Uploaded by lahore gg

TSP2018 template A4 2 columns for author names (2)

advertisement
Sign Language Recognition System
Mr. Usman Latif, Mr. Mubashar Ali khan, Mr. Ammar Rafiq, Miss Amna Yasin
Abstract—Hearing impeded individuals consistently utilizes
signs to speak with one another. In any case, it is consistently
hard for an ordinary people to comprehend gesture based
communication. Sign based communication is meaningful and
normal way for communication among normal and hearing
crippled people (information altogether went on through the
hand movement).Thus, an interpreter is expected to comprehend
what they talk and speak with us. The entire thought is to
manufacture a framework that empowers correspondences between a conference debilitated and ordinary individuals. The Sign
Language Recognition (SLR) framework makes an interpretation
of the gesture based communication to message. Proposed work
plans to assemble a framework that will have the option to arrange the different hand offers of gesture based communication.
Proposed framework is able to do recognize alphabets, digits,
some keywords and also capable of concatenation of alphabets
to make words. In proposed work, set of image data is trained
using machine learning algorithms that is Convolutional Neural
Network (CNN) and testing is done on runtime. Image processing
and Feature Extraction techniques are applied to simplify the
images. By using CNN model during the training and validation
of data set system gained accuracy up to 80%. By using the SLR
system it becomes easy for normal people to understand sign
language of hearing impaired people.
Index Terms—Convolutional Neural Network (CNN), Concatenation of Alphabets, Sign Language Recognition.
I. I NTRODUCTION
People express their feelings, express their thoughts, their
opinions, and experiences from the people around them “by
speech”. They communicate by listening and talking, but there
are a lot of people out there unable to hear or speak. It would
not be right for us to overlook the individuals who are denied
of this valuable blessing. Hearing impaired (HI) individuals
find it difficult to convey their message to ordinary individuals.
The only communication method available for HI individuals
is “sign language”. But the problem is that ordinary individuals
cannot understand the sign language. So it is difficult for
ordinary people to understand hearing impaired individuals.
Using sign language they are limited to their own world.
Hard of hearing is a handicap that disable their hearing and
make them unfit to hear, while quiet is a handicap that weaken
their talking and make them unfit to talk. Both are as it were
crippled at their hearing as well as talking, consequently can at
present do a lot different things. The main thing that separate
them from ordinary people is communication. If there is a
route for typical individuals and hard of hearing individuals to
communicate, the hard of hearing individuals can undoubtedly
live like an ordinary individual. Furthermore, the main way
for them to impart is through communication via gestures.
While gesture based communication is difficult to understand
for ordinary individuals, is as yet getting little consideration
from the ordinary individuals. We as the typical individuals
will in general overlook the significance of communication via
gestures, except if there are friends and family who are hard of
hearing. One of the solution for speak with the hard of hearing
individuals is by utilizing the administrations of gesture based
communication mediator. In any case, the utilization of sign
language translator can be exorbitant. Modest arrangement is
required so the hard of hearing and ordinary individuals can
impart typically.
Thusly, researchers need to find a way for the hard of hearing quiet people so they can examine viably with customary
person. The disclosure for this is the utilization of Sign Language Recognition System. The proposed work expect to see
the gesture based communication, and make a translation of
it to the local language through substance or talk. Regardless,
developing this structure cost a great deal and is difficult to be
applied for ordinary use. Early investigates have known to be
powerful in Sign Language Recognition System by using data
gloves. However, the critical cost of the gloves and wearable
character make it difficult to be promoted [1]. Understanding
that, examiners by then endeavor to develop an unadulterated
vision Sign Language Recognition Systems. In any case, it is
additionally going with inconveniences, especially to follow
hands development.
The issues of creating communication via gestures range
from the picture securing to the classification cycle. Specialists
are as yet finding the best technique for the picture securing.
