Вы находитесь на странице: 1из 7

International Journal of Computer Information Systems and Industrial Management Applications.

ISSN 2150-7988 Volume 10 (2018) pp. XXX-XXX


© MIR Labs, www.mirlabs.net/ijcisim/index.html

Alphabet and Number Recognition of Indonesian


Sign Language Using SIFT Algorithm
Isa Mahfudi1, Moechammad Sarosa2, Rosa Andrie Asmara3 and Muh Bambang Purwanto4
14
Graduate Student of Electrical Engineering Department, State Polytechnic of Malang,
Malang 65141, Indonesia
Isa_mahfudi@polinema.ac.id 1 and muh_bambang@polinema.ac.id 4
2
Department of Electrical Engineering, State Polytechnic of Malang,
Malang 65141, Indonesia
msarosa@polinema.ac.id 2
4
Department of Information Technology, State Polytechnic of Malang,
Malang 65141, Indonesia
rosa_andrie@polinema.ac.id 3

MIR Labs, USA


2

Abstract: The inability to generally used them in form of a text or sound. methods are still not perfect
hear or talk to cause people communicating. They used Approach to the and will continue to do
to become isolated and sign language as the recognizing of sign/gesture development[6] [7].
feeling lonely. It really gives Generally be classified into Scale Invariant Feature
the influence to their social
primary language in doing
life. They are the deaf/mute their activity. Standard sign three, namely: based on data Transform Algorithm is one
disabilities. They used sign language in the country of glove [2][3][4], using of method for detecting
language as the primary Indonesian using sensors EMG object [1]. This method has
language in doing their Indonesian Sign Language (electromyography) [3][5] a good performance to
activity. Standard sign System (ISLS) and based on vision [6][7] changes in scale, rotation,
language in the country of Dictionary[1]. Sign/Gesture [8][9]. brightness and noise This
Indonesian using Indonesian
has content specific This method of data method will have an added
Sign Language System
(ISLS) Dictionary. However, information, sometimes glove is a glove consist of advantage by combining the
not all people understand used by normal people to strain gauge sensor and features of another
sign language, and normal communicate as a hand tracker. Strain gauge matching technique. In this
people who does not communication sensor is used to detect hand paper, researchers proposed
understand sign language, nonstandard. this is and bend the finger, and the SIFT method to use on
will find difficult to tracker is an electronic the recognition of Number
generally used to interact
communicate with people
with the computer[2]. system used to provide and Letter on Indonesian
who use sign language to
communicate. To resolve this information location, sign language.
problem, sign language orientation, speed and
translation allows to help direction of the hand. This
deaf and mute people to glove data produce good II. Related Work
communicate with normal performance and have a
people. Sign/Gesture high reliability in the Many researchers have been
recognition is a technique to done based on sign
recognize certain sign/gesture
recognition of sign
language. This method language recognition using
and the end result can be vision based. For hand
translated into form of a text eliminated the phases of
detection technique
or sound. Scale Invariant pre-processing on vision
Feature Transform
described by [10] and [11],
technique. The methods has both use the HSV colour
Algorithm is one of method deficiency, namely: users
for detecting object. This models and YCbCr colour
need to wear glove, this model for the process of
method has a good
performance to changes in system is quite expensive segmentation on the skin
scale, rotation, brightness because uses many sensor. color of the hand.
and noise This method will Many electronic devices Researcher [11] using
have an added advantage by embedded in gloves and a generic fourier descriptor
combining the features of lot of cables are connected and nearest neighbour on
another matching technique. between the computer and feature extraction process,
In this paper, researchers
the gloves, as well as have this research produces the
proposed the SIFT method to
use on the recognition of the potential to inhibit the accuracy 86% for static data
Number and Letter on comfort and movement and 69% for dynamic data.
Indonesian sign language. when using this gloves data In [12] using labview
The best performance on the [2][3][4]. For vision based recognize of sign language.
recognitoin of the character technique, input device Researchers using static
sign language (letters and using a camera that was data and dynamic data, the
numbers) on the threshold Figure 1. Indonesian Sign used to observe the hand accuracy obtained 97.79%
value of 0, 65 namely: “B, H, for static sign and 80.38%
“1”, “2”, “4”, “5” and “10” Language System (ISLS) and its movement. In this
for dynamic sign.
with its 100% accuracy. Dictionary[1]. method, the signer is in
Researchers [9] using edge
Change the threshold value front of the camera and the
orientation histogram. This
that gives best accuracy is at However, not all people camera captures sign image,
the threshold value of 0,45 method produces
understand sign language, and then in the process of recognition level of 88.26%
which produces 79,71% using vision method for
accuracy. and normal people who within recognition time of
does not understand sign recognized and translated. 0.5 second. those results are
Keywords: SIFT method, language, will find difficult This method is relatively tested on a complex
Indonesian Sign Language, to communicate with people inexpensive because background and mixed
Threshold. who use sign language to requires only a camera for lightning condition.
communicate. To resolve pictures and many Researchers [13] combine
this problem, sign language developed by the between edge detection
I. Introduction translation allows to help researchers. For researchers algorithm and skin
deaf and mute people to is a challenge in perfecting detection algorithm for
The inability to hear or the introduction in any use gesture identification, and
communicate with normal
talk to cause people to of the application. using ANN algorithm for
people. Sign/Gesture
become isolated and feeling Researchers have many Static hand gesture
recognition is a technique to
lonely. It really gives the presents various methods of recognition. In [14],
recognize certain
influence to their social life. introduction sign/gesture. researchers develop
sign/gesture and the end
They are the deaf/mute Until now, the existing recognition hand signals in
result can be translated into depth data using dynamic
disabilities. Sign language
How to Format Your Paper for JIAS 3

