You are on page 1of 4

Traffic Sign Image Recognition Using Gabor

Wavelet and Principle Component Analysis


Immawan Wicaksono
Hendra Kusuma Tri Arief Sardjono
Department of Electrical Engineering
Department of Electrical Engineering Department of Electrical Engineering
Faculty of Electrical Engineering
Faculty of Electrical Engineering Faculty of Electrical Engineering
Institute Teknologi Sepuluh Nopember
Institute Teknologi Sepuluh Nopember Institute Teknologi Sepuluh Nopember
Surabaya, Indonesia
Surabaya, Indonesia Surabaya, Indonesia
immwnw@ee.its.ac.id
hendraks@ee.its.ac.id sardjono@elect-eng.its.ac.id

Abstract- The image recognition system of traffic signs is


one of the important elements needed by autonomous vehicles,
a match of 90% for a background similar to a database, and
which also these signs are a guide for autonomous vehicles. In has decreased to 75% when the background taken is
this paper we propose a traffic signs image recognition by different from the database. Tong Guofeng et all [3] used
implementing Gabor wavelet feature extraction. The images of SVM (Support Vector Machine) HOG (Histogram of
traffic signs will be captured by cameras. These images will Oriented Gradient) features and CNN (Convolutional Neural
then be processed for recognition. Where in the process the Network). The HOG feature technique calculates the
features of the traffic sign images will be extracted so that the appearance of gradient orientations in localized images and
identifier for each traffic sign is obtained. We use 8 SVM is an effective classifier. Hengliang Luo et all [3] used
orientations and 5 scales Gabor bank and PCA in recognition multi-task convolutional neural network, which is trained
process to get a good accuracy as well as efficiency. Easiness with a large amount of data and Chunsheng Liu et all [4]
and high accuracy was the avantages from proposed Gabor used a novel region-of-interest (ROI) extraction method
method. This method as seen from the experimental results called the high-contrast region extraction (HCRE), the split-
demonstrate fair reasonable in terms of preciseness, with flow cascade tree detector (SFC-tree detector), and a rapid
90.55% overall average recognition rate are obtained for the occlusion-robust traffic sign classification method based on
data of 100 traffic sign images. the extended sparse representation classification (ESRC).
Keywords— traffic sign, feature extraction, gabor wavelet III. PROPOSED FRAMEWORK
component.
The traffic sign recognition system that we proposed has
I. INTRODUCTION several details that will be explained in this section, i.e.
PCA, Gabor wavelet features then nearest neighbor
Traffic signs are one of the road equipment in the form
recognition. The processing’s steps are shown in figure 1.
of symbols, letters, numbers, sentences and its combinations
such as warnings, prohibitions, orders or directions for road
users. By recognizing traffic signs, the vehicle is expected
to provide information to the driver. In addition, automotive
manufacturers are currently aggressively developing
autonomous vehicles in which the detection and
introduction of digital images has a very important role for
driver assistance systems, and autonomous driving systems.
Fig.1: The Block Diagram of traffic sign Recognition System
The image recognition system of traffic signs is one of the
important elements needed by the autonomous vehicle, Note: after the picture of the traffic sign is taken, all traffic
where these signs are also a guide for autonomous vehicles sign images are cropped and resized to 200x200 pixels,
during trips on the road. and converted to grayscale images..
This research aims to develop a traffic signs recognition
system with a high degree of accuracy in recognizing traffic 3.1 Gabor Filter Construction
sign even when they are color degradation, dirty and various The first step of using gabor wavelet is Gabor filter
forms of traffic signs. We used Gabor wavelet as a feature construction. Gabor wavelet is very usefull tool as localize
extraction method in this research. when used in face characteristics in both frequency and spatial domain analysis
recognition biometric recognition system Gabor wavelet has in image analysis and machine vision, [1][5][6][7]. The
been proven to have excellent performance[1]. Gabor
equation of 2D Gabor filters can be defined as [8][9],
wavelet is a powerful tool for representing complex texture
features so this method intended to be precise and accurate
when applied to traffic signs recognition system. The
recognition is done by using 3 (three) shape of traffic sign (1)
such as rhombus, round and square In this research.
II. RELATED WORKS , , ,
Several works have been conducted to recognize traffic
sign. Sridevi et al [2] used the red mask technique to detect
traffic signs that looking red color in the image because all .
traffic signs in India are red and NCC (Normalized Cross-
Correlation) as a recognition system. The use of NCC shows

266
978-1-5386-8448-1/19/$31.00 ©2019 IEEE
Authorized licensed use limited to: Universitas Indonesia. Downloaded on March 23,2024 at 03:04:39 UTC from IEEE Xplore. Restrictions apply.
The scale and orientation of the Gabor wavelet filters
represented respectively by u and ν. The part between the
Gaussian envelope’s size and centre frequency the are
detemined by the symbol κ and η respectively. The
parameters values of κ and η are generally set to κ = η =√2
fmax that indicates the maximum frequency of the filters and
usually is set to fmax = 0.25. mostly image recognition
researchs use eight orientations, v ϵ {0, 1, 2, 3, 4, 5, 6, 7}
and five scales, u ϵ {0, 1, 2, 3, 4} , which causes a Gabor
filter bank of 40 filters [9][10].