Social event pictures utilizing camera gives the challenges of
picture pre-handling. In the interim, utilizing dynamic sensor
gadget can be expensive. Classification techniques additionally
give specialists a few downsides. Wide decision of recognition
strategy makes researcher incapable to focus on one best
technique. Picking one strategy to be centered around will in
general make other technique that may be better suit for Sign
Language Recognition, not being tried. Evaluating different
strategies makes researcher scarcely creates one technique to
its fullest possibilities.
This task attempted to eliminate this obstruction between
typical what’s more, the hard of hearing people by building
up a framework which can make an interpretation of communication through gesture into text. We have proposed a
convolutional neural organization model which can perceive
the signs with precision.
II. R ELATED W ORK
Al-Ahdal and Tahir [2], acquainted the novel’s methodology
with planning a SLR framework dependent on EMG sensors
with information glove. This technique depends on the indications of electromyography recorded in the possession of
the muscles to spread the limits of the expressions of word
streams in constant SLR. SLR System is sought after because
of its capacities beating the hindrance between the almost
totally senseless. At present, powerful SLR is not, at this
point accessible world in light of numerous deterrents. Also,
as we probably am aware, Sign Language acknowledgment
has risen as a need significant examination territories in the
field of human PC (HCI). Consequently, this paper presents
an outline of much examination is dealing with the reception
of Sign Language framework, and the serious framework was
isolated by an image talks about photography procedures and
survey methods. Shortcomings and shortcomings that add to
the framework full usefulness or in any case will be featured by
asking the serious issues related with cutting edge frameworks.
Next, a novel method to manufacture a SLR framework
dependent on Includes EMG sensors and information glove.
This technique depends on electromyography signals recorded
from hand muscles to share the limits of the names of surges
of words in persistent SLR. The proposed program is required
to take care of the spelling issue, which will contribute in data
created with nonstop sign quality mindfulness program.
Iwan Njoto Sandjaja and Nelson Marcos [3], proposed
shading glove strategy that eliminates key highlights from
video utilizes a multicolor tracking algorithm. The recognition
framework establishes the framework for manual recognition
that tends to genuine and current issues by enlisting the hard
of hearing network and prompting compelling use. Information number for gesture based communication obtaining the
program is a video record number for 5,000 Filipino original
copies Frame size 640 x 480 pixels with 15 edge/second.
Shaded gloves utilize less shading than hued ones gloves in
existing exploration. Framework redistributing is basic Features from the video utilize a multicolor following calculation
quicker than the shading following calculation accessible in
light of the fact that didn’t utilize a monotonous cycle. Next,
the program peruses again observes the estimation of Filipino
communication via gestures and preparing the exploratory
stage utilizing Markov’s concealed model. The framework
utilizes Markov’s shrouded model (HMM) for the preparation
and testing stage. Highlight deliveries can follow 92.3 percent
everything being equal. The showcase can likewise observe
Filipino gesture based communication an aggregate of 85.52
percent with moderate precision.
Dibya Jyoti Bora1, Anil Kumar Gupta and Fayaz Ahmad
Khan [4], presents the paper while picking the correct shading
space is the most significant issue for shading picture partition
measure. Normally L * A * B * and HSV are the most
widely recognized chosen shading spaces. In this paper a
near examination done between these two shaded territories
according to shading picture partition. By estimating their
presentation, we take a gander at the boundaries: mse and
psnr. Things being what they are, HSV shading space works
superior to L * A * B *.
Hand and late disclosure PC research, there has been a great
deal of work in the previous a hand motion was performed.
Numerous new specialist co-ops they offer continuous gadgets
and convenient innovation. A considerable lot of these are
business the accessible innovation is costly and costly. The
reason for this work was to make ongoing other regular
activity recognition webcam. Through the investigation cycle
we exhibit that we recommend that the framework have the
option to distinguish and identify contact 80% precision. The
proposed cycle can proceed with created by creating following
techniques utilizing a model based following [5].
