Multi-Dimensional Features Reduction of Consistency Subset Evaluator on Unsupervised Expectation Maximization Classifier for Imaging Surveillance Application

Chue-Poh Tan, Ka-Sing Lim and Weng-Kin Lai
International Journal of Image Processing, Volume (2) : Issue (1) 18
Multi-Dimensional Features Reduction of Consistency Subset
Evaluator on Unsupervised Expectation Maximization Classifier
for Imaging Surveillance Application
Chue-Poh TAN chue.poh@mimos.my
Centre for Advanced Informatics
MIMOS Berhad
Technology Park Malaysia
57000 Kuala Lumpur, Malaysia.
Ka-Sing LIM kslim@mmu.edu.my
Faculty of Engineering,
Multimedia University,
Persiaran Multimedia,
63100 Cyberjaya, Selangor, Malaysia.
Weng-Kin LAI (Dr.) weng.kin@mimos.my
Centre for Advanced Informatics
MIMOS Berhad
Technology Park Malaysia
57000 Kuala Lumpur, Malaysia.
Abstract
This paper presents the application of multi dimensional feature reduction of
Consistency Subset Evaluator (CSE) and Principal Component Analysis (PCA)
and Unsupervised Expectation Maximization (UEM) classifier for imaging
surveillance system. Recently, research in image processing has raised much
interest in the security surveillance systems community. Weapon detection is one
of the greatest challenges facing by the community recently. In order to
overcome this issue, application of the UEM classifier is performed to focus on
the need of detecting dangerous weapons. However, CSE and PCA are used to
explore the usefulness of each feature and reduce the multi dimensional features
to simplified features with no underlying hidden structure. In this paper, we take
advantage of the simplified features and classifier to categorize images object
with the hope to detect dangerous weapons effectively. In order to validate the
effectiveness of the UEM classifier, several classifiers are used to compare the
overall accuracy of the system with the compliment from the features reduction of
CSE and PCA. These unsupervised classifiers include Farthest First, Density-
based Clustering and k-Means methods. The final outcome of this research
clearly indicates that UEM has the ability in improving the classification accuracy
using the extracted features from the multi-dimensional feature reduction of CSE.
Besides, it is also shown that PCA is able to speed-up the computational time
with the reduced dimensionality of the features compromising the slight decrease
of accuracy.

Keywords: Consistency Subset Evaluator, Principal Component Analysis, Unsupervised Expectation
Maximization, Classification, Imaging surveillance
1. INTRODUCTION
Security surveillance systems are becoming indispensable in scenarios where personal safety
could be jeopardized due to criminal activities [1]. Conventional security surveillance systems
require the constant attention of security personnel, who monitor several locations concurrently
[2,3]. Hence, the advancement in image processing techniques has become an advantage to the
security surveillance systems to improve on the operational activity for monitoring purpose.
Image classification is an essential process in image processing and its major issue lies in
categorizing images with huge number of input features using traditional classification algorithm.
These algorithms tend to produce unstable prediction models with low generalization
performance [4]. To overcome high dimensionality, image classification usually relies on a pre-
processing step, specifically to extract a reduced set of meaningful features from the initial set of
huge number of input features. Recent advances in classification algorithm have produced new
methods that are able to handle more complex problems.
In this paper, we emphasize on the analysis and usage of the multi-dimensional features
reduction on advanced classification method of Unsupervised Expectation Maximization (UEM) to
classify dangerous weapons within an image. In order to validate the effectiveness of the feature
reduction method and classifier, several classifiers such as Farthest First, Density-based
Clustering and k-Means methods are utilized to compare the overall accuracy of the classifiers.
Finally, the study depicts the comparative analysis of different classification techniques with
respect to the robustness of the meaningful extracted features. The classification process
comprised of four steps, which are feature extraction, training, prediction and assessing the
accuracy of the classification. Analysis on the features is done to ensure the robustness and
usefulness of each feature to differentiate classes effectively. The details of the classification will
be discussed in this paper.
This paper is divided into four sections. Section II presents the methodology and the dataset
used in this paper. In this section, the basic concept of Consistency Subset Evaluator (CSE),
Principal Component Analysis (PCA), Expectation Maximization (UEM), Farthest First, Density-
based Clustering and k-Means methods are discussed. Section III describes the results and
discussion for the findings of the classification process using the aforementioned classifiers. The
accuracy assessment with the comparisons between the classifiers is discussed in this section. In
Section IV, we conclude this paper with the suggestion on future work.
2. METHODOLOGY
2.1 Data Description
In this paper, we utilized on a set of data which was available freely in the internet [5] to carry out
some experimental research on the classification. We evaluated the selected algorithms using the
training dataset which contains 13 features (attributes value of the image objects) with their
associate class labels (Human, Bull, Child, Dog, Duck, Knife classes). Besides, 6 test dataset that
contain the same features value of the image objects for each class have been identified. Feature
extraction process was carried out to extract all useful features from 128 binary images (black
and white images) to represent the characteristics of the image object. From the image analysis
and feature extraction, 13 important and useful features of the image object as the attributes of
the dataset were obtained. In this case, the extracted features must be robust enough and RST
(rotation, scale and transition) invariant. A very adaptive feature would be RST-invariant, meaning
that if the image object is rotated, shrunk or enlarge or translated, the value of the feature will not

