Deep convolutional neural network for hand sign language recognition using model E

Bulletin of Electrical Engineering and Informatics
Vol. 9, No. 5, October 2020, pp. 1873~1881
ISSN: 2302-9285, DOI: 10.11591/eei.v9i5.2027  1873
Journal homepage: http://beei.org
Deep convolutional neural network for hand sign language
recognition using model E
Yohanssen Pratama, Ester Marbun, Yonatan Parapat, Anastasya Manullang
Faculty of Informatics and Electrical Engineering, Institut Teknologi Del, Indonesia
Article Info ABSTRACT
Article history:
Received Nov 3, 2019
Revised Feb 3, 2020
Accepted Mar 15, 2020
An image processing system that based computer vision has received many
attentions from science and technology expert. Research on image processing
is needed in the development of human-computer interactions such as hand
recognition or gesture recognition for people with hearing impairments
and deaf people. In this research we try to collect the hand gesture data
and used a simple deep neural network architecture that we called model E to
recognize the actual hand gestured. The dataset that we used is collected
from kaggle.com and in the form of ASL (American Sign Language)
datasets. We doing accuracy comparison with another existing model such as
AlexNet to see how robust our model. We find that by adjusting kernel size
and number of epoch for each model also give a different result.
After comparing with AlexNet model we find that our model E is perform
better with 96.82% accuracy.
Keywords:
Convolutional neural network
Gesture
Hand recognition
Hand sign
Image processing
This is an open access article under the CC BY-SA license.
Corresponding Author:
Yohanssen Pratama,
Faculty of Informatics and Electrical Engineering,
Institut Teknologi Del,
Jl. Sisingamangaraja, Sitoluama, Toba Samosir, Sumatera Utara, Indonesia.
Email: yohanssen.pratama@del.ac.id
1. INTRODUCTION
Based on the disability data from the Ministry of Social Affairs in 2012, they figures that in
33 provinces of Indonesia, 223,655 people were deaf (around 10.52%), 151,371 were speech impaired
(around 7.12%), and were deaf and speech impaired (dumb deaf) as many as 73,560 (around 3.46%) [1].
If calculated, around 21.10% of Indonesia's population experiences hearing and speech problems. This is
very concerning because around 21.10% of the population in Indonesia has difficulty in communicating with
normal people, as well as to their fellow deaf and speech impaired people. This difficulty is caused by
the inability of the other person to recognize sign language issued by speech impairments. If you want to
recognize sign language from the communicator (speech impaired), then the communicant must learn
and train themselves to understand the meaning of the sign language [2]. This is the reason why the research
team feels the need to make a technology that can help deaf and speechless people communicate with each
other as well as normal people.
At present the development of computer technology is rapid, making computing capabilities on
computers also increase [3]. Machine learning technology has many benefits for people's lives in different
aspects [4-6]. Machine learning technology is used to identify objects in an image, conversations into text,
match news, upload products according to user interests, and select relevant results on a search [7]. Some of
the algorithms/models used in machine learning technology include neural network and deep learning.
Learning representation (representation learning) is a method that allows a machine to process raw data
and automatically find the representation needed for detection or classification. Deep learning is a method of
learning representation that allows computational models that are composed of many layers of processing to

