Hidden Markov Model Classification And Gabor Filter Preprocessing

Jurnal Ilmiah Komputer dan Informatika(KOMPUTA)
1
Edisi...Volume..., Bulan 20..ISSN :2089-9033
IMPLEMENTATION OF HIDDEN MARKOV MODEL AND GABOR
FILTER TO DETECT PUBLIC VEHICLE VIOLATIONS
Kukuh Setiawan
Teknik Informatika – Universitas Komputer Indonesia
Jl. Dipatiukur 112-114 Bandung
Email : kukusetiawan@email.unikom.ac.id
ABSTRACT
The number of traffic violations from year to
year has increased. It is known from analysis and
evaluation data of traffic offences there is a rise in
2015, traffic violations committed riders if compared
with the period in 2014. Then the relation to this
research, that recognize the type of offence by way
of a Hidden Markov Model as classification methods
and Gabor Filter in image processing based on the
problems presented. Things to note in knowing
violation, infringement upon a signpost to classify
the image around the vehicle. And within the
boundaries of the problem observed research focus
to signs "don’t enter" and "don’t turn back". So the
offense that was the Foundation of the research
using methods to be able to generate conclusions
this type of offence. When encountered with such a
pattern of that signs, then direction violations can be
detected from the classification of the extraction of
the characteristics at the time of image preprocessing
stage. Then the methods used to detect violations of
the image Hidden Markov Model in the process of
classification and image of Gabor Filters as one way
the preprocessing image. By implementing these
methods are intended to be able to meet the need to
detect the violations committed by the vehicle. The
results of this research in the form of data
classification training which is used to detect data
testing, and produce an output type of offence
intended. And testing of K-Fold Cross Validation
using data obtained 82, average accuracy value
70.31% correct classification and 29.69% invalid.
Keywords
:
Gabor filter, Hidden Markov Models,
image classification, image processing,
violation of vehicle.
1. INTRODUCTION
Traffic regulations made in the public interest with
the goal of creating order in the drive, because the
rules are obeyed, it will create a comfortable
atmosphere while using traffic facilities. But still,
there are those who commit the offense, and the
amount of vehicle traffic violations from year to
year increase. It was known from the data analysis
and evaluation of traffic violations in 2015, there is a
rise in the number of traffic violations committed
rider when compared with the period in 2014. [1]
And in the city of Bandung in one week alone there
were an average of 1200 docket last violation cross
on trial. [2]
In terms of this study, which detects the image
processing vehicle violations that will be a study to
determine the types of violations that occurbased on
the issues presented. Things to consider in knowing
violation of an object is a vehicle and symbol signs
in the vicinity. Violations that will be used as a
research method to produce conclusions of
violations. When encountered signs with patterns
such as banned "Don’t Enter" or forbidden "turning
toward" the offense can be detected using pattern
recognition. The method used to detect the image of
this offense, namely Hidden Markov Model as
image classification and Gabor filter as one image
preprocessing, by implementing these methods are
intended to be able to meet the needs of detected
offenses committed by vehicles.
Regarding the Hidden Markov Model (HMM) is a
method that can determine patterns of violations of
the vehicle based on probability, it can be analyzed
that the vehicle has a traffic violation. Or sequence
feature extraction matrix formed image
preprocessing results will be stored in the database
as a value classification and HMM working hides
another image that does not fit in the database image
edge detection value of training with a given state,
so it will detect the violation in question. [3]
Then explained that the Hidden Markov Model
(HMM) is a probability model that describes the
statistical relationship between the sequence of
observations and order state which is not observed
"hidden" or hidden. [4] In HMM, the state is not
seen directly, but its output is dependent on the
circumstances the look is in this case a matrix of
extraction. And to make image classification, the
need for preprocessing stage beforehand that lets me
switch condition where if the original image data or
noise can still be done testing. And in previous
studies on the noise detection code such as the