change. We took the invariance of each feature into consideration and the features comprised of
compactness, elongation, ratio of major axis length and minor axis length, hull ratio, moment,
area ellipse ratio, axis ratio, ratio between area of the bounding box minus area of the blob and
area of the bounding box, ratio between the height and the width of the bounding box, ratio
between the squared perimeter and the area of the blob, roughness, ratio of the area of the blob
and the area of the bounding box and compactness circularity of the blob.
2.2 Multi-dimensional Feature Reduction Methods
Feature reduction process can be viewed as a preprocessing step which removes distracting
variance from a dataset, so that classifiers can perform better. In this paper, we present two multi-
dimensional feature reduction methods, namely Consistency Subset Evaluator (CSE) and
Principal Component Analysis (PCA).
2.2.1 Consistency Subset Evaluator (CSE)
Class consistency has been used as an evaluation metric by several approaches to attribute
subset evaluation [6-8]. Attribute subset evaluation is done to look for combinations of attributes
whose values divide the data into subsets containing a strong single class majority [9]. The
search is in favor of small feature subsets with high class consistency. This consistency subset
evaluator uses the consistency metric presented by H. Liu et al. as shown in Equation (1)
N
MD
yConsistenc
J
i ii
s
∑=
−
−= 0
1
where s is an attribute subset, J is the number is the number of distinct combinations of
attribute values for s , iD is the number of occurrences of the i th attribute value combination,
iM is the cardinality of the majority class for the i th attribute value combination and N is the
total number of instances in the data set [9].
To use the Consistency Subset Evaluator, the dataset needs to be discretized with numeric
attributes using any suitable method such as the method of U. M. Fayyad et al. [10]. The search
method that can be used is the forward selection search which is to produce a list of attributes
[11]. The attributes are then ranked according to their overall contribution to the consistency of
the attribute set.
2.2.2 Principal Component Analysis (PCA)
Principal component analysis (PCA) is one of the most popular multi dimensional features
reduction products derived from the applied linear algebra. PCA is used abundantly because it is
a simple and non-parametric technique of extracting relevant information from complex data sets.
The goal of PCA is to reduce the dimensionality of the data while retaining as much as possible of
the variation in the original dataset.
Suppose 1x , 2x , … Nx are N x 1 vectors.
Step 1: Mean value is calculated with Equation (2).
∑=
=
N
i
i
x
N
x
1
1
Step 2: Each features is used to subtract the mean value, shown in Equation (3).
xxii −=Φ
Step 3: Matrix [ ]NA ΦΦΦ= Κ21 is generated with N x N matrix and covariance matrix
with the same dimension size is computed as Equation (4) [12].
TT
i
N
i
i AA
M
C =ΦΦ= ∑=1
1
(2)
(4)
(1)
(3)

The covariance matrix characterizes the distribution of the data.
Step 4: Eigenvalues is computed:
NC λλλ >>>= Κ21
Step 5: Eigenvector is computed:
NuuuC ,,, 21 Κ=
Since C is symmetric, N
uuu ,,, 21
Κ form a basis, ( )( xx − , can be written as a linear combination
of the eigenvectors):
∑=
=+++=−
N
i
iiNNi
ububububxx
1
221 ...
Step 6: For dimensionality reduction, it keeps only the terms corresponding to the K largest
eigenvalues [13]
∑=
=−
K
i
iiubxx
1
where K << N
The representation of x into the basis Kuuu ,,, 21 Κ is thus