time warping. average


recognition rate of 92.4%.
Weighted hidden markov
models (WHMM) is also
used in research [15].
Weighted hidden markov
model is claimed capable of
overcoming the signs from
different signers. It is able
to provide accuracy of
94.74% every experiments.
The SIFT has on (a)
implementations in different
fields. Researchers [16]
have conducted
implementation SIFT for
the automated recognition Figure 4. Pre-Processing
of Batik patterns with batik and Hand Segmentation
data collected as many as Process
120 image that has 20 basic
batik pattern. The use of The YCbCr is a family of
SIFT Matching method colour spaces, Y is the luma
produces an equal error rate component and Cb and Cr
of 8.47%. In [17] the use of Figure 2. Block diagram of are the blue-difference and
SIFT combined with the Sign Language Recognition red difference chroma
technique of bag of features (b) components. Calculation on
extracted at the recogntion A. Image Acquicision Figure 3. Sign image the process of
of batik pattern produces a data. (a) Number Sign transformation from RGB
The first step would
good performance. This definitely be capturing all language (b) Letter Sign colour space to YCbCr
technique produces an character of sign image by Language. colour space shown in
average accuracy of webcam for further Equation 1[10].
97.67%, 95.47% and 79% processing. all sign image B. Pre-Processing and
in normal image, rotated were collected, it aims to Y  0,299  R  0,587  G  0,1
Hand Segmentation
image and a scaled image. create a database for Cb  0,1687  R  0,3312  G
The images were captured
training and testing data. Cr  0,5  R  0,4183 * G  0
by the webcam and then
III. Proposed System Sign language used on this
done the segmentation (1)
research is standard of
Sign language recognition process to eliminate the
indonesia sign language Skin detection
system is composed of four background and to get the
(ISLS). Total of sign processing shown in Fig 5.
main phases[1]. Image hand object. This process is
character used is 34 static Transformation result into
Acquisition phase, pre- usually called skin
data include: the letter YcbCr colour space shown
processing and hand detection. The sign image
consists of 24 (without 2 in Fig 5(b). Transformation
segmentation phase, hand captured is composed of
letter: J and Z) and number result from YCbCr into
posture recognition phase RGB color spaces, This
consist of 10 digits (1-10). RGB and as result of skin
and finally classification means that sign image
We have been collecting 340 area detection shown in Fig
gesture phase. Figure 2 combined with red (R),
images characters of sign 5(c). Calculation on the
shows the stages of the green (G) and blue (B) as of
(letters and numbers), process of transformation
process of the sign language three primary colours.
consist of 170 sample image from YCbCr colour space to
recognition. Luminance and
for training data and 170 RGB colour space shown in
chrominance are not
image for testing data. Equation 2[10].
separated in this colour
Each sign image has a
space. The value of R, G
resolution of 96 dbi (dots
and B are linear
per inch) and 24-bit mode
combination against the
with jpg image type. all
luminance of a pixel of
datasets image are
RGB.Researchers propose
conditioned with the same
using YcBcr method on
background and lighting.
research that can take a
Some of sign image
hand and removes
captured are shown in
unwanted background.
Figure 2.
Steps of pre-processing are
shown in Figure 4.
4

