the image with rate and false negative rate based on classification. By the
normal cloud model which is an uncertainty transformation enlightenment of these research theses, this paper
model between quantities and qualities conception. We proposed a method to classify the image with Cloud
develop the algorithm for classification based on normal model. We design an algorithm to develop normal cloud
cloud model. Finally we perform experiments on an generation based on an artificial database which includes
artificial trademark image database. The results show the train samples and test samples. Then classify the test
advantages of the cloud model in the process of sample to its class. The experiments results are showed in
classification. section4. The rest of the paper is organized as follows.
Keywords-data mining; thecloud model; classification Section2 briefly introduces cloud model. The algorithm to
algorithm; trademark image retrival structure cloud model classifier is showed in Section3.
Finally Section5 concludes the paper based on the
I. INTRODUCTION
experiments results and the future research direction.
Data classification is a well-recognized operation in
data mining and knowledge discovery research field and it II. CLOUD MODEL
has been studied extensively in statistics and machine Deyi Li. define the cloud model following the
learning literature [1]. There are some methods on data important characteristics of linguistic variables and
classification. Quinlan [2] developed a decision-tree terms[1,6]. Normal Cloud is widely used in knowledge
classification method, which is a supervised learning representation. Let U be the set U={u},as the universe of
method that constructs decision trees from a set of discourse, and T a term associated with U. The
examples. Shan.et.al. [3,4] proposed an approach, which membership degree of u in U to the term T, CT(u)∈[0,1],
uses rough sets to ensuring the completeness of the is a random number with a stable tendency. The cloud of
classification and the reliability of the probability estimate T is a mapping from the universe of discourse U to the
prior to rule induction. In [5] proposed statistical unit interval [0,1]. That is:
approaches to classification. Jianhua Fan and Deyi Li CT(u):U→[0,1] (1)
present a new mathematical representation of qualitative u∈U u→CT(u) (2)
concepts—Cloud Models. With the new models, mapping The normal cloud (NC) is based on normal
between quantities and qualities becomes much easier and distributions. It can be characterized by three digital
interchangeable. Based on the cloud models, a novel parameters: A (Ex, En, He). The expected value Ex points
qualitative strategy for data classification in large out the center of gravity of a cloud. The entropy En is a
relational databases is proposed. Then, the algorithms for measure of the fuzziness of the concept over the universe
classification are developed, such as cloud generation, of discourse, showing how many elements could be

978-0-7695-3876-1/09 $26.00
$25.00 © 2009 IEEE 74
DOI 10.1109/ICIII.2009.328

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on July 12,2010 at 08:11:31 UTC from IEEE Xplore. Restrictions apply.
accepted by the term A. The hyper entropy He is a − ( x − Ex ) 2
2 ( E ′N ) 2
measure of the uncertainty of the entropy En. The larger y=e (3)
the value of He, the more random the set of membership take y as the degree of membership;
degrees is distributed. Given the digital characteristics Ex, 4) Let(x, y) be a cloud drop in the universe of discourse;
En and He represent a linguistic term. For a cloud first of 5) Repeat 1-4 until the number of drops required all
all, the mapping from all u in U to the interval [0, 1], is generated.
one-point to multi-point transition, producing a B Normal Cloud Recognition Algorithm
membership cloud, rather than a membership curve. Lots of normal cloud models will be trained, and we
Secondly, any particular drop of the cloud may be paid used these models to classification. It is important to say
little attention to, however, the total shape of the cloud, again the entropy En is a measure of the fuzziness of the
which is visible, elastic, boundless and movable, is most concept over the universe of discourse, showing how
important. Thirdly, the mathematics expected curve many elements could be accepted by the term A, so we
(MEC) of a membership cloud may be considered as its use the En to control the test result. The specific algorithm
membership function from the fuzzy set theory point of showed as following:
view. Finally, the definition has effectively integrated the Normal Cloud classification Algorithm:
fuzziness and randomness of a linguistic term in a unified Input:
way. the expected value of developed cloud Ex, the
entropy of developed cloud En, the hyper entropy of
III. CLOUD MODEL CLASSIFIER
developed cloud He, the awaiting data to classify x’ .
Cloud model curve obey normal distribution which Output:
reflect daily fuzzy concept. And in the cloud model for The degree of membership y’ corresponding the
random x, the degree of membership will change within awaiting data x’.
limits to any concept, which reflects different person has 1) for a random object x’ to classify, we computer y’
different viewpoint for the same thing in the daily life. In according the formula,
this paper we train the normal cloud model for − ( x ′ − Ex ) 2
2 ( E ′N ) 2
classification. y′ = e (4)
A Normal Cloud Generation 2) for every model computer y’;
Given three digital characteristics Ex, En, and He, to 3) order all y’ form large to small;
represent a linguistic term, a set of cloud drops may be 4) the class of larger y’ corresponding is taken for the
generated by the following algorithm: class of x’ residing in.
Normal Cloud Generation Algorithm:
Input: IV. EXPERIMENTS
The expected value of cloud Ex, the entropy of cloud The artificial image database which selected from
En, the hyper entropy of cloud He, the number of drops N. http://image.google.cn and we draw some similar image.
Output: The database includes 200 sheets of image. We use
A normal cloud with digital characteristics Ex, En, Matlab7.0 to receive the Tamura texture feature [7, 8] ,
and He. and for every class image use coarseness, contrast,
1) Produce a value x which satisfies with the normal directionality train the cloud model respectively. For
distribution probability of mean=Ex, and standard example as Figure1, we develop cartoon calf cloud model
error=En, select x aimed to the center of class; based on coarseness of the training
2) Produce a random value En’ which satisfies with the samples.
normal distribution probability of mean=En, and standard
error=He;
3)Calculate