In [6], the creators examined the shading structures of hsv
space with an accentuation on visual view of assorted variety
in Hue, Saturation and pixel intensity of the picture. Eliminate
pixel includes by choosing Hue or Quantity as the prevailing
resource as far as Saturation number of pixels. The segment
that utilizes this strategy gives better ID of articles in the
picture contrasted with those created utilizing RGB shading
space.
In [7], the creators recommended shading partition they
moved toward where they began changing over the rgb picture
into a solitary hsv. At that point utilize the majority of Otsu’s
obstructing on V channel to get a pleasant fringe from the
photograph. The subsequent picture is then separated into KMeans incorporation for ceaseless combination far off locales
coming about because of the utilization of Many Otsu block. In
the end they made a foundation evacuation and morphological
revision. The consequence of this technique is discovered to
be good regarding MSE and PSNR values are remembered for
the review.
In [8], the creators initially changed over the first picture
from RGB structure HSV structure. At that point utilize the
move you mean with FELICM rather than Hue, Saturation
and Value Components. Last photographs acquired from move
with FELICM at blend. The proposed technique shows better
execution the level has past calculations.
In [9], the creators proposed another quantization cycle HSV
shading space to produce shading histogram once a dark histogram of K-Means combination, working across various size
in HSV shading space. Thusly, usage of centroids and number
of assortments naturally estimated. The foundation preparing
channel says acquainted with effectively dispose of low-weight
locales. This a technique is accessible to accomplish high
computational speed again the outcomes are shut to public
remark. Also, about thusly, can be given in key districts of
photographs effectively.
In [10], the creators changed the picture from the rgb
shading space to l * a * b * shading space. In the wake of
separating the three channels of l * a * b *, one course is
chosen by the shading underneath thought. From that point
forward, a hereditary calculation is utilized for that channel
picture. This strategy is discovered to be viable in isolates
complex foundation pictures.
In [11], the picture is changed over from rgb to l * a *
b * space, and afterward k implies that the calculation is
applied to the outcome picture. The pixels are then named
with an alternate picture. In the long run pictures were made
that isolated the genuine picture by shading. As a novel
consolidating shading picture made by a blend a split is
proposed. Here, the shading space for l * a * b * is chosen.
At that point there is the consolidated exertion of the KMeans calculation, the sobel channel and the water calculation
is utilized for dividing. The outcome is discovered to be
acceptable as indicated by MSE and PSNR rates.
In [12,] the creators proposed a powerful combination
calculation when the consolidating is done in the shading space
l * a * b *. Picture division is accessible straightforwardly
by setting every pixel and its comparing assortment. The
calculation is applied for detachment and picture testing the
outcomes plainly show the order of what you like.
In [13], the creators proposed an assortment of ants system
corresponding to the shading region of the CIE Lab where
CMC separation is utilized to figure the separation between
pixels as this separation estimation is accessible to deliver the
best outcomes according to the CIE lab shading field. Execution of this technique is contrasted with the MSE boundary
and is accessible at fulfill.
Asanterabi Malima, Erolozgur, and Mujdatcetin [14], This
methodology contains ventures for portioning the hand district,
finding the fingers ,lastly ordering the signs. The calculation
is invariant to interpretation, turn, and size of the hand. This
calculation can be reached out in various manners to perceive a
more extensive arrangement of motions. The division segment
of calculation is excessively basic, and should be improved if
this method should be utilized in testing working conditions.
Solid execution of hand signs acknowledgment methods in
an overall setting require managing impediments, transient
following for perceiving dynamic motions, just as 3D demonstrating of the hand, which are still generally far off the present
status of the workmanship.
Mark Batcher [15], Gripsee is the name of the Robot whose
implementation is debated in this paper, robot specifically use
to understand an article, manage it, and transfer it from one
place to another. It fills in as a multi-skill Robot that can play
out various undertakings, robot is used as a services robot.