features are extracted using element feature vector for


Scale Invariant Feature L(x, y,  ) = G(x, y,  )  I(x,each
y) keypoint.
Transform SIFT method
(3) 5) Feature Vektor
[18]. The basic idea is to
2) Keypoint Matching
extract the invariant key
Localization As each feature point
point which
The detected local extrema corresponds to a 128
represents/identifies hand
are good candidates for element feature vector,
from the segmented image.
keyponts. However, they searching for matching
For this purpose of hand
need to be exactly localized points from two images
detection, SIFT features are
by fitting a 3D quadratic becomes searching for
first extracted from a set of
function to the scale-space matching feature vectors.
reference images and stored
local sample point. The Euclidean distance is used
in a database. An image
quadratic function is for measuring the similarity
frame is matched by
computed using a second degree between two feature
individually comparing
order Taylor expansion vectors, the ratio between
each feature from the image
having the origin at the the closest Euclidean
frame to this previous
sample point. Then, local distance and that of the
database and finding
extrema with low contrast second closest are computed
candidate matching features
and such that correspond to and if the ratio is lower than
(a) based on Euclidean distance
edges are discarded a certain threshold [18].
(b) (c) of their feature vectors.
because they are sensitive When the ratio of two
Figure 5. Skin Detection D. SIFT Algorithm to noise.
Process (a) Original Image distances is less than a
(b) The features are invariant to 3) Orientation given threshold, accept the
image scaling, translation, Assignment match points. The Euclidian
R  Y  1,4022  Cr and rotation, and partially distance of two N dimension
Local extrema detected in
invariant to illumination functions f1(x) and f2(x) is
G  Y  0,3456  Cb  0,7145  Cr DOG scale-space are called
changes and affine or 3D as follows[20]:
keypoints after the
B  Y  1,7710  Cb projection[18]. These
operations of improving d((fx), g(fx)) = (x1 - g(x1 )) 2 +
(2) features share similar
positioning accuracy and
properties with neurons in n
The image has noise after inferior temporal cortex that
the skin detection will be are
eliminating low-contrast
points. To determine the
=  f (x 1  g ( x1 )) 2
used for object i 1
done the process of removal recognition keypoint orientation, a
in primate
of noise using vision. SIFT algorithm gradient orientation
histogram is computed in In this paper use 0.65 the
morphological filtering includes five steps:(1)
the neighborhood of the vectors are considered to
namely: Morphology Scale-Space Local Extrema
keypoint. The contribution
opening (erosion followed Detection (2) Keypoint of each neighboring pixel is be   matched   with   each
by dilation). This process is Localization (3) Orientation weighted by the gradient other.   Example   matching
just used to the assignment (4) Keypoint magnitude and a Gaussian
segmentation phases. Descriptor (5) Feature result   with   Threshold   =
window with a value of δ
vektor matching. 0.65 is shown figure 7.
that is 1.5 times the scale of
the keypoint. Peaks in the
1)
histogram correspond to
1)
dominant orientations. A
1)
separate keypoint is created
1)
for the direction
1)
corresponding to the
1)
histogram maximum, and
1)
(a) any other direction within
Space-space Local
(b) 80% of the maximum value.
Extrema Detection
All the properties of the
Figure 6. Result process This stage of the filtering
keypoint are measured
Figure 7. Matching
attempts to identify those
of mophology (a) Original relative to the keypoint results with T=0.65.
locations and scales that
Image (b) After morphology orientation, which provides
are identifiable from
process. invariance to rotation[19]. IV. Eksperimental
different views of the same
object. This can be 4) Keypoint Descriptor Result
C. Hand Posture efficiently achieved using a
Recognition The descriptor is formed This stage it will be shown
"scale space" function. from a vector containing the
The image has been Further it has been shown the result of experiments to
values of all the orientation the recognition of sign
processed segmentation under reasonable histogram entries. The best
which show a picture of a assumptions it must be language using SIFT
results are achieved with a algorithm with a variety of
hand object and background based on the Gaussian 4×4 array of histograms
has been removed. The next function. The scale space is characters. Figure 8. shows
with 8 orientation bins in a matching result with the
process is to change this defined by the function[18]: each histogram[20]. this
image to grayscale. The value Threshold of 0.65.
paper use a 4×4×8 = 128
converted gray scale image
is normalized. Invariant
How to Format Your Paper for JIAS 5

system makes some obtain   best   result   with


mistakes while recognition. T=0.45   that   produces   the
For the characters such as
“B, H, “1”, “2”, “4”, “5” value   recognition   rate   to
and “10” the images of 79,71%.   The   illustrate   of
hand gesture are good results   sign   language
quality after preprocessing
recognition   number   shown
stage. Thus the system
achieves high success rate at Figure 9.
in recognition for these
characters. Table 2 show
test results matching by
changing the value of the
threshold.