2
signature pattern recognition HMM only has an
average of 43.35% CER with the explanation that
the smaller the CER (Character Error Rate), the
better the resulting accuracy, while the introduction
of handwritten letters and numbers achieve an
accuracy of 74%. That means the introduction of the
resulting better than the preprocessing. [4] In the end
to assist in refining the HMM weakness then the
need for other methods mangkus in the image
preprocessing stage before image classification.
Concerning Gabor filter is a method with object
recognition which can be defined as the process of
determining the identification of the object. This
method can detect any dots produced with high
accuracy [5] [6] The image before entering them
into a database, the image will first be processed
using the method of preprocessing and one of them
is Gabor FIlter. Gabor convolution calculation filter
will produce a specific value called Gabor jet
responses to achieve thresholding.
It is necessary both methods with different
functionalities that Gabor filter to enhance the
preprocessing stage and HMM to generate
classification.
1.1 Gabor Filter
As the filter is used Gabor 2D filter kernel obtained
by modulating a sine wave at a frequency 2D and
specific orientation with Gaussian envelope. Basic
equation Gabor filter kernel function 2D equation
shown in the following equation:
(1)
Then x and y is the standard deviation of the
Gaussian envelope dimensions x and y.  and n is
the wavelength and orientation of the 2-D sine wave.
The deployment of the Gaussian envelope is defined
in the form of a sine wave .. Rotation of x - y by
angle produce n Gabor filter at orientation. Here is
an equation calculating the standard deviation:
(2)
The default parameters used to calculate Gabor
filters can be found in the documentation accord
library.[12]
Orientation ͠θn = 0.6
Gaussian Variance Sigma = 2
Aspect Ratio Gamma = 0.3
Kernel Size fx=3
Wave length  = 4
As in calculating the final result can be performed
on Gabor value equation formula (3):
(3)
If all Gabor filters with a variation of the wavelength
() and orientation (n) is applied at a certain point
(x, y), then obtained a lot of response filter to the
point, for example: use four wavelengths ( = 3, 5,
7, 10) and eight orientation, it will produce thirty-
two response filter for each image point that
dikonvolusikan the filters.
1.2 Hidden Markov Model
Hidden Markov Model or betterknown as the
Hidden Markov Model (HMM) is a statistical model
of a systemthat is assumed to be a process of
Markov with a parameter that is not known, and the
challenge is to determine the parameters of the
hidden (state)of parameters that can be observed (
observer). The parameters set can then be used for
further analysis,for example for pattern recognition
applications (Pattern Recognition). A HMM can be
considered as a dynamic Bayesian Network
simplest.
In general Markov models (Vanilla / Visible Markov
Model), its state can be observed directly, therefore
the probability of the state transition (state) being the
only parameter. Inside the hidden Markov model, its
state can’t be observed directly, but can be observed
are the variables that are affected by the state. Each
state has a probability distribution over the tokens
output that may arise. Therefore, a series of tokens
generated by HMM gives some information about
the sequence of state-state. [4]
Figure 1 HMM Model