Kevin Gabayan, Steven Lansel [16], proposed the paper
that deals with the lively time sign acknowledgment approach
including single sign medium. Model, a sensor collaboration
prototyping programming and equipment condition, as of
now utilizes a unique time wrapping motion acknowledgment
approach including single sign channels. Creator use a five
channel accelerometer and whirligig blend board to test translational and rotational expanding speeds, and a microcontroller
to achieve simple to advanced transformation also transfer
approaching signs. Layout coordinating by means of direct
time traveling and lively time traveling are disconnected,
just as fortification learning through the algorithm HMM
continuously.
M. Ebrahim Al-Ahdal and Nooritawati Md Tahir [17],
proposed a paper that represent a diagram of fundamental
exploration that dependent on the gesture Language acknowledgment framework, and the formed framework ordered into
the gesuture catching strategy and acknowledgment methods
is talked about. The qualities and impediments that add to the
framework working impeccably or in any case will be featured
by summoning serious issues related with the created frameworks. Next, a novel strategy for planning SLR framework
dependent on consolidating EMG sensors with an information
glove is suggested. This strategy depends on electromyography
signals captured from hands for allotting word limits for floods
of words in nonstop system.
Iwan Njoto Sandjaja and Nelson Marcos [18], proposed a
work in which Sign language acknowledgment framework that
detects numbers sets down establishment for hand movement
acknowledgment which tends to genuine and latest issues in
marking hearing impaired network and prompts pragmatic
applications. The contribution of gesture based communication
number acknowledgment framework is 5000 Filipino gesture
based number video document with 640 x 480 pixels outline
size and 15 casing/second. A shading coded gloves utilizes
minor shading contrasted and other shading coded gloves
in the current exploration. The framework extricates significant highlights from the video utilizing multicolor following
calculation which is quicker than existing shading following
calculation since it didn’t utilize recursive strategy. After that,
framework grasp and perceives the Filipino gesutre based
number in preparing and test stage utilizing HMM. The
framework utilizes HMM to prepare and test stage. The component extraction can follow 92.3% all things considered. The
recognizer additionally could perceive Filipino communication
via gestures number with 85.52% normal precision.
Noor Adnan Ibraheem and Rafiqul Zaman Khan [19],
proposed a review for different earlier motion acknowledgment
approaches is given specific accentuation close by hand signs.
An audit of static hand act strategies are clarified with various
instruments and calculations applied on signal acknowledgment framework, including connectionist models, HMM, and
fuzzy clustering. Difficulties and future enhancement are additionally featured.
Archana S. Ghotkar, Rucha Khatal , Sanjana Khupase,
Surbhi Asati and Mithila Hadap [20], proposed different
authentic foundation, required, degree and worry of ISL are
mentioned. Vision based hand motion acknowledgment framework debated as hand plays crucial correspondence mode.
Pondering earlier reported work, various techniques open for
hand following, division, include extraction and order are
recorded. Vision based framework have difficulties over conventional equipment based methodology; by productive usage
of PC vision and example acknowledgment, it is conceivable
to take a shot at such framework that will be common and
acknowledged, all in all.
Paulraj M P, Sazali Yaacob, Mohd Shuhanaz bin Zanar
Azalan, Rajkumar Palaniappan [21], proposed a basic gesture
based communication acknowledgment framework that has
been created utilizing skin shading division and ANN. The sec-
ond invariants highlights removed from the privilege and left
hand motion pictures are utilized to build up an organization
model. The framework actualized and tried for its legitimacy.
Framework shows recognition rate of 92.85% after the testing.
Divya Deora1, Nikesh Bajaj,k [22] presents that every
gesture based communication system is prepared to perceive
explicit arrangements of gesture that likely yield the gesture
in necessary configuration. These frameworks are worked with
ground-breaking picture preparing strategies. The communication via gestures acknowledgment frameworks are fit for perceiving a particular arrangement of marking motions and yield
the comparing text/sound. The majority of these frameworks
include the methods of discovery, division, following, motion
acknowledgment and arrangement.