(a) Table 1. The Success Rate


of Recognition
Characte Correct Wrong
r (times) (times)
A 9
B 10
C 8
D 9 (a) Exact matching
E 8
F 5
results
G 9
H 10
I 9
K 9
L 9
M 4
N 4
O 7
P 6
Q 6
R 7
S 4
T 4
U 8
V 4
W 8 (b) False matching
X 8
(b)
Y 8 results
Figure 8. Matching results
with the value of threshold
1 10 Figure 9. Ilustrate
2 10
0.65 3 8 results Matching
4 10 performed by SIFT
The next stage will be 5 10
shown the test results with Method.
6 8
340 image dataset that 7 4
consists 170 sample images 8 4
testing data and 170 sample 9 5 V.Conclusions
images trainning data. 10 10
Result of the proposed This paper proposes SIFT
Average method for Indonesian sign
system is shown on Table 1.
there are 34 characters language recognition.
Table 2. Recognition results Change the value of the
(Letter and Number)
of sign language number threshold provide influence
Indonesian sign language
with total success rate up to Value on accuracy of recognition.
of Threshold
74,12%. The results show 0.45 0.55 The best performance on the
that the success rates of Recognition 79,71 76,47 recognition of the character
characters “M”, “N”, “S”, Rate (%) sign language (letters and
“T”, “V”, “3”,”6”, “7”, “8” numbers) on the threshold
and “9" are low because value of 0,65 namely: “B,
Table   2   shows   the H, “1”, “2”, “4”, “5” and
they are shape-like.Thus the
recognition   rate   with “10” with its 100%
different   threshold.   We accuracy. Change the
6

threshold value that gives pattern “Pattern Matching 110, 2004.