75

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on July 12,2010 at 08:11:31 UTC from IEEE Xplore. Restrictions apply.
which maps between quantities and qualities become
much easier and interchangeable. The aim of our method
is recognize sample belong to which class. For promoting
the classification precision, on the one hand, we train more
class cloud sets for one hand, the other for one class image
use various feature to train various normal cloud model.
The precision is conditioned by the rule or the threshold
value to determine the sample whether belongs to some
class or not. At the same time it is the weakness in defect
of objectivity. In the future we will improve the decision
rule with objectivity.
(a) Cloud model parameter

(b) Normal cloud model


Figure 1. Cloud model based on coarseness (a) A calf image to test with calf cloud model
We can use the train sample gradually modulate the
En. Then test according the front algorithm. For example
we select a calf and a tree image to classification, the
result showed as Figure 2 (a) and (b) respectively. The red
line shows location of the test data in the cloud model.
From the results we can define a rule or a threshold
value which prescripts the sample whether belongs to this
class or not. Because a feature of image shows a aspect of
characteristic, so for one class image we train various
cloud model based on various feature. For a image to
classify can be performed with weighing the result from
various cloud model.

V. CONCLUSION
Data classification is a well-recognized operation in
data mining and knowledge discovery research field and it (b) A tree image to test with calf cloud model
has been studied extensively in statistics and machine Figure 2. Classification results
learning literature. We described a novel approach for
image classification. We have presented Cloud Models

76

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on July 12,2010 at 08:11:31 UTC from IEEE Xplore. Restrictions apply.
ACKNOWLEDGEMENTS
The research is supported by The Science and
Technology Project of Beijing Education Committee
(18000109008).

REFERENCES
[1] Jianhua Fan and Deyi Li, Mining Classification Knowledge Based
on Cloud Models, Proceedings of PAKDD99, Beijing, China,
April.26-28,1999,pp 317-326.
[2] Shan,N.,Ziarko,W.,Hamilton,H.J.,and Cercone, “Discovery
Classification Knowledge in Databases Using Rough Sets”, In
Proceedings of KDD-96:Second International Conference on
Knowledge Discovery&Data Mining,Menlo Park,CA:AAAI
Press,1996.
[3] Shan,N.,Ziarko,W.,Hamilton,H.J.,and Cercone,“Using Rough Sets
as Tools for Knowledge Discovery”.In Proceedings of the First
International Conference on Knowledge Discovery and Data
Mining(KDD-95),Montreal,Canada,1995.
[4] G.Piatetsky-Shapiro,“Discovery,analysis,and presentation of
strong rules”, In G. Piatetsky-Shapiro and
W.J.Frawley,editors,Knowledge Discovery in Databases,
AAAI/MIT Press,1991,p.229-238.
[5] Li Deyi,Shi Xuemei,Meng Haijun.“Membership clouds and clouds
generators ” ,The Research and Development of
Computers,1995,42(8):32-41.
[6] T Deselaers, D Keysers, H Ney. Features for image retrieval: A
quantitative comparison. Lecture Notes in Computer Science,
Springer,2004,Volume 3175/2004, pp.228-236.
[7] H.Tamura, S.Mori, T.Yamawaki. Textural Features Corresponding
to Visual Perception.IEEE Transaction on Systems,Man,and
Cybernetcs,1978,Vol.SMC-8, No.6, pp.460–472.
[8] WANG Si-peng, Detecting Technology Based on Cloud Mode,
Computer Technology and Development, Jun.2006,
Vol.16,pp:187-190.

77

Authorized licensed use limited to: EASWARI COLLEGE OF ENGINEERING. Downloaded on July 12,2010 at 08:11:31 UTC from IEEE Xplore. Restrictions apply.

Você também pode gostar