Yikai Fang, Kongqiao Wang, Jian Cheng and Hanqing Lu
[23], proposed a solid progressing hand gesture identification
technique. In this proposed work, directly off the bat, a
specific movement is expected to trigger the hand ID followed
by following; by then hand is divided using development
and concealing prompts; finally, to break the limitation of
perspective extent experienced in most of learning based hand
signal methods, the scale-space incorporate revelation is joined
into movement affirmation. Applying the proposed procedure
to course of picture examining, preliminary outcomes show
that our method achieves great execution.
J. H. Kim, N. D. Thang, and T. S. Kim [24], proposed
a work in which creator built up a 3D hand movement
following and sign acknowledgment framework through an
information glove (to be specific the KHU-1 information glove
comprising of three trihub accelerometer sensors, one regulator, one Bluetooth). The KHU-1 information glove equipped
for sending hand movement gestures to computer through
remote correspondence by means of Bluetooth. Additionally
we have executed a 3D computerized hand model for hand
movement following and acknowledgment. The actualized 3D computerized hand model depends on the kinematic chain
hypothesis using ellipsoids and joints. At last, creator have
used a standard based calculation to perceive straightforward
hand signs to be specific scissor, rock, and paper utilizing the
3D advanced hand model and the KHU-1 information glove.
Various starter trial results are introduced in the paper.
J. Weissmann and R. Salomon [25], proposed a paper
that investigates the utilization of gestures of hand as a
methods for human and PC collaborations for augmented
reality applications. For the application, explicit hand signs,
for example, ”clench hand”, ”pointer”, and ”triumph sign”,
have been characterized. Various exisiting approaches utilize
different camera based acknowledgment frameworks, which
are fairly expensive and delicate to ecological changes.
MIE324 Final Report [26], Sign Language Recognition
Anna Deza ( ) and Danial Hasan ( ) Decemeber second 2018
Word Count: 1993 Penalty: 0%. The objective of this undertaking was to fabricate a neural organization ready to characterize
which letter of the American Sign Language (ASL) letter set is
being marked, given a picture of a marking hand. This venture
is an initial move towards building a potential gesture based
communication interpreter, which can take correspondences
in communication via gestures and make an interpretation of
them into composed and oral language. Such an interpreter
would incredibly bring down the boundary for some hard
of hearing and quiet people to have the option to all the
more likely speak with others in everyday collaborations.
This objective is additionally roused by the separation that
is felt inside the hard of hearing network. Forlornness and
sadness exists in higher rates among the hearing impaired
populace, particularly when they are drenched in a meeting
world. Enormous obstructions that significantly influence life
quality originate from the correspondence disengage between
the hearing impaired people. A few models are data hardship,
restriction of social associations, and trouble incorporating in
the public eye. Most exploration executions for this assignment
have utilized profundity maps produced by profundity camera
and high goal pictures. The target of this venture was to
check whether neural organizations can arrange marked ASL
letters utilizing basic pictures of hands taken with an individual
gadget, for example, a PC webcam. This is in arrangement
with the inspiration as this would make a future usage of a
continuous ASL-to-oral/composed language interpreter viable
in a regular circumstance.
III. P ROPOSED M ETHODOLOGY
In proposed methodology a system is developed that takes
gestures of user making particular sign through webcam and
converts them into text without use of any data glove or
sensor-based glove. In this system a Convolutional Neural
Network (CNN) model is implemented using both Tensorflow
and Keras. Then the model was trained using Keras.
A. CNN Model of proposed work
B. Convolution
Convolution is a way to capture information about the
ordering of pixels. The type of convolutions are interested in
2d discrete convolutions, which act like a weighted sliding
sum over an area of pixels. For instance, a matrix called a
kernel slides across the pixels in an image. At each point, it
calculates the weighted sum of the kernels’ values and pixel
in chunk of the image. The sum is then put in the first value
of the output image. The kernel then slides over one pixel and
repeats the process for every pixel in the image.This process
incorporates information about a pixels’ neighboring values
into its own value. If we compare the original image to the
output image, the result looks like a low-budget photoshop
filter.