K
i
b
b
b
Κ
2
.
2.3 Classification Methods
The aim is to do comparison of supervised classification methods for classification of the image
object to their known class from the reduced multi-dimensional features dataset. The issue in
identifying the most promising classification method to do pattern classification is still in research.
Therefore, we are interested in predicting the most promising classification method for pattern
classification in terms of the classification accuracy achieved in detecting dangerous weapons.
The algorithms considered in this study are UEM, Farthest First, Density-based Clustering and k-
Means. The methodology for each classifier is presented with basic concept and background.
2.3.1 Unsupervised Expectation Maximization (UEM)
The algorithm is in a model-based methods group which hypothesizes a model for each of the
clusters and finds the best fit of the data to the given model [14]. Expectation Maximization
performs the unsupervised classification or learning based on statistical modeling [15].
A cluster can be represented mathematically by a parametric probability distribution
( ) ( ) ( ) ( )
( )i
kik
ikki
xp
CxpCp
xCpCx
|
| ==∈Ρ
where each object ix is assigned to cluster kC and ( ) ( )( )ikkki xEmNCxp ,| = follows the
normal distribution around mean, km , with expectation, kE [16]. The entire data is a mixture of
these distributions where each individual distribution is typically referred to as a component
distribution which makes use of the finite Gaussian mixture models. So, clustering the data can
be done by using a finite mixture density model of k probability distribution [17].
This algorithm can be used to find the parameter estimates for the probability distribution. It
assigns each object to a cluster according to a weight representing the probability of membership
[16]. Basically the algorithm consists of two main steps which are the Expectation step and the
Maximization step. The Expectation step calculates the probability of cluster membership of
(9)
(8)
(5)
(6)
(7)
(10)

object ix , for each cluster and these probabilities are the expected cluster membership for
object ix . On the other hand, the Maximization step uses the probability estimates to re-estimate
the model parameters. The Expectation step can be interpreted as constructing a local lower-
bound to the posterior distribution, whereas the Maximization step optimizes the bound, thereby
improving the estimate for the unknowns [18]. The parameters found on the Maximization step
are then used to begin another Expectation step, and the process is repeated [19].
2.3.2 Farthest First Classifier
Farthest First is a unique clustering algorithm that combines hierarchical clustering and distance
based clustering. It uses the basic idea of agglomerative hierarchical clustering in combination
with a distance measurement criterion that is similar to the one used by K-Means. Farthest-First
assigns a center to a random point, and then computes the k most distant points [20].
This algorithm works by first select an instance to be a cluster centroid randomly and it will then
compute the distance between each remaining instance and its nearest centroid. The algorithm
decides that the farthest instance away from its closed centroid as a cluster centroid. The process
is repeated until the number of clusters is greater than a predetermined threshold value [21].
2.3.3 Density-based Clustering Classifier
Density based algorithms typically regard clusters as dense regions of objects in the data space
that are separated by regions of low density [22]. The main idea of density-based approach is to
find regions of low and high density. A common way is to divide the high dimensional feature
space into density-based grid units. Units containing relatively high densities are the cluster
centers and the boundaries between clusters fall in the regions of low-density units [23].
This method of clustering also known as a set of density-connected objects that is maximal with
respect to density-reachability [22]. Regions with a high density of points depict the existence of
clusters while regions with a low density of points indicate clusters of noise or clusters of outliers.
For each point of a cluster, the neighbourhood of a given radius has to contain at least a
minimum number of points, which is, the density in the neighbourhood has to exceed some
predefined threshold. This algorithm needs three input parameters, which comprised of the
neighbour list size, the radius that delimitate the neighbourhood area of a point, and the minimum
number of points that must exist in the radius that deliminate the neighborhood area of a point
[24].
2.3.4 K-Means Classifier
K-Means is one of the simplest unsupervised learning algorithms that solve clustering problem. K-
Means algorithm takes the input parameter and partitions a set of n objects into k clusters so that
the resulting intracluster similarity is high but the intercluster similarity is low [25]. Cluster
similarity is measured in regard to the mean value of the object in a cluster which can be viewed
as the centroid of the cluster.
The k-Means algorithm randomly selects k of the objects, each of which initially represents a
cluster mean or center. For each of the remaining objects, an object is assigned to the cluster to
which it is the most similar based on the distance between the object and cluster mean. Then, it
computes the new mean for each cluster and this process iterates until the criterion function
converges [26]. The algorithm works well when the clusters are compact clouds that are rather
well separate from one another. The method is relatively scalable and efficient in processing large
data sets because the computational complexity of the algorithm [27-28].