3
1.2.1 Markov Chain
Markov chains (Markov Chains) is a mathematical
technique that is commonly used to perform
modeling (modeling) a wide variety of business
systems and processes. This technique can be used
to estimate changes in the future in a dynamic
variables on the basis of the changes of the dynamic
variables in time plainly. This technique can be used
also for analyzing events in the future times
mathematically.
Markov Chain Model (Markov Chains) discovered
by a Russian named Andrey Markov in 1906,
namely: "For every time t, when the event is Kt and
all previous events are 𝐾𝑡(𝑗)... , 𝐾𝑡(𝑗−𝑛) that occurs
from a process is known, the probability of the
whole incident to come Kt (j) only depends on the
incidence of 𝐾𝑡(𝑗−1) and does not rely on previous
events that 𝐾𝑡(𝑗−2), 𝐾𝑡(𝑗−3),..., 𝐾𝑡(𝑗−𝑛).”.
Markov chains (Markov Chains) when given the
input state of the current state, the state will come is
unpredictable and can be separated fromthe state in
the past. That is, a description of the current state
captures all the information that affects the evolution
of a system in the future. In other words, the future
condition addressed by using probability.
That is about markov chain an then movements of
variables in future able to predictable according
variables movements at the past. K_t4 influenced by
events K_t3, K_t3 K_t2 influenced by events and so
on which this change occurred because of the role of
the transition probability (transition probability).
Genesis K_t2 for example, will not affect the
incidence of K_t4.
Figure 2 Markov Chains
Markov Chains helpful to calculate the probability
of an event is observed that in general can be
formulated as follows:
𝑃( 𝜎𝑖
) = 𝑃(𝜎𝑡|𝜎𝑡−1, 𝜎𝑡−2, 𝜎𝑡−3, … , 𝜎𝑡−𝑛) (4)
σ_t is the current state, and σ_t is the condition at a
particular time associated with σ_t. While σ_ (t-i) is
the condition before σ_t. Then it can be assumed that
the right of the equation is invariant, ie,
hypothesized in entire of system, the transition
between the specific circumstances remained the
same in its probabilistic relationships. Based on
these assumptions, it can be the formation of a stet
state transition probability between two states K_i
and t_j:
Transition probability =
𝑃( 𝜎𝑡 = 𝐾𝑡| 𝜎𝑡−1 = 𝑡𝑗 ), 1 ≤ 𝑖, 𝑗 ≤ 𝑁 (5)
1.2.2 Parameter HMM
HMM mempunyai parameter-parameter distribusi
sebagai berikut :
a. Probabilitas Transisi (A)
Parameter A merupakan parameter dengan ukuran
MxM dengan M adalah jumlah state yang ada,
parameter transisi dapat dituliskan dalam bentuk
matriks seperti berikut:
𝐴𝑖𝑗 =
⌈
⌈
⌈
⌈
𝐴11 𝐴12 𝐴13 ⋯ 𝐴1𝑛
𝐴21 𝐴22 𝐴23 ⋯ 𝐴2𝑛
𝐴31
⋮
𝐴 𝑛1
𝐴32
⋮
𝐴 𝑛2
𝐴33
⋮
𝐴 𝑛3
⋯
⋱
⋯
𝐴3𝑛
⋮
𝐴 𝑛𝑛 ⌉
⌉
⌉
⌉
𝐴 = { 𝑎𝑖𝑗 } , 𝑎𝑖𝑗 = 𝑃𝑟 ( 𝑥 𝑡+1 = 𝑞 𝑗
| 𝑥 𝑡 = 𝑞𝑖), 1 ≤
𝑗, 𝑖 ≤ 𝑁 (6)
b. Probability Of Observation (B)
Parameter B is an observation probability or
likelihood state is the appearance of a state existing
rows of the entire state. Parameter B in HMM
written in matrix form column with MX1, where M
is the sum of all existing state. Parameter B can be
written in matrix form as follows:
𝐵 =
⌈
⌈
⌈
⌈
𝑏1
𝑏2
𝑏3
⋮
𝑏𝑛⌉
⌉
⌉
⌉
𝐵 = { 𝑏𝑖
}, 𝑏𝑖
( 𝑘) = 𝑃𝑟
( 𝑂𝑡 = 𝑉𝑘
| 𝑥 𝑡 =𝑞𝑖 𝑡)
(7)
c. Initial State Distribution (𝜋)
Parameter pi (π), referred to as the initial parameters,
the probability of occurrence of a state in the
beginning. Similarly, the B parameter, parameter π
is also written in the form of a column matrix with
the size of the MX1, where M is the number of their
state, parameter π can be written in the following
form:
𝜋 =
⌈
⌈
⌈
𝜋1
𝜋2
𝜋3
⋮
𝜋𝑛⌉
⌉
⌉
𝜋 = { 𝜋𝑖
}, 𝜋𝑖 = 𝑃𝑟 (𝑥0 = 𝑞𝑖) (8)
While certain parameters HMM there are two,
namely N and M:

4
N, the number of states the model. Notated finite set
of possible states is Q = {q_1, ..., q_N} (9)
M, the number of observation symbols / state,
discrete font size. Symbol observation relates to the
physical output of the modeled system. Notated
limited set of observations is probably the
V = {V_1, ..., V_M} (10)
HMM parameters are estimated based on the criteria
of maximum likelihood (ML) and the Baum-Welch
algorithm (EM = Expectation Modification).
1.2.3 HMM Problems
There are 3 basic human rights issue that must be
solved for the model applied :
a. calculating 𝑃 = ( 𝑂| 𝜆) 𝑂 = 𝑂1, 𝑂2, … , 𝑂 𝑇 dan
λ = (A, B, π). (10)
Solution :
How common is commonly used to examine every
possible state sequence along T N (number of
observations). This is not possible because the
calculation is less efficient. There are other
procedures more simple and efficient is the use of
forward procedure.
- forward procedure
forward αt(i) at ke-t and state ke-i define with. αt (i)
= P (O1,O2,...,OT, qt=i | λ). Functions can be solved
forward opportunities for the N T symbol
observation state and inductively with the following
steps:
o Initialitation, 𝛼 𝑡
(𝑖) =
πibi(O1), 1≤ i ≤ N (11)
o Induction, αt+1 (j) = [∑ αt
(i)αij
N
i=1 ]
bj (Oi+1),
1≤ t ≤ T-1; 1 ≤ j ≤N
(12)
o Termination, P(O|λ)=
∑ 𝛼 T(𝑖)𝑁
𝑖=1 (13)
The calculation of odds forward based on the pattern
of trellis diagram. There are N points each time slot
on the pattern, all possible row combined state as a
point regardless of length N rows of observation. At
the time t = 1, calculated the value of O1 (i), 1≤i≤N.
At time t = 2,3, ..., T is only required calculation of
the value αt (j) where 1 ≤j≤ N.
Each calculation requires as much as the previous
value of αt N-1 (i) for each N point can only be
connected to the N points on the previous time slot.
Figure 3 Serial of State
2. Research
2.1 Overview
The main objective of this thesis is to create
applications that can perform the introduction of a
traffic violation is prohibited "Don’t Enter" and
prohibited "reverse direction" in the vehicle. The
process flow diagram modeling applications on the
following picture:
Figure 4 Flowchart of Application