Fig. 3. The output of filtering round traffic sign with the


whole of 40 Gabor filters

3.3 Principal component analysis


The total dimensions of Gabor Wavelet Features
Vector sets are considerable, and directly use the vectors
would lead to get various kinds of dimensional problem.
That problem can be prevented by using Subspace LDA as a
dimensionality reduction techniques. LDA provides a
mapping process that generate the within-class scatter small
and the between class scatter large. Through PCA this
process has shown could improve classification
Fig. 2. The real part of Gabor Wavelets bank with 5 achievement. On the other hand, it depends on a large
frequencies and 8 orientations training sample set to get good results. To deal analogous
problems, Zhao et al [9] before performing LDA, he using
3.2 Gabor Wavelet Features PCA to decrease feature size.
After we construct gabor filter, next step is make Gabor PCA aims to get a set of projection vectors that
wavelet features. By filtering an image R(x,y) using gabor project into a RPCA - dimensional feature space from the
filter that we construct with various scales and orientations original R-dimensional image space, in which RPCA << R. In
, we can received Gabor wavelet features. The PCA, the Ɲ images of the training set are merged to Ɲ
process of extracting features can be mathematically column vectors xi of length P accord to the number of
expressed as, pixels. If the total scatter is set by the P x P and mean is ,
i.e.
(2)
(4)

A set of eigenvectors can be found from PCA, i.e. {𝐰𝟏,


Where * stands for the convolution operator and 𝐰𝟐,…, 𝐰𝐃𝐏𝐂𝐀} of 𝑆𝑇, was used to construct the projection
stands for the complex output which can be broke down into
matrix 𝑊𝑃𝐶𝐴 = [𝐰𝟏, 𝐰𝟐, …, 𝐰𝐃𝐏𝐂𝐀] pursuant to the 𝐷𝑃𝐶𝐴
its imaginary parts and real in this largest eigenvalues. As a result of that, vectors of new
way: feature we get , i.e.

and (5)

(3) The purpose of LDA is to find the best set of separate


vectors that maps a M𝐿𝐷𝐴-dimensional feature domain from
the initial M-dimensional image domain (M𝐿𝐷𝐴≪M ) such as
that images of the same class are more closed and images
from various classes are more separated. A matrix projector
W is the objective of LDA, this matrix extend the result of
Sb and Sw [9],

(6)

Sw and Sb are the within-class scatter and between-class


distribute respectively. given Nc be the number of samples
267 in class c and let a C class patterns, a set of M training
Authorized licensed use limited to: Universitas Indonesia. Downloaded on March 23,2024 at 03:04:39 UTC from IEEE Xplore. Restrictions apply.
patterns from the C class can be assigned as
. The Sw first traffic sign image set, was used for training and
and Sb from a training set can be figured as : enrollment, and the second set were used for testing.
The classification tests were achieved by Subspace
(7) LDA used several numbers of eigenvectors. The recognition
was achieved using a cosine - based distance measure and a
(8) NN classifier, which is express as,

variable µc is the mean for the class c µ is the mean of all (11)
the training set, µ is the mean of all the training set. As can
be seen in [10] that from the eigenvectors of Sw-1Sb, the we used 8 orientations (v = 8) and 5, 4, and 3 filter scales
matrix projector W can be projected. Let we say WPCA is the (u=5, 4, 3) to construct The Gabor Feature,.
matrix projector from the PCA subspace of initial image
domain, WLDA (LDA matrix projector) is developed from the 1.3 Experiment Result
eigenvectors of . The Experiment results from the traffic sign images
final projection matrix W subspace LDA is show at the recognition can be seen on figures 5, 6 and 7.
equation 9,
Fig. 5 The Experimental Result Using 5 Scales and 8
(9) Orientations Gabor Wavelet

therefore a smaller traffic sign feature vector Fig. 6 The Experimental Result Of Using 4 Scales 8
are provided as, Orientations Gabor Wavelet

(10)

IV. Experimental Results And Discussion


To show the performance of our suggestion method,
we used our traffic sign image datasets.