 ISSN: 2302-9285
Bulletin of Electr Eng & Inf, Vol. 9, No. 5, October 2020 : 1873 – 1881
1874
learn data representation with many levels of abstraction [8]. This method significantly increases performance
in the introduction of conversations, visual object recognition, object detection and so on [9]. One of the deep
learning methods that are often used for image recognition is deep convolutional neural networks.
Hand gesture recognition technology made based on image processing and computer vision.
Computer vision is a system in image processing that obtained an image from electronic cameras and similar
to human vision systems where the brain processes images from the eye. To enhance raw image from camera
so we can improve the pictorial information we need an image processing [10]. At present, computer vision
is a topic that is widely discussed by us originating from electrical engineering, computer scientist, etc.
This system is used in many ways, such as checking object size, food quality, detecting faces automatically,
recognizing humans through iris, etc. The technology that will be developed by the research team to help
speech impaired, deaf and normal people to communicate is a technology that applies computer vision,
namely hand recognition using the CNN algorithm. The purpose of hand recognition is to provide natural
interactions between humans and computers that can process data and then process it into information.
Of course, this is very useful for the speech impaired people and they companion because communicants can
understand sign language delivered by communicators (speech impaired). Unlike the hearing aids that have
been circulating in the market such as SIGNLY (gloves for detecting movement and finger positioning),
Kinect, Albab, etc., the technology that will be developed by the research team is technology that will change
sign language from speech impaired people becomes a sentence that can be understood by opponents
of communication.
2. RESEARCH METHOD
Convolutional neural network (CNN) is an approach that builds the invariance properties into
the structure of a neural network and it was based a neocognitron model which is also hierarchical
and multilayered structured [11-13]. CNN have succeeded in the problem of image recognition
and classification, and have been successfully implemented for the introduction of human body movements
in recent years. In particular, experiments have been carried out in the field of introduction of sign language
using CNN, with the condition of inputs that are sensitive to background change. By using a camera that can
detect the depth and contours, the process of recognizing sign language is made easier through the development
of depth characteristics and movement of profiles. For each sign language introduction, the use of sensing
technology is rapidly gaining popularity, and other tools have been incorporated into processes that have
proven successful. Developments such as specially designed color gloves have been used for hand recognition
processes and make feature extraction steps more efficient by making certain gesture units easier to identify
and classify [14, 15].
However, until now, the method of introducing automatic sign language cannot utilize sensing
technology that is usually carried out daily. Previous works used camera technology that was very basic to
produce simple image data sets, without information on depth or contours available, only pixels exist. Efforts
to deal with the task of classifying several ASL letter movement images have been successful but using
the AlexNet architecture that has been previously trained.
In this paper we proposed a simple model E architecture to detect the hand gesture and varying
the filter size for the input. The Model is based on traditional convolutional neural network which being
used for many applications such as house numbers digit classification and has a good performance [16].
The architecture could be seen in Figure 1, that from the first convolution layer until the fourth we modified
the filter size (1x1, 3x3, 9x9, and 11x11) and also we did the experiment that using 5x5 filter sizes for all
convolution layer except the fourth layer that using 4x4 filter size. We have model E with different filter
there are: model E which have 1x1 filter size, model E with 3x3 filter size, model E with 9x9 filter size,
and model E with 11x11 filter size. Also we have model E with 5x5 filter size except its fourth layer with
4x4 filter size. In this research, we will see which filter gives the best accuracy result by doing some test
experiments. After that we compared our model with previous AlexNet architecture to get accuracy. Estimating
the accuracy of our model is central for this research, so we can measure our system performance [17].
In this research we use a collection of SIBI (Indonesian Language Signal System) dataset that taken
from online datasets sourced kaggle.com and in the form of ASL (American Sign Language) datasets which
consisting of 29 objects, 26 letters (a-z), nothing, delete, and space [18]. This is because SIBI has adopted
ASL to become a standard sign language and standard in Indonesia. In carrying out this research,
we conducted several steps as follows:

Bulletin of Electr Eng & Inf ISSN: 2302-9285 
Deep convolutional neural network for hand sign language recognition... (Yohanssen Prataman)
1875
Font
RGB (Red, Green,
Blue)
Grayscale Binary
F
P
R
Figure 1. Model E architecture with 1x1 filter size
2.1. Collecting a dataset
The SIBI dataset, shown in Figure 2 (a) that has been obtained divided into two categories with
a ratio of 80:20 using holdout method in cross validation. The data is consist of 80% training dataset and 20%
testing dataset. Furthermore, the data were analyzed for physical forms such as file size, pixel number, image
clarity (noisy) to facilitate the training process and data validation test. After that in Figure 2 (b), the colored
dataset (RGB/red green blue) undergoes a color transformation stage to grayscale (gray) and then undergoes
transforms again to black and white (binary image) to separate between background and foreground image
using a background subtraction [19, 20]. Here is the SIBI dataset that we used:
Figure 2. Dataset, (a) SIBI dataset, (b) Conversion of RGB images to binary images
2.2. Image preprocessing
In this stage we try to convert images from RGB to binary color space. Image preprocessing is done
to improve the computer accuracy when recognizing images and the time efficiency to recognize images.
In the image preprocessing stage, several steps are carried out such as:
‒ Segmentation: good segmentation process leads to perfect feature extraction process and the later play
an important role in a successful recognition process [21, 22]. So for segmentation we use thresholding
method to find the pattern and the background image. The value for the background image is 0
and foreground (hand) image is 1 as shown in Figure 3.
(a) (b)