5
In the modeling workflow application are described
there are 4 (four) stages, namely the input image,
preprocessing, feature extraction and classification.
From each of these stages can be analyzed process is
further from the input data in the form of the image
of the violation, then into the preprocessing is to
convert the image to grayscale, decisive edge
detection Canny, filtering using Gabor filter and
resize result of Gabor into 128x128. After that, look
for the feature extraction to be included in the
classification using HMM.
2.2 Input
Data input will be taken on this application is an
image that is taken when a violation is prohibited
"Don’t Enter" and prohibited "reverse direction". So
clearly the image of an offense, and in this study
determines the identification of categories of
offenses contained in the image.
Each process on image processing using pixel values
of the image for calculation. Each pixel in the image
is composed of three elements of basic colors
namely Red (R), Green (G) and Blue (B), which is
often called RGB, and each element has a value
between 0-255 which became the code color depth
of each of the elements , For example, the
calculation of the processes of image processing,
image used "vehicle violations" to do with the size
varying input. The image of a violation can be seen
as in the following picture:
Figure 5 Image Infringement prohibited "reverse
direction"
Figure 6 Violations Prohibited "don’t enter"
Of the input image size is different because of the
dimensions of each image acquisition can be
performed with a variety of tools that notabennya
not have the same pixel size, it is necessary to resize
the preprocessing to match the size of the image
pixel dimensions. In this research will be conducted
resize be 128x128 pixels.
2.3 Testing
Application testing is a stage that has the purpose to
find flaws in the software being tested. The test aims
to determine the software that made already meet the
criteria in accordance with the purpose of designing
and testing this software using black box testing.
Black box testing focuses on functional
requirements of the software, and also to the
performance testing conducted to test the
classification menghasil value.
Black box testing is used to determine the testing
performed is divided into several scenarios. As seen
below.
2.3.1 Scenario of Testing
Here is a test scenario as shown in the table below:
2.3.2 Fungsional Scenario
Functional Scenario describes the functional testing
of applications used. Here are the results of testing
on the application:
Testing
Class
Process Type
Images
Test input of
image
Blackbox
Training
data
Manage the data
from image of
train
Blackbox

6
Fungtion
al
activity Testing Conclusio
n
Grayscale Change
image to
grayscal
e
[√ ] accept
[ ] reject
Canny
Edge
Detection
Change
grayscal
e to
canny
edge
detection
.
[√ ] accept
[ ] reject
Gabor
Filter
Canny to
gabor
filter.
[√ ] accept
[ ] reject
Resize Rezing
to be
128x128
pixel.
[√ ] accept
[ ] reject
2.3.3 Cases and Test Results
Based on the test plan has been finalized, it
can be tested as described in the table below:
Result
Input Activity Observation Conclusion
Image detected
“Don’t
Enter”
The apps
show “Don’t
Enter” from
input.
[√ ]
accept
[ ] reject
detected
“turn
back”
Show “turn
back” from
input.
[√ ] accept
[ ] reject
2.4 Pengujian Performansi
Performance testing is used to determine the
accuracy of which is done using k-fold cross
validation, with k values as much as 10 fold. Aiming
to test the stability of accuracy when tested with the
training data and test data are different. The use of a
10-fold is recommended as the best to fold the
number of test validity. Tests conducted on methods
of Hidden Markov Model (HMM). To test of the
method of Hidden Markov Model (HMM) data that
is used by 82 data is divided into two subsets with
categories of violations "Do not Enter" and "reverse
direction", and used in performance testing using
iterations 50 times. The accuracy of the test scenario
performasi fold cross validation method as follows:
a. Testing Result Fold 1
The test results are not correct in the classification
results in a subset 1 is 20. So the value of accuracy
in experiment fold 1, which is
2
22
∗ 100% = 9.09%.
b. Testing Result Fold 2