4.1 Traffic sign image database


The traffic sign database consists of 100 images for 20
signs (5 images for each sign). The dimension of each image
is 200x200. Every specific traffic sign image is shown on
fig. 4.

Fig. 4. Sample traffic sign dataset

4.2 Experiment Setting


First step all traffic sign images were converted to
grayscale and then cropped to a size of 200 × 200 pixels to
make sure all those Traffic sign images have same size.
Next step is all the images were separated into two sets. The

268

Authorized licensed use limited to: Universitas Indonesia. Downloaded on March 23,2024 at 03:04:39 UTC from IEEE Xplore. Restrictions apply.
[3] T. Guofeng, C. Huairong and L. Yong, "Traffic Sign Recognition
Based on SVM And Convolutional Neural Network," 12th IEEE
Conference on Industrial Electronics and Applications, pp. 2066-
2071, 2017.
[4] H. Luo, Y. Yang, B. Tong, F. Wu and B. Fan, "Traffic Sign
Recognition Using a Multi-Task Convolutional Neural Network,"
IEEE Transaction on Image Processing, vol. 19, pp. 1100 - 1111,
2017.
[5] L. C and W. H., "Gabor Feature Based Classification Using the
Enhanced Fisher Linear Discriminant Model for Face Recognition,"
IEEE Transaction on Image Processing, vol. 11, pp. 467-476, 2002.
[6] L. Huang, A. Shimizu and H. Kobatake, "Classification-Based Face
Detection Using Gabor Filter Features," 5th IEEE International
Conference on Automatic Face and Gesture Recognition, pp. 397-402,
2004.
Fig. 7 The Experimental Result Of Using 3 Scales 8 [7] V. Kyrki, J. Kamarainen and H. Kalviainen, "Simple Gabor feature
space for invariant object recognition," Pattern Recognition Letters,
Orientations Gabor Wavelet vol. 25, pp. 310-317, 2005.
[8] V. Struc and N. and Pavesic, "From Gabor Magnitude to Gabor Phase
Features: Tackling the Problem of Face Recognition under Severe
4.4 Discussion Illumination Changes," in Face Recognition, intechopen online books,
The Traffic sign image recognition results obtained 2010.
from applying Gabor wavelet features extractors combined [9] H. Kusuma, Wirawan and A. Soeprijanto, "Face Recognition Against
with PCA has shown the good recognition performance on Varying Lighting Conditions Using Oriented PCI Features," Journal
traffic sign image database. of Theoretical & Applied Information Technology, p. 88, 2016.
[10] C. Wei-lun, "Gabor Wavelet Transform and Its Application,"
R98942073.
V. Conclusions
The potential method for traffic sign image
recognition have been proposed in this paper. Gabor
wavelet is used for extracting traffic sign features. The PCA
used to reduce matrix data that we process, for classification
all those feature vectors we used Subspace LDA. 24 Gabor
filters ( 3 scales, 8 orientations) that we used show the best
traffic sign image recognition performance,. The usefullness
of the traffic sign recognition method was evaluated on our
private dataset, i.e. traffic sign image dataset. As shown
from the experimental results, our proposed approach
demonstrates an acceptable traffic sign recognition
achievement, particularly to assure an excellent recognition
performance in the presence of different traffic sign shape.
Our method can reach good recognition accuracies which
are, 95%, 90%, 95% Using 8 Orientations And 5 Scales
Gabor Wavelet; 85%, 75%, 75% Using 8 Orientations And
4 Scales Gabor Wavelet; 71%, 55%, 60% Using 8
Orientations And 3 Scales Gabor Wavelet. The data shows
that the best results are obtained using Gabor wavelet with 8
orientations and 5 scales without reducing any number of
eigenvectors. side effects from without this reduction is a
large database but has the best recognition.

REFERENCES

[1] L. L. Shen and L. Bai, "A fast and robust Gabor feature based
method for face recognition," Special Issue on Crime Detection and
Prevention, Pattern Recognition Letters, 2005.
[2] M. Sridevi, N. Sankaranarayanan and A. Jyothish, "Automatic Traffic
Sign Recognition System Using Fast Normalized Cross Correlation
And Parallel Processing," International Conference on Intelligent
Communication and Computational Techniques, pp. 200-204, 2017.
269

Authorized licensed use limited to: Universitas Indonesia. Downloaded on March 23,2024 at 03:04:39 UTC from IEEE Xplore. Restrictions apply.

You might also like