International Journal of Computer Applications (0975 – 8887)

Volume 25– No.5, July 2011

Image Annotations using Machine Learning and Features

of ID3 Algorithm

D.V.N Harish Y. Srinivas K.N.V.S.S.K Rajesh P. Anuradha

M. Tech CSE, Dept of IT, Dept of CSE, Department of CSE,
KIET-Engg. College GITAM-University KIET-Engg College, Gitam University

ABSTRACT algorithm is tested on the constructed animal database. The

With the rapid technological growth, availability of digital results were compared and tabulated.
images in world wide web has increased , since the data
levels available in these databases is also increasing 2. IMAGE SEGMENTATION
enormously, it is very difficult to recognize or retrieve a Image segmentation is the prime dissertate for image analysis,
image of interest instantaneously. Many methodologies have it is a process of converting the heterogeneous data into
been proposed but still a lot of new technologies are to be homogenous data [2]. Segmentation can be done in two ways.
proposed in order to retrieve the images of interest basing on Non-parametric based segmentation and parametric based
the pattern or feature at low cost and with minimum time segmentation. In non parametric based segmentation, the
period. The main disadvantages of the existing methods segmentation process is carried out by using edges, threshold
present in the literature is, the retrieval of images are done based etc., where as in parametric segmentation, the
with mapping of images using one to one mapping strategy, parameters inside the image are obtained and using these
this method is time consuming ,hence to overcome this parameters, the segmentation is carried out. Parametric based
disadvantage an attempt is made in this paper . This paper segmentation is considered to be the efficient when compared
focuses on extracting the image if interest across the web in to non parametric based methods, because it involves
an efficient manner with the pattern and behavior as the key parameters while segmenting [2]. Lot of segmentation
elements. algorithms were proposed in literature, but the main
disadvantage of segmentation is that the segmentation
algorithms differ from application to application [3], another
Key Words disadvantage is that, to segment the image it is needed to
ID3, Image Annotations, Feature set, Machine Learning, initialize the number of clusters, if the number of clusters are
Image segmentation, K-Means algorithm not initialized then the segmentation results differ, hence it is
needed for initialization of the number of image components,
1. INTRODUCTION for this purpose, generally K-Means algorithm is used. The
With the extensive development in technology, in particular to main disadvantage of K-Means algorithm is that if the ‘K’ is
digital imaging and the internet technologies, the user can not initialized properly, then it will lead into spherical clusters
easily acquire a huge dataset of images. On the other hand, and finally results into a single cluster, hence to overcome this
retrieval of an image is very difficult due to the exponential disadvantage the image segments are first initialized by using
growth of image database available in the web. It is therefore a Histogram, this value is passed as the initial estimate for K-
needed to develop new methodologies which help to retrieve Means algorithm.
the image or content from the image collections in an efficient
and elegant manner with minimum time complexity. Hence, 3. K – MEANS ALGORITHM
image annotations play a vital role in this regard. The input The K-Means clustering is a popular approach to segment the
image is first segmented and the features are extracted and image into K-Clusters. The steps to be performed are:
these features are annotated, and after post-processing the
labels are identified. Lot of research is available in literature Step:1. Begin with initial value of
regarding image retrievals and feature vector identification k=Number of clusters.
using Bayesian classifiers, SVM’s, SVD. But, the main Step:2. Select the number of clusters k
disadvantage in this methods is that Bayesian classifiers are with Initial cluster centroids Vi; i=1, 2…k.
not best suited if the dataset is large[1].SVM and SVD’s Partition the input pixels into k clusters by assigning
classify the image on either way for those set of data which lie each pixel xj to the closest Cluster centroid Vi using
near the boundaries, hence in either way the main the selected distance measure, e.g. Euclidean
disadvantage of image retrievals using the above foresaid distance defined as:
methods is not accurate. Hence in this paper we proposed and dij = || xj - Vi ||…… (1)
develop a new methodology for image annotations using ID3 Step:3. Compute a cluster assignment
algorithm. The paper is organized as follows section- 2 deals matrix U representing the partition of the pixels
with a brief discussion about in a segmentation, section-3 with the binary membership value of the jth pixel to
deals with K-Means algorithm, section- 4 deals with ID3 the ith cluster such that:
algorithm, section- 5 deals with machine learning, section-6 U = [µij]……………….. (2)
deals with classification with a comparative study section-7 Where,
deals with proposed algorithm and experimentation, section-8
   
 ………… (3)
deals with testing and results, section-9 deals with conclusions
section-10 deals with future extensions. The developed

45
 International Journal of Computer Applications (0975 – 8887)
Volume 25– No.5, July 2011

(3) The information gained is calculated using

Gain =I (p,n)-E(A)………………………(8)