 ISSN: 2302-9285
1876
(a) (b)
Figure 3. Thresholding process, (a) Thresholding output, (b) Output binary representation
‒ Resizing: resizing to help image processing performance where the image size is adjusted to the standard
height and width of the source image input. Each hand dataset could have a different resolution, so to
maintain the input size we need to standarize the image resolution for 100x100 pixels.
‒ Rescaling: re-scaling to prevent image distortion.
‒ Noise removal: use an efficient method of removing noise from image, so we will get the better analysis
result from the system. In this research we use BM3D filter [23].
2.3. Modeling and training
At this stage, we built a simple model called Model E and the processing is done by using NVIDIA
Geforce 930mx software that is operating under the Windows 10 Pro operating system. This model consists
of 4 convolutional layers, 4 max pooling 2D pieces, and 1 fully connected layer. We use 4 max pooling
because it can greatly improve the statistical efficiency of the network [24]. The training process uses a 5x5
matrix filter on the first layer to the third layer and uses a 4x4 filter in the fourth layer, shown in Figure 4.
Furthermore, the number of epochs used is 150 with a step count of 1000 and this is done to produce high
data accuracy. We take conventional approach to look for similar problems and deep learning architectures,
which have already been shown to work, because there is no general answer to determine hyperparameter
such as kernel size, output maps, and CNN layers. Then a suitable architecture can be developed by
experimentation. The model E is a simple model of conventional CNN that improve based on experiment.
If we compare a smaller filter size to larger filter size, the larger one extracted quite generic features that
spread across the image and capture the basic component of image. Model E use larger filter size from 1st
to
3rd
layer because the hand sign does not have a complex features, although the amount information that
extracted is lesser but it will use lesser memory for backpropagation. In the 4th
layer we used smaller filter
because we will have a better weight sharing.
Figure 4. Model E architecture with 4x4 filter size in 4th
convolutional layer
2.4. Testing
After the model has been built and the data is already passed the training phase, then the next step is
testing. At this stage we tried to classifying test images using 2 models with different filters. A common
problem that found during training neural networks is the choice of number training epochs that will used.
Too many epochs can lead to overfitting of the training dataset, whereas too few may result in an underfit
model [25]. So we try various number of epochs and see when the model performance stops improving.
We choose the number of epoch based on the best model performance.

1877
3. RESULTS AND DISCUSSION
The results of this research will consist of 3 main parts, there are data preparation stages, modeling
stages, and testing stages.
3.1. Data preparation
The dataset is consist of 3000x29 imagery of hand gesture language. The image size is converted to
100x100 pixels with the binary format and stored in .png format.
3.2. Modeling
The experiment is conducted by using model E which uses several types of filter sizes. Model E
implement the stochastic gradient descent (SGD) method with learning rate=0.01 and 50 epochs.
Then the highest result of the fifth filters will be compared to AlexNet model. The graph of the training
process can be seen in the Figure 5. From Figure 5 within the five filters, filter with combination 4x4 and 5x5
kernel size has the best accuracy when the epoch reaches 50.
(a) (b)
(c) (d)
(e) (f)
Figure 5. Experiment results based on the dataset type, (a) Model E with filter 1x1, (b) Model E with filter
combination 4x4 and 5x5, (c) Model E with filter 3x3, (d) Model E with filter 11x11,
(e) Model E with filter 9x9, (f) Model E AlexNet with filter 9x9