7
results in a subset 2 is 2. So the value of accuracy in
experiments fold 2 that is
20
22
∗ 100% = 90.9 %.
c. Testing Result Fold 3
results in a subset 3 is 0. So the value of accuracy in
experiments is
22
22
∗ 100% = 100 %.
Hasil Percobaan Fold 4
Hasil test yang tidak benar dalam hasil klasifikasi
pada subset 4 yaitu 3. Sehingga nilai akurasi dalam
percobaan fold 1 yaitu
13
16
∗ 100% = 81.25%.
Accuration Result
Fold Accuration
Fold 1 50%
Fold 2 30%
Fold 3 40%
Fold 4 30%
Average
Of
Accuration
9.09% + 90.9% + 100% + 81.25%
4
= 70.31%
Of testing C-Fold Cross Validation uses 82 data got
an average accuracy value of 70.31% correct
classification and 29.69% wrong. Factors causing
the lack validan in classifying using Hidden Markov
Model (HMM) because it depends on the traits that
are used as training data.
3. Conclusion
Conclusions from the study entitled:
"Implementation of Hidden Markov Model method
and Gabor Filter To Detect Vehicle Traffic
Violations" is as follows:
1. Research on offense vehicle image processing can
be used to detect the image of the type of offense the
vehicle is prohibited "Do not Enter" and "reverse
direction".
2. Method Hidden Markov Models can be applied in
the process of determining the classification of the
type of offense "reverse direction" and "Do not
Enter" made vehicle on vehicle violations image
processing with accuracy results average 70.31%
correct classification.
4. LITERATUR
[1] Website Resmi Korps Lalu Lintas POLRI. 2015.
http://lantas.polri.go.id.
[2] Website Resmi Pemerintahan Jawa Barat. 2015.
http://www.jabarprov.go.id.
[3] D Sudian, Arman dan P Priambodo. “Aplikasi
Pengenalan Wajah (Face Recognition) Menggunakan
Metode Hidden Markov Model (HMM)”. Teknik
Elektro, UI.
[4] E Yuwitaning, B Hidayat dan N Andini. “Implementasi
Metode Hidden Markov Model Untuk Deteksi Tulisan
Tangan”. Teknik Elektro, Universitas Telkom.
[5] H Kekre and V Bharadi. 2010. “Gabor Filter Based
Feature Vector for Dynamic Signature Recognition.”.
[6] D Murugan, S Arumugam, K Rajalakshmi dan Manish.
2010. “Performance Evaluation of Face Recognition
Using Gabor Filter, Log Gabor Filter and Disctere
Wavelet Transform”.

8
[7] Sepritahara. 2012. “Aplikasi Pengenalan Wajah (Face
Recognition) Menggunakan Metode Hidden Markov
Model (HMM)”. Skripsi. Jakarta: Fakultas Teknik,
Universitas Indonesia.
[8] A Margono, I Gunawan dan R Lim. 2004. “Pelacakan
dan Pengenalan Wajah Menggunakan Metode
Embedded Hidden Markov Models”.
[9] A Agung, Fazmah A Yulianto dan W Maharani. 2011.
“Pengenalan Wajah Menggunakan Psedo-2D Hidden
Markov Model”.
[10] P Dymarski. 2011.“Hidden Markov Model, Theory and
Aplications”. India: InTech.
[11] A Khandual, G Baciu and N Rout. 2013. “Colorimetric
Preprocessing o
f Digital Colour Image”.
[12] Gabor Filter Imaging Filter. 2015. http://accord-
framework.net/docs/html/T_Accord_Imaging_Filters_G
aborFilter.htm.
[13]
Canny Edge Detector Class. 2015.
http://www.aforgenet.com/framework/docs/html/e08cae
30-7a37-db9f-cede-05cf6521343f.htm
[14]
P Mishra, R Chatterjee and V Mahapatra. 2010.
“Texture Gabor Filter Using Gabor Filter and
Wavelets”.

Hidden Markov Model Classification And Gabor Filter Preprocessing

Recommended

More Related Content

Similar to Hidden Markov Model Classification And Gabor Filter Preprocessing (20)

More from Kukuh Setiawan (20)

Recently uploaded (20)

Hidden Markov Model Classification And Gabor Filter Preprocessing