(4) ID3 algorithm chooses the attribute with maximum Gain (A)
Step:4. Recomputed the centroids using as the root node, which means the attributes with the
The membership values as minimum
E (A).
……… (5)
Step:5. If the cluster centroids or the 6. CLASSIFICATION
assignment Matrix does not change from the To classify the feature set obtained after clustering using the
previous iteration, stop else go-to step 3. k- means algorithm, ID3 algorithm is used, the main purpose
of ID3 algorithm is to find all the possibilities in the decision
4. INDUCTION DECISION TREE (ID3) trees which enables as to classify the training set more
This is a iterative model, whose main purpose is to generate accurately and exactly. The experimentation is carried out by
all possible decision trees that correctly classify the training taking a sample animal database of size 20 with 4 different set
set and to select the simplest of them. The number of such of animals namely Mammals, Birds, Aquatic, Amphibians,
trees is finite but very large, so this approach would only be this data set is given as an input to the K Means Algorithm
feasible for small induction tasks. In order to overcome this and the individual clusters. We then train the dataset with the
disadvantage we need to construct reasonably good decision three different concepts namely color, texture, behavior.
tree without much computation. A subset of the training set
called the window is chosen at random and a decision tree The feature set can be of two types either a local feature set or
formed from it; this tree correctly classifies all objects in the a global feature set, we have designed a questioner for the
window. All other objects in the training set are then color discrimination, texture discrimination and also behavior
classified using the tree. If the tree gives the correct answer discrimination. To classify the color we have considered the
for all these objects then it is correct for the entire training set animal color into three types i.e., black, brown and yellow.
and the process terminates. If not, a selection of the The texture is considered while classifying are three types i.e.,
incorrectly classified objects is added to the window and the Plain, Stripes and Spotted, to identify the Behavior we have
process continues [4] [5] [6] [7] [8] [9]. considered the animal behaviors into two types i.e., carnivores
and herbivores.
5. MACHINE LEARNING Input
Machine learning is mainly used to recognize complex
patterns and make intelligent decisions based on data; the
Pre Feature Annotation Post
difficulty is with regard to the size of the data (training data) Processing extraction Processing
• ID3
that is used. The image is segmented into basic blocks, which • Segmentation • Colors • Machine • Converting
the feature set
classifies according to the type of content, based on color, • Dividing • Textures Learning
into image
texture and behavior. A lot of research is projected towards
image annotation based on (support vector machines) SVM Output
and Bayesian networks.

However the main disadvantage of classifying the data using

SVM is that it classifies the data is marginally and the main Figure 1: Block diagram of the image annotation system
disadvantage of SVM classifier is the data which is nearest to
the line will be classified in either way. Hence, there is every An intuitive answer to this question is to segment the image
possibility of misclassification where as in Bayesian into regions, cluster similar regions and then use the regions
classification, it needs a very complex data set and time taken as a vocabulary hence we obtained a good vocabulary.
to classify the data is marginally high. Hence, in a proposed
method we develop a methodology for image annotation
based on ID3 algorithm the results obtained from this method
7. PROPOSED ALGORITHM
are tabulated.
1. Consider the image database
ID3 algorithm based on the following two
2. Apply the k means algorithm to identify the cluster
suppositions:
3. Obtain the features color, texture and behavior
(1)The class probability that a correct decision tree classify to
4. Apply the ID3 algorithm for classification of images
random example set is consistent with the probability of
basing on the feature set deriving on step 2
positive-example and negative-example in vector space E.
5. The step 4 is continued until all the images are classified,
…… ……….. (6) all the image in the data set using the features.
A is the event, Where p is the positive example i.e., the
possibility of happening and n is the negative example i.e., The results of the experimentation using the above
possibility of not happening algorithm is carried out in the image data set and is
presented in table 1 the output images obtained are
(2) The information entropy needed for a decision making, presented in figure-2 represents the data set figure-3 and
and is calculated as follows figure-4 represents the process of classification is
.. (7) represented in figure-5.