 ISSN: 2302-9285
1878
3.3. Testing
In our experiment, we have succeeded in classifying images that taken from the dataset. The image
data was taken from ASL (American Sign Language) hand motion images.
3.3.1. Dataset test
The results of our experiment can be seen in Table 1 which is the experimental results based on
the different dataset type. In Table 1 can be seen that there are three types of color spaces that are compared,
there are RGB (red, green, blue), grayscale (gray), and binary (black and white). In this comparison,
the parameters that we used has the same value, such as the kernels/filters having a matrix of size 1x1, epoch
numbers are 100, and the total steps are 1000. If we sorted by the accuracy of the training data, the highest
accuracy is in the RGB type with values 0.8759 and the lowest accuracy is in the Binary type which
is 0.6776. However, at this initial comparison stage, we only take the lowest time/duration during
the computational process to produce fast output without considering the accuracy of each type of dataset.
Table 1. Experiment results based on the dataset type
Dataset type Computing time duration (hour) Kernel Epoch Step Accuracy Loss
Grayscale 49.52 1x1 100 1000 0.8538 0.3961
Binary 30.49 1x1 100 1000 0.6776 0.9380
RGB 42.84 1x1 100 1000 0.8759 0.4316
In this comparison process, the kernel was intentionally implemented following the SqueezeNet
model measuring 1x1 to speed up the computing [26]. After the training phase is done, it turns out that
the type of dataset that has the fastest computing duration is the “binary type” with an execution time for
30.49 hours and the duration of computing is longer for the grayscale type which is for 49.52 hours. Based on
the results that already obtained, we decided to use binary data types to speed up the process of training
and testing of sign language data.
3.3.2. Filter size test
Next, the dataset in the form of binary is compared again to determine the most effective filters to
apply in the process of recognizing sign language. In Table 2 we could see an accuracy comparison based on
filter size. From Table 2 can be seen that the number of the epoch, step, and convolutional layers that we
used has to be in the same amounts. The epochs numbers are 50, steps are 10, and convolutional layers are 4.
This experiment was done to test the accuracy of the model with a different number of filters. The difference
in filter size can affect the accuracy of training data and testing data. This is proofed by the difference in
accuracy that we acquired. The highest level of accuracy is in the 5x5 filter combined with 4x4 filters, which
are equal to 0.19, while the lowest level of accuracy is in the 11x11 filter, which is equal to 0.09. Based on
these results, we decided to use a 5x5 filter combined with a 4x4 matrix.
Table 2. Experiment results based on filter size
Conv. layer Kernel Epoch Learning rate Step Accuracy Loss
4 1x1 50 0.01 10 0.0938 3.2911
4 5x5 (3 layer) +4x4 50 0.01 10 0.1938 2.9334
4 3x3 50 0.01 10 0.1156 3.2564
4 11x11 50 0.01 10 0.0906 3.2578
4 9x9 50 0.01 10 0.0305 3.3674
3.3.3. Epoch size test
After do some experiment for dataset type and filter size, we also conducted a comparison for
the number of epochs to obtain the highest accuracy from the training and testing data. The results of this
experiment can be seen in Table 3.
Table 3. Experiment results based on epoch amount
Conv. layer Kernel Epoch Learning rate Step Accuracy Loss
4 5x5 (3 layer) +4x4 50 0.01 150 0.7388 0.7651
4 5x5 (3 layer) +4x4 100 0.01 150 0.8596 0.4024
4 5x5 (3 layer) +4x4 150 0.01 150 0.9379 0.1777
4 5x5 (3 layer) +4x4 200 0.01 150 0.8360 0.4860
4 5x5 (3 layer) +4x4 250 0.01 150 0.8981 0.2879

1879
In Table 3, it can be seen that the number of convolutional layers, steps, and filters are in the same
size, the convolution consists of four layers, 1000 total steps, and 5x5 filters size for the first into third
convolutional layer and 4x4 sized matrices in the fourth convolutional layer. Based on these results,
the highest accuracy is at the 150th epoch with an accuracy rate of 0.9379 and the lowest accuracy is at
the 50th epoch with an accuracy rate of 0.7388. The higher the number of epochs that are applied during
the training and testing process, the higher the level of accuracy of the data, so that we decide to use epoch
which amounts to 150 to be used as parameters when performing the recognition process.
3.4. Discussion of differences in 2 models
In determining whether model E is a suitable and relatively efficient model for performing
recognition processes on SIBI, we compare model E with a pre-existing model and are generally used for
recognition, the AlexNet model. The AlexNet model is often used in the process of hand gesture recognition
for sign language. This is due to several advantages of AlexNet such as simple, using the CNN architecture
that is basic, and effective. An accuracy and loss comparison of model E with the trained Alexnet model can
be seen from the following TensorBoard graph in Table 4.
Table 4. Comparison of model E and AlexNet
Aspect Model E Model AlexNet
Accuracy
Loss
Both models use the same number of epochs as 50, step 150, and learning rate 0.01, but use different
numbers of convolutional layers, five layers in the AlexNet model and four layers in model E. The different
layers make the AlexNet model require longer computation time to do the training process compared to
model E, where AlexNet requires 1.72 hours while model E takes 0.71 hours.
Comparison of model E with the AlexNet model can be seen that the performance of model E is
more efficient than the AlexNet model because in model E, accuracy tends to rise from the lower left to
the lower right, while in the AlexNet model accuracy tends to increase and decrease in succession. In
the AlexNet model, the accuracy of the 4th
epoch increase in training reaches accuracy at 0.0531
and decreases at 0.02 at the 5th
epoch, and at the 7th
epoch decreases to 0.01. Then, at the 48th epoch,
it increases to 0.0625, drops back to the 49th
epoch to 0.0438, and drops again to 0.0250 at the 50th
epoch.
In model E, the training accuracy is 19.38% and validation accuracy is 30.94%. Then, training loss
is 2.9334 and the validation loss is 2.4456. In the AlexNet model, the training accuracy is 2.50%
and validation accuracy is 3.28%. Then, the training loss is 3.3700 and the validation loss is 3.3669. Based on
these data, the model E has better prediction compare to AlexNet Model because the total loss in model E is
smaller than the AlexNet model.
Based on the results of a comparison between model E and AlexNet, we decided to use model E in
classifying and predicting sign language (SIBI). Then, we optimizes the model again by increasing
the number of epochs to 100, the stepper epoch being 960, and the validation step being 2600, resulting in