3. RESULTS AND DISCUSSION
In this study, before any classification is applied on the dataset, CSE and PCA are used to
explore the usefulness of each feature and reduce the multi dimensional features to simplified
features with no underlying hidden structure. The distributions of each feature are drawn and
analyzed statistically. Figure 1 shows the distributions for the features which are discarded after
CSE implementation. These features include ratio of major axis length and minor axis length,
ratio between the squared perimeter and the area of the blob and ratio of the area of the blob and
the area of the bounding box. On the other hand, Figure 2 shows the distributions for the features
which are discarded after PCA implementation and these features comprised of hull ratio, axis
ratio, ratio between area of the bounding box minus area of the blob and area of the bounding
box, ratio of the area of the blob and the area of the bounding box and compactness circularity of
the blob
(a) (b) (c)
FIGURE 1: The distributions of features which are being discarded after CSE implementation (a) ratio of
major axis length and minor axis length, (b) ratio between the squared perimeter and the area of the blob
and (c) ratio of the area of the blob and the area of the bounding box
(a) (b) (c)
(d) (e)
FIGURE 2: The distributions of features which are being discarded after PCA implementation (a) hull ratio,
(b) axis ratio, (c) ratio between area of the bounding box minus area of the blob and area of the bounding
Value of
feature
Number of data
Value of
feature
Number of data
Value of
feature
Number of data
Value of
feature
Number of data
Value of
feature
Number of data
Value of
feature
Number of data
Value of
feature
Number of data
Value of
feature
Number of data
Legend:
X - Human
X - Bull
X - Child
X - Bull
X - Dog
X - Knife

box, (d) ratio of the area of the blob and the area of the bounding box and (e) compactness circularity of the
blob
The unsupervised classification algorithms, including UEM, Farthest First, Density-based
Clustering, and k-Means classifiers are applied to the datasets. In order to validate the impact of
multi dimensional feature reduction methods of CSE and PCA, four types of dataset are utilized ,
namely the original data, data produced after CSE method, data produced after PCA method and
data produced after CSE and PCA methods. The classifiers are analyzed and the accuracy
assessment is as shown in Table 1 with the computational speed (shown in bracket). In this
study, the model with the highest classification accuracy is considered as the best model for
pattern classification of this dataset.
Original data
(13 features)
CSE +
Classifier
PCA +
Classifier
CSE + PCA +
Classifier
Expectation
Maximization
93.33 %
(8.12ms)
95 .83%
(7.18ms)
90.12 %
(6.21)
92.29 %
(4.88ms)
Farthest First
81.88 %
(7.33ms)
83.54 %
(6.09ms)
82.08 %
(5.65ms)
86.25 %
(4.26ms)
Density based
Clusterer
85.21 %
(8.35ms)
88.33%
(7.27ms)
87.71 %
(6.51ms)
80.21%
(4.93ms)
K-Means
86.04 %
(7.45ms)
86.88 %
(6.15ms)
89.38 %
(5.69ms)
81.67%
(4.37ms)
Table 1 : Accuracy Assessment and Computational Speed of Experimental Methods on Different
Datasets
Based on Table 1, we can see that CSE + UEM classifier achieve the highest overall
classification accuracy of all the different datasets. As the dataset we used in this study is quite
small and based on our research, UEM classifier is best applied to small dataset. On the other
hand, the classifiers with features generated from PCA provide slightly less accuracy and
computational speed compared to the classifiers using the predefined number of features. This is
due to the reduced dimensional features offered by PCA which allow only the useful key features
to participate in the classification process.
4. CONCLUSION
The project is aimed to investigate the performance and impact of CSE and PCA on classification
in the aspect of accuracy and computational speed. The potential of each classifier has been
demonstrated and the hybrid method of CSE and UEM has shown a desirable result in detecting
weapons compared to other classifiers. Our future work shall extend this work to multiple type of
images and real-time signal data.
5. REFERENCES
1. A.T. Ali, and E.L. Dagless. “Computer vision for security surveillance and movement control”,
IEE Colloquium on Electronic Images and Image Processing in Security and Forensic Science, pp.
1-7. 1990.
2. A.C.M. Fong. “Web-based intelligent surveillance systems for detection of criminal
activities”, Journal of Computing and Control Engineering, 12(6), pp. 263-270, 2001.