It perform different convolutions on input, where every
activity utilizes an alternate channel. This outcomes in various component maps. At last, it take these element guides
and set up them as the last yield of the convolution layer.
Much the same as some other Neural Network, it utilize an
activation function to make yield non-direct. On account of
a Convolutional Neural Network, the yield of the convolution
will be gone through the activation work. This could be the
ReLU enactment work.
C. Relu
Fig. 1. CNN Model of proposed work
This model is manufactured utilizing 3 convolutional layers
with different sizes advancing from 2 x 2 to 5 x 5, activation
function ReLu, and the standard Max Pooling and Dropout.
This manufactured model is fed to fully connected layer which
yield the output of 26 classes of Alphabets, 10 classes of
digits and 7 classes of specific words. The design followed
by imagining that the first layer with little kernel size would
catch little features, for example, hand layout, finger edges and
shadows. The bigger kernel size ideally catches blends of the
little features like finger crossing, edges, hand area, etc.
An extra activity called ReLU has been utilized after each
Convolution activity ReLU refers to Rectified Linear Unit and
is a non linear movement. ReLU segments clever movement
(applied per pixel) and replaces all negative pixel regards in
component map by zero. The purpose behind ReLU is to
introduce non-linearity in CNN, since most of this current
reality data we would require CNN model to learn would
be non-direct (Convolution is a straight activity component
shrewd grid augmentation and expansion, so we represent nonlinearity by presenting a non-direct capacity like ReLU).
D. Pooling
In CNN, pooling is utilized to diminish the spatial size of the
convolved include. There are basically two sorts of pooling,
for example, max pooling and normal pooling. In max pooling,
a window moves over the input matrix and makes the matrix
with maximum values of those windows.
In average pooling, it is similar to max pooling but uses
average instead of maximum value. The window moves according to the stride value. If the stride value is 2 then the
window moves by 2 columns to right in the matrix after each
operation. In short, the pooling technique helps to diminish the
computational force needed to break down the information.
E. Classification
Fig. 2. Convolution Neural Network Working
Here, the completely associated layers will fill in as a
classifier on head of these extracted features. They will allocate
a likelihood for the item on the picture being what the
calculation predicts it is. Adding a Fully-Connected layer is
an (ordinarily) humble strategy for learning non-direct blends
of the huge level features as addressed by the yield of the
convolutional layer. The Fully-Connected layer is learning a
maybe non-direct capacity in that space.
The fully connected layers gain proficiency with a (perhaps
non-direct) work between the significant level highlights given
as a yield from the convolutional layers. Over a movement of
ages, the model can perceive overpowering and certain lowlevel features in pictures and describe them using the Softmax
Classification methodology.
C. Error Rate of Alphabets
Figure 4 shows the error rate of alphabets. The error rate
for alphabets lies between 0% to 50%. Some alphabets have
0% error rate as the system predicts them correctly. The error
rate of other alphabets is due to same shape of the signs. The
system get confuse to predict some alphabets due to their same
shape.
IV. R ESULTS
By using the CNN model during the traning and validation
of data set the system have gained accuracy upto 80%. This
system includes 43 classes. Accuracy is different for different
classes. Some signs have the recognition rate of 100% while
some have the recognition rate of 50% to 90%.
A. Recognition and Error Rate
1) Recognition Rate: The Accuracy of each class is calculated by the following formula:
Recognition Rate = (Number of correct recognition / Total
images) * 100
In this Total images is the No.of images taken as input
and the No. of correct recognition is the prediction that the
system made correctly. It is then multiplied by 100 to get the
percentage.