 ISSN: 2302-9285
1880
a 96.83% accuracy. This quite high result indicates that the model has a good learning ability so that it has
a good level of prediction. The model E that has been improved is then used by us to recognize sign
language, and then combine the results into a sentence.
4. CONCLUSION
The model E is the model that based on simple CNN architecture and uses a number variation
of filter size. The model E is proven to have a better performance with larger filter size from 1st
to 3rd
layer
because the hand sign does not have a complex features, although the amount information that extracted is
lesser but it will use lesser memory for backpropagation. In the 4th
layer the smaller filter is being used here
because it will have a better weight sharing. For the epoch variable already tried a various number of epochs
and seen when the model performance stops improving, for model E the number epoch that give the best
performance is 150. As the result we suggest to use 150 number of epoch for hand gesture experiment with
model E. Moreover, CNN can be managed to recognize or classify hand gestures especially in the Indonesian
Language Signal System with accuracy of 96.82% by using model E. With the same number of epoch for
AlexNet model we found our model E was perform slightly better with delta 16.88%. For further research,
we will try a model that can recognize sign language by observing static to dynamic gestures. So the gesture
that observed from the camera could directly recognize into a language.
ACKNOWLEDGEMENTS
This work was supported in part by Institut Teknologi Del.
REFERENCES
[1] Ministry of Social Affairs Republic of Indonesia, “Ministry of social affairs in social welfare development
numbers,” in Bahasa “Kementrian sosial dalam angka pembangunan kesejahteraan sosial,” Jakarta, 2012.
[2] V. Priyadharshni, M. S. Anand, and N. M. Kumar, “Hand gesture recognition system using hybrid technology for
hard of hearing community,” Int. J. of Eng. Mathematics & Computer Science, vol. 1, no. 2, pp. 50-63, 2013.
[3] A. Chaudhary, J. L Raheja, K. Das, and S. Raheja, “Intelligent approaches to interact with machines using hand
gesture recognition in natural way: A survey,” Int. J. of Comp. Sci. & Eng. Survey, vol. 2, no. 1, pp. 122-133, 2011.
[4] P. Yohanssen, I. G. B B. Nugraha, and E. T. Samosir, “Vehicle counting and classification for traffic data
acquisition,” Jurnal Teknologi, vol.78, no. 6-3, pp. 77-82, 2016.
[5] P. Yohanssen and P. R. Puspoko, “The addition symptoms parameter on sentiment analysis to measure
public health concerns,” TELKOMNIKA Telecommunication Computing Electronics and Control, vol. 15, no. 3,
pp. 1301-1309, 2017.
[6] P. Yohanssen and S.S. Hadi, “Cassava quality classification for tapioca flour ingredients by using ID3 algorithm,”
Indonesian Journal of Electrical Engineering and Computer Science, vol. 9, no. 3, pp. 799-805, 2018.
[7] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp.436-444, 2015.
[8] M. Tanaka and M. Okutomi, “A novel inference of a restricted boltzman machine,” 22nd International Conference
on Pattern Recognition, pp. 1526-1531, 2014.
[9] A. Deshpande, “A begginner's guide to understanding convolutional neural networks,” Retrieved March, vol. 31, 2016.
[10] B. Chitradevi and P. Srimathi, “An overview on image processing techniques,” International Journal of Innovative
Research in Computer, vol. 2, no. 11, pp. 6466-6472, 2014.
[11] C. M. Bishop, “Pattern recognition and machine learning,” Springer, pp. 227-229, 2006.
[12] L. V. Fausett, “Fundamental of neural networks: Architecture, algorithm, and application,” Prentice-Hall Inc., 1994.
[13] K. Fukushima, “Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition
unaffected by shift in position,” Biol. Cybernetics, vol. 36, pp. 192-202, 1980.
[14] J. F. Paul, D. Seth, C. Paul, and J. G. Dastidar, “Hand gesture recognition library,” Int. J. of Science and Applied
Information Technology, vol. 3, no. 2, pp. 44-50, 2014.
[15] Y. Xu and Y. Dai, “Review of hand gesture recognition study and application,” Contemporary Engineering
Science, vol. 10, no. 8, pp. 375-384, 2017.
[16] P. Sernamet, S. Chintala, and Y. LeCun, “Convolutional neural networks applied to house numbers digit
classification,” Proc. Of the 21st International Conference on Pattern Recognition, pp. 3288-3291, 2012.
[17] E. Platanios, H. Poon, T. M. Mitchell, and E. J. Horvitz, “Estimating accuracy from unlabeled data: A probabilistic
approach,” 31st Conference on Neural Processing Systems, (NIPS 2017), pp. 4361-4370, 2017.
[18] J. Davis and M. Shah., “Recognizing hand gestures,” European Conference on Computer Vision, pp. 331-340, 1994.
[19] R. Suguna and P. S. Neethu, “Hand gesture recognition using shape features,” Int. J. of Pure and Applied
Mathematics, vol. 117, no. 8, pp. 51-54, 2017.
[20] S. D. Badgujar, G. Talukdar, O. Gondhalekar, and S. Y. Kulkarni, “Hand gesture recognition system,” Int. J. of
Scientific and Research Publications, vol. 4, no. 2, pp. 1-5, 2014.
[21] R. Z. Khan and N. A. Ibraheem, “Hand gesture recognition: A literature review,” Int. J. of Artificial Intelligence &
Application, vol. 3, no. 4, pp. 161-174, 2012.