46
 International Journal of Computer Applications (0975 – 8887)
Volume 25
25– No.5, July 2011

Name of the Feature sets Output image in

Sl.No
image Color Texture Behavior database
Carnivores White tiger
Stripes
Herbivores Zebra
Carnivores White cheetha
1 Black Spotted
Herbivores White horse
Carnivores Panther
Plain
Herbivores Bull
Carnivores Long lizard
Stripes
Herbivores Squirrel
Carnivores Anaconda
2 Wild Brown Spotted
Herbivores Indian cow
Carnivores Bear
Plain
Herbivores Australian horse
Carnivores Tiger
Stripes
Herbivores African deer
Carnivores Hyena
3 Yellow Spotted
Herbivores Giraffe
Carnivores Lion
Plain
Herbivores Donkey
Carnivores Owl
Stripes
Herbivores Sparrow
Carnivores Vulture
4 Black Spotted
Herbivores Parrot
Carnivores Eagle
Plain
Herbivores Peacock
Carnivores Vampire bat
Stripes
Herbivores Brown bull
Carnivores African eagle
5 Bird Brown Spotted
Herbivores Brown pelican
Carnivores Indian vultures
Plain
Herbivores Gyrfalcon
Carnivores Australian eagle
Stripes
Herbivores Pigeon
Carnivores American owl
6 Yellow Spotted
Herbivores Quail
Carnivores Bad bird
Plain
Herbivores Woodpecker
Carnivores Lizard
Stripes
Figure-22 representation of Data Set Herbivores Frog
Carnivores Dragon
7 Black Spotted
Herbivores Turtle
Carnivores Snake
Plain
Herbivores Sea turtle
Carnivores Gladiator
Stripes
Herbivores Seal
Carnivores Crocodile
8 Amphibians Brown Spotted
Herbivores Penguin
Carnivores Godzilla
Plain
Herbivores Ortonil
Carnivores Atractus schach
Stripes
Herbivores Plica
Carnivores Iphasa elegans
9 Yellow Spotted
Herbivores Femoralis
Carnivores Cetnophryne geayi
Plain
Herbivores Magnussoni
Carnivores Gaint isopod
Stripes
Herbivores Fish
Carnivores Octopus
10 Black Spotted
Herbivores Gelly Fish
Carnivores Lazydesis
Plain
Herbivores Sea horse
Carnivores Sea elephant
Stripes
Herbivores Sea horse
Carnivores Sea lion
11 Aquatic Brown Spotted
Herbivores Anger fish
Carnivores Brown skate
Plain
Herbivores Skate
Carnivores Killer whale
Stripes
Herbivores Whale
Carnivores Golden fish
12 Yellow Spotted
Herbivores Yellow shark
Carnivores Isopod
Plain
Herbivores Good fish
Table: 1 The classification and differentiation between images
using the image data base after and before applying the
algorithm

Figure-3
Retrieving the images using classification models

47
 International Journal of Computer Applications (0975 – 8887)
Volume 25
25– No.5, July 2011

9. CONCLUSION
A novel algorithm for image annotation using ID3 is
presented since the input image is un supervised data K means
algorithm is used for clustering so that the classification
classi
technique is applied the feature vector in our mythology
consists of both local and global feature sets using IF-THEN
IF
loops to identify the feature set we have used three feature
sets for extraction of image accurately i.e.., using color texture
behavior
ehavior the output obtained from the mythology consists of a
image with the above feature set the performance of the
developed algorithm is evaluated by the image fidelity this go
for further extension is also presented

10. FUTURE EXTENSIONS

We want to enhance ce the development of our algorithm using
the concept of boundary value problem where we want to
have much more effective classification and retrieval using
the technique Boundary Value Problems(BVP)
roblems(BVP) , this is
Figure-4: Showing the difference between two sets of images Another way to produce the annotation of the image is
in behavior boundary value problem in electrical circuits mainly to detect
the boundaries of the circuit, first the pixels should be
The output results are evaluated using image fidelity which calculated for the image next the adjacent pixels should be
consists of two criteria’s, the objective criteria and the connected with a line,, and the adjacent lines should be
objective criteria we have used the subjective criteria where connected to get the shape of the images
the set of input images and the retrieved images are showed to
the viewers the outcome of the result shown is very good
classification rate 11. APPENDIX
Partial Pseudo Code for ID3
If the animal is in jungle
Then go to wild animal ()
If the animal is in air
Then go to bird ()
If the animal in water
Then go to aquatic ()
If the animal is in both water and land
Then go to amphibian ()

Wild animal ()

If image color code is#000000

Then go to Black
If image color code is#660000
Figure-5 classification in diagrammatic manner Then go to Brown
If image color code is#FFFF00
8. TEST RESULTS Then go to Yellow
{
In order to demonstrate our model we have considered a data
Black ( )
set of 72 images consisting of animals, birds, aquatic and
{
amphibians. We have considered the features based on color,
If image is flesh eating
texture and behavior. We have taken into consideration two
Then go to Carnivores
types of behavior classes mainly flesh eating (carnivores) and
If image is Grass eating
grass eating (herbivores). The database is first segmented
Then go to Herbivores
basing on the query of interest i.e., either to retrieve animals,
{
birds, aquatic or amphibians and the k means algorithm is
Carnivores ( )
utilized
zed for this purpose. In order to classify or in order to
{
retrieve the query of interest exactly we have used the ID3
If the image having stripes
algorithm with color, texture and behavior as the patterns for
Then it is a jaguar
the efficient retrieval. The query processed and retrieved
If the image having plain
images are presented in figure-2 and figure-33 the output
Then it is a bear
results from these two figures clearly show that query of
If the image having spotted
interest is retrieve accurately. The efficiency of the algorithm
Then it is a killer whale
is clearly seen from the above figures and our algorithm
}
works more exactly and accurately since we are considering
Herbivores ( )
multiple patterns to retrieve a query of interest.
{