2) Error Rate: In predective model none of the recognition system recognize the signs with 100% accuracy, as the
accuracy depends on many factors. The error rate in our SLR
system is calculated by the following formula
Error Rate = 100-Recognition rate
Fig. 4. Error rate of Alphabets
D. Accuracy Rate of Digits
Figure 5 shows the accuracy rate of digits. Some digits have
the accuracy rate of 100% while some have the accuracy rate
of 50% to 80%. This system is fairly good at predicting the
correct digits as shown in figure.
B. Accuracy Rate of Alphabets
Figure 3 shows the accuracy rate of alphabets. Some alphabets have the accuracy rate of 100% while some have the
accuracy rate of 50% to 90%. This framework is genuinely
acceptable at predicting the right alphabets. Investigating some
of confounded letters in order, similar to V and W, we see that
the two letters are comparative fit as a fiddle. This discloses
to us why the model may confound these two letters, as a W
resembles a rehashed V.
Fig. 5. Accuracy rate of Digits
E. Error Rate of Digits
Figure 6 shows the error rate of digits. The error rate for
digits lies between 0% to 50%. Some digits have 0% error
rate as the system predicts them correctly.
Fig. 3. Accuracy rate of Alphabets
H. Comparison of Recogniton and Error rate of Alphabets
Figure 9 shows the comparison of recogniton and error rate
of alphabets. In the Graph, blue line shows the accuracy rate of
alphabets which is different for every alphabet. Some alphabets
have the accuracy rate of 100% while some have the accuracy
rate of 50% to 90%. Red line shows the error rate of alphabets
which is between 0% to 50%.
Fig. 6. Error rate of Digits
F. Accuracy Rate of Specific Words
Figure 7 shows the accuracy rate of specific words. Five out
of seven specific words have the accuracy rate of 100% while
other two specific words have the accuracy rate of 50% and
60%. This system is good at predicting the correct words as
shown in figure.
Fig. 9. Recogniton and Error rate of Alphabets
I. Comparison of Recogniton and Error rate of Digits
Figure 10 shows the Comparison of Recogniton and Error
rate of digits. In the Graph, blue line shows the accuracy rate
of digits which is different for every digit. Some digit have
the accuracy rate of 100% while some have the accuracy rate
of 50% to 90%. Red line shows the error rate of digits which
is between 0% to 50%.
Fig. 7. Accuracy rate of Specific Words
G. Error Rate of Specific Words
Figure 8 shows the error rate of specific words. Only two
specific words have the error rate of 40% and 50% while other
words have 0% error rate as the system predicts them correctly.
Fig. 10. Recogniton and Error rate of Digits
J. Comparison of Recogniton and Error rate of Specific Words
Figure 11 shows the comparison of recogniton and error rate
of specific words. In the Graph, blue line shows the accuracy
rate of specific words five out of seven specific words have
the accuracy rate of 100%, while only two specific words have
error rate of 40% and 50% which is shown by the red line.
Fig. 8. Error rate of Words
Fig. 11. Recogniton and Error rate of Specific Words
V. C ONCLUSION
Proposed work aims to make communication simpler between hearing impaired people and normal people by introducing computer in communication path so that signs can
be captured, recognized, translated to text and displayed it
on screen. Sign Language Reconition (SLR) system is implemented by using OpenCV, Tensorflow, Keras and CNN
model.Deep Learning is turning into an exceptionally famous
subset of AI because of its significant level of execution
across numerous sorts of information. An incredible method to
utilize deep learning is to figuring out how to classify pictures
to manufacture a convolutional neural Network (CNN). The
Keras library in Python makes it really easy to fabricate a
CNN. Keras is a ground-breaking and simple to-utilize free
open source python for making and evaluating profound learning models. It wraps the powerful numerical count libraries
Theano and TensorFlow and grants you to plan neural association models in precise code. The principle goal of proposed
work is to plan a framework that makes an interpretation
of static hand signs into English text continuously, utilizing
webcam of PC. By utilizing Convolutional neural organization
the propose framework can perceive static signs upto 80%
accuracy. This framework can also be additionally enhanced
by making dataset more diverse.
Download