3. Y. T. Chien, Y. S. Huang, S. W. Jeng, Y. H. Tasi, and H. X. Zhao. “A real-time security
surveillance system for personal authentication”. IEEE 37
th
Annual 2003 International
Carnahan Conference on Security Technology 2003, pp. 190-195, 2003.
4. P. Geurts. “Contribution to decision tree induction: bias/ variance tradeoff and time series
classification”. PhD. Thesis, Department of Electrical Engineering and Computer Science,
University of Liege, May 2002.
5. http://www.cs.sdce.edu/ShapeMatcher/.
6. H. Almuallim and T. G. Dietterich. “Learning with many irrelevant features". Proceedings
of the Ninth National Conference on Artificial Intelligence, pp. 547-552, 1991.
7. H. Liu and R. Setiono. “A probabilistic approach to feature selection". Proceedings of the
13
th
International Conference on Machine Learning. pp. 319-327, 1996.
8. M. A. Hall and G. Holmes. “Benchmarking Attribute Selection Techniques for Discrete
Class Data Mining”. IEEE Transactions On Knowledge And Data Engineering, 15(3),
2003.
9. I. Kononenko. “Estimating attributes: Analysis and extensions of relief". Proceedings of
the Seventh European Conference on Machine Learning, pp. 171-182, 1994.
10. U. M. Fayyad and K. B. Irani. “Multi-interval discretisation of continuous-valued
attributes”. Proceedings of the Thirteenth International Joint Conference on Artificial
Intelligence, pp. 1022-1027, 1993.
11. M. A. Aizerman, E. M. Braverman, and L.I. Rozoner. “Theoretical foundations of the
potential function method in pattern recognition learning”. Automation and Remote
Control, 25:826-837, 1964.
12. M. E. Tipping and C. M. Bishop. “Mixtures of probabilistic principal component analyzers”.
Neural Computation, 11(2): 443-482, 1999.
13. M. E. Tipping and C. M. Bishop. “Probabilistic principal component analysis”. Journal of
the Royal Statistical Society, 61(3): 611, 1999.
14. A. Basilevsky. “Statistical Factor Analysis and Related Methods”. Wiley, New York, 1994.
15. J. Han and M. Kamber. “Data Mining: Concepts and Techniques”. Morgan Kaufmann,
San Francisco, CA (2001).
16. I. Borg and P. Groenen. ”Modern Multidimensional Scaling: Theory and Applications”.
Springer (1997).
17. T. F. Cox and M. A. A. Cox. “Multidimensional Scaling”. Chapman and Hall (2001).
18. S. T. Roweis and L. K. Saul. “Nonlinear dimensionality reduction by locally linear
embedding”. Science, 290(2):2323-2326, 2000.
19. F. Dellaert. “The Expectation Maximization Algorithm, College of Computing, Georgia
Institute of Technology”. Technical Report, 2002.
20. S. D.Hochbaum and B. D. Shmoys. “A Best Possible Heuristic for the k-Center Problem”.
Mathematics of Operational Research, 10(2): pp. 180-184,1985.
21. S. Dasgupta and P. M. Long. “Performance guarantees for hierarchical clustering”.
Journal of Computer and System Sciences, 70(4):555-569, 2005.
22. X. Zheng, Z. Cai and Q. Li. “An experimental comparison of three kinds of clustering
algorithms”. IEEE International Conference on Neural Networks and Brain, pp. 767 -771,
2005.
23. M. Rehman and S. A. Mehdi. “Comparison of density-based clustering algorithms”.
Lahore College for Women University, Lahore, Pakistan, University of Management and
Technology, Lahore, Pakistan.
24. M. Ester, H. P. Kriegel, J. Sander and X. Xu, “A density-based algorithm for discovering
clusters in large spatial databases with noise”. The 2nd International Conference on
Knowledge Discovery and Data Mining, Portland, Oregon, USA, 1996.
25. K. Alsabti, S. Ranka and V. Singh. “An efficient k-Means clustering algorithm”. Available
online at http://www.cise.ufl.edu/~ranka/1997.
26. T. H. Cormen, C. E. Leiserson and R. L. Rivest. “Introduction to algorithms”. McGraw-Hill
Book Company, 1990.
27. R. C. Dubes and A. K. Jain. “Algorithms for clustering data”. Prentice Hall (1998).

28. T. Zhang, R. Ramakrishnan and M. Livny. “An efficient Data Clustering Method for very
large databases”. Proceedings of the 1996 ACM Sigmod International Conference on
Management of Data, Montreal, Canada, pp.103-114, 1996.

Multi-Dimensional Features Reduction of Consistency Subset Evaluator on Unsupervised Expectation Maximization Classifier for Imaging Surveillance Application

Recommended

More Related Content

What's hot (17)

Viewers also liked (13)

Similar to Multi-Dimensional Features Reduction of Consistency Subset Evaluator on Unsupervised Expectation Maximization Classifier for Imaging Surveillance Application (20)

Recently uploaded (20)

Multi-Dimensional Features Reduction of Consistency Subset Evaluator on Unsupervised Expectation Maximization Classifier for Imaging Surveillance Application