best accuracy is at the classification using for Automatic Sign [19] W. Wang, J. Hong,
threshold value of 0, 45 spectral estimation Language and Y. Tang, “Image
which produces 79,71% and neural Translation System matching for
accuracy. network,” SICE, using LabVIEW geomorphic
2007 Annu. Conf., Keywords :,” pp. measurement based
pp. 1108–1111, 660–665, 2007. on SIFT and
Acknowledgment 2007. [13] S. K. Yewale and P. RANSAC
The Authors would like to [6] S. S. Rautaray and K. Bharne, “Hand Methods,” Proc. -
thank to The State A. Agrawal, “Vision gesture recognition Int. Conf. Comput.
based hand gesture using different Sci. Softw. Eng.
Polytechnic of Malang for
recognition for algorithms based on CSSE 2008, vol. 2,
supporting to attend this
human computer artificial neural no. 50505051, pp.
research. interaction : a network,” 2011 Int. 317–320, 2008.
survey,” 2012. Conf. Emerg. [20] J. Zhu, S. Wang,
References [7] N. A. and S. S. Trends Networks and F. Meng, “SIFT
Pragati Garg, Comput. Commun., method for paper
[1] I. Mahfudi, M. “Vision Based Hand no. 1998, pp. 287– detection system,”
Sarosa, R. A. Gesture 292, 2011. 2011 Int. Conf.
Asmara, and M. A. Recognition,” Int. J. [14] G. Plouffe and A.- Multimed. Technol.
Gustalika, Comput. Inf. Eng., M. Cretu, “Static ICMT 2011, pp.
“Indonesian Sign vol. 3, no. 1, pp. and Dynamic Hand 711–714, 2011.
Language Number 186–191, 2009. Gesture Recognition
Recognition using [8] J. Singha, A. Roy, in Depth Data
SIFT Algorithm
Author Biographies
and R. H. Laskar, Using Dynamic
Indonesian Sign “Dynamic hand Time Warping,” Isa Mahfudi,
Language Number gesture recognition IEEE Trans. received the
Recognition using Bachelor of
using vision-based Instrum. Meas., vol. Applied Science,
SIFT Algorithm,” approach for 65, no. 2, pp. 305– in electrical
vol. 336, 2018. human–computer 316, 2016. engineering from
[2] R. M. Prakash, State Polytechnic
interaction,” Neural [15] W. Yang, J. Tao, C. of Malang,
“Gesture Comput. Appl., vol. Xi, and Z. Ye, “Sign Indonesia in
Recognition and 29, no. 4, pp. 1129– Language 2015. He is
finger tip detection 1141, 2018. Recognition System currently a
for Human magister student
[9] J. Pansare and M. Based on Weighted at electrical
Computer Ingle, “Vision- Hidden Markov engineering
Interaction,” in Based Approach for Model,” 2015. departement in
International American Sign [16] I. Nurhaida, A. State Polytechnic
Conference on of Malang,
Language Noviyanto, R. Indonesia. Her
Innovations in Recognition Using Manurung, and A. research interests
Information, Edge Orientation M. Arymurthy, include antenna,
Embedded and computer vision,
Histogram,” 2016 “Automatic image processing,
Communication Int. Conf. Image, Indonesian’s Batik Microcontroller
Systems (ICIIECS), Vis. Comput., pp. Pattern Recognition and signal
2017. 86–90, 2016. Using SIFT processing.
[3] M. E. Al-ahdal, N. [10] A. S. Konwar, B. S. Approach,” Moechammad
Tahir, and U. T. Borah, and C. T. Procedia Comput. Sarosa, received
Mara, “Review in Tuithung, “An Sci., vol. 59, no. the diploma of
Sign Language engineering
American Sign Iccsci, pp. 567–576, technology from
Recognition Language detection 2015. Université de
Systems,” IEEE system using HSV [17] R. Azhar, D. Nancy I, France
Symp. Comput. color model and Tuwohingide, D. in 1989. His
Informatics, pp. 52– master and
edge detection,” Kamudi, doctoral degrees
57, 2012. 2014 Int. Conf. Sarimuddin, and N. in computer
[4] V. A. Farooq Commun. Signal Suciati, “Batik technology is
Husain, Shivani Process., pp. 743– Image Classification received from
Gandhi, Tanisha Bandung Institute
747, 2014. Using SIFT Feature of Technology,
Nijhawan and S. P. [11] A. Izzah and N. Extraction, Bag of Indonesia, in
Sehba Khatun, Suciati, Features and 2002 and 2007,
“Gesture respectively. He
“Translation of Sign Support Vector is currently
Recognition System Language Using Machine,” Procedia lecturer in
Using Matlab: A Generic Fourier Comput. Sci., vol. Department of
Literature Review,” Descriptor and 72, pp. 24–30, Electrical
vol. 2, no. 11, pp. Engineering of
Nearest Neighbour,” 2015. both Brawijaya
425–432, 2012. Int. J. Cybern. [18] D. G. Lowe, University and
[5] K. K. Jung, J. W. Informatics, vol. 3, “Distinctive Image State Polytechnic
Kim, H. K. Lee, S. of Malang. His
no. 1, pp. 31–41, Features from,” Int. current research
B. Chung, and K. 2014. J. Comput. Vis., vol. interests lie in
H. Eom, “EMG [12] A. Domingo, 60, no. 2, pp. 91– information
How to Format Your Paper for JIAS 7

communication
technology,
decision support
system, artificial
intelligence, and
mobile
computing.

Rosa Andrie
Asmara,
received the B.E.
degrees, in
electronics
engineering from
Brawijaya
University,
Indonesia in
2004 and M.S.
degree in
Multimedia
engineering from
Institute of
Technology
Sepuluh
Nopember,
Surabaya,
Indonesia in
2009, and Dr.
Eng degrees from
Saga University,
Japan in 2013.
He is a lecturer in
the Informatics
Management
field, Department
of Information
Technology at
State
Polytechnics of
Malang,
Indonesia. His
research interests
include signal
processing, image
processing,
parallel
processing,
pattern
recognition, and
computer vision.

Muh Bambang
Purwanto,
received her
Bachelor of
Engineering, in
electronics
engineering from
University
Merdeka Malang,
Indonesia in
2015. Currently
he has been
studying in
Electrical
Engineering, at
the third semester
of Master Degree
in State
Polytechnic of
Malang,
Indonesia. Her
research interests
include Control,
computer vision
and image
procesing.

Вам также может понравиться