48
 International Journal of Computer Applications (0975 – 8887)
Volume 25– No.5, July 2011

If the image having stripes }

Then it is a Zebra
If the image having plain // End of the Pseudo code
Then it is a buffalo
If the image having spotted 12. REFERENCES
Then it is a goat [1] “Automatic image annotation Based on Decision Tree
} Machine learning “Lixing Jiang, JinHou, Dengsheng
Brown ( ) Zhang
{
If image is flesh eating [2] “Unsupervised image segmentation using finate doubly
Then go to Carnivores truncated Gaussian Mixture Model and Hierarchical
If image is Grass eating clustering ” Srinivas Y Journal of Current Science
Then go to Herbivores Vol.93 No.4 Aug-2007 PP507-514
{
[3] “A review on image segmentation techniques process
Carnivores ( )
pattern recognition” Pal N.R, Pal S. K –1993
{
If the image having stripes [4] “Induction of Decision trees” J.R.QUINLAN Centre for
Then it is a wild cat Advanced Computing Sciences, New South Wales
If the image having plain Institute of Technology, Sydney 2007,Australia
Then it is a dinosaur
If the image having spotted [5] “Some studies in machine learning using the game of
Then it is a killer snake checkers II: Recent progress.”, Samuel, A. (1967). IBM
} J. Research and Development 11. balancing poles
Michie, 1982,
Herbivores ( ) [6] "Experiments on the mechanisation of game-learning 2 -
{ Rule-based learning and the human window."Michie, D.
If the image having stripes (1982). Computer Journal 25
Then it is a cat
If the image having plain [7] "A task-independent experience gathering scheme for a
Then it is a yak problem solver."Quinlan, J.R. (1969). Proceedings of the
If the image having spotted First International Joint Conference on Artificial
Then it is a deer Intelligence. Washington, D.C.: Morgan Kaufmann
} [8] "Learning structural descriptions from
Yellow ( ) examples."Winston, P.H. (1975). In P.H. Winston (Ed.),
{ The psychology of computer vision. McGraw-Hill
If image is flesh eating
Then go to Carnivores [9] “Performance of a reading task by an elementary
If image is Grass eating perceiving and memorizing program," Feigenbaum, E.A.,
Then go to Herbivores & Simon, H.A. (1963). Behavioral Science, 8.
{
[10] H. Almuallim and T. G. Dietterich . Learning with many
Carnivores ( )
irrelevant features. In Proc. 9th Nat. Conf. on Artificial
{
Intelligence, volume 2, pages 547–552, Menlo Park, CA,
If the image having stripes
Then it is a tiger July 1991. AAAI Press.
If the image having plain [11] L. Breiman, J. Friedman, R. Olshen, and C. Stone.
Then it is a lion Classification of Regression Trees. Wadsworth, 1984.
If the image having spotted
Then it is a cheetah [12] W. Buntine and T. Niblett. A further comparison of
}Herbivores ( ){ splitting rules for decision tree induction. Machine
If the image having stripes Learning, 8:75– 85, 1992.
Then it is a yellow zebra [13] U. Fayyad, G. PiatetskyShapiro, and P. Smyth.
If the image having plain Knowledge discovery and data mining: Towards a
Then it is a rhino unifying framework. In Proc. 2nd Intl. Conf. on
If the image having spotted Knowledge Discovery and Data Mining (KDD’96),
Then it is a giraffe pages 82–88, Portland, Oregon, 1996.
}}}}}
Bird() [14] U. M. Fayyad. Branching on attribute values in decision
{ tree generation. In Proc. 1994 AAAI Conf., pages 601–
--- 606, AAAI Press, 1994.
}
[15] U. M. Fayyad, S. G. Djorgovski, and N. Weir.
Aquatic()
Automating the analysis and cataloging of sky surveys.
{
In U. Fayyad, G. PiatetskyShapiro, P. Smyth, and R.
---
Uthurusamy, editors, Advances in Knowledge Discovery
}
and Data Mining, pages 471–493. AAAI/MIT Press,
Amphebian() 1996.
{
---

49