1881
[22] X. Li, “Gesture recognition based on fuzzy c-means clustering algorithm,” Department of Computer Science,
University of Tennessee Knoxville, 2003.
[23] M. Verma and D. J. Ali, “A comparative study of various types of image noise and efficient noise removal
techniques,” International Journal of Advanced Research in Computer Science and Software Engineering, vol. 3,
no. 10, pp. 617-622, 2013.
[24] I. Goodfellow, Y. Bengio, and A. Courville, “Deep learning,” MIT Press, pp. 204-210, 2016.
[25] F. Chollet, “Deep learning with python,” Manning Publications, 2018.
[26] W. Rawat and Z. Wang, “Deep convolutional neural networks for image classification: A comprehensive review,”
Neural Computation, vol. 29, no. 9, pp. 2532-2449, 2017.
BIOGRAPHIES OF AUTHORS
Yohanssen Pratama, S.Si., M.T. Current Faculty Members & Researcher in Del Institute of
Technology. 4+ years experience specializing in back-end/infrastructure, analytical tools
development and computer programming. Teach academic and vocational subjects to
undergraduate also pursue my own research to contribute to the wider research activities of my
department.
Ester Enjela Marbun, S.Tr.Kom Currently, work as a Software Engineer at PT. Citra Global
Dinamika (Escurity), Batam Center. Have experience at image processing using neural network,
automation testing, web & desktop developing, and digital designing.
Yonatan Vikario Resha Parapat, S.Tr.Kom Currently, work as a IT Consultant Analyst at
Mitra Integrasi Informatika( subsidiary of PT. Metrodata Electronics Tbk.) APL Tower 37th
Floor Suite 3 Jl. Letjen S. Parman Kav. 28, RT.13/RW.7, Jelambar Baru, Grogol petamburan,
RT.10, RT.12/RW.6, Tj. Duren Sel., Kec. Grogol petamburan, West Jakarta. Have experience at
image processing using neural network, Networking, Web PenTest, and Mobile developing.
Anastasya Pehulisa Manullang, S.Tr.Kom Currently, work as a Product Owner at PT.
Moonlay Technologies, Sudirman Central Business District, South Jakarta. Have experience at
image processing using neural network, automation testing, web & desktop developing, digital
designing and system analysis.

Deep convolutional neural network for hand sign language recognition using model E

More Related Content

What's hot (19)

Similar to Deep convolutional neural network for hand sign language recognition using model E (20)

More from journalBEEI (20)

Recently uploaded (20)

Deep convolutional neural network for hand sign language recognition using model E