2023 IEEE International Students’ Conference on Electrical, Electronics and Computer Science

Deep Learning Approach for Sign Language

Recognition using DenseNet201 with Transfer
2023 IEEE International Students' Conference on Electrical, Electronics and Computer Science (SCEECS) | 979-8-3503-9874-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/SCEECS57921.2023.10063044

Learning
Yasir Altaf Abdul Wahid
Department of ComputerScience Department of Computer Science and Infromation Technology
Maulana Azad National Urdu University Maulana Azad National Urdu University
Hyderabad, India Hyderabad, India
yasiraltafwani66@gmail.com wahidabdul76@yahoo.com

Mudasir Manzoor Kirmani

Department of Computer Science, Division of Social Science
FoFy SKUAST-K
Srinagar, India
mmkirmani@gmail.com

Abstract—Transfer learning has been utilized to solve many people, (b) assisting speech and hearing loss people in their
complex real-world problems. Over the last several years, trans- daily life activities, such as studying, learning, travelling,
fer learning had many applications in image and video recogni- marketing, banking, and health quires; (c) accelerating the
tion. To get higher recognition rates, deep and wider architectures
of the convolutional neural networks (CNN) have been designed. teaching learning at schools meant for speech and hearing-
In this research, we proposed a novel transfer learning-based impaired students, and (d) improving the digital literacy among
model using a popular CNN architecture called DenseNet201 for deaf-mute people and also improving their lifestyle with
the recognition of Indian Sign Language (ISL) hand gestures. We enhanced communication and understanding. A number of
applied transfer learning to DenseNet201 by freezing some of its automatic sign language recognition and translation systems
layers to retain its knowledge of generalization and fine-tuning
the remaining layers for ISL dataset. Pre-trained DenseNet201 have been developed over the last several years to meet the
was used to extract the features of the gesture images. To classify communication needs of deaf and hearing-impaired individuals
the ISL gesture, custom layers were added to the pretrained [1], [2].This area is still an active area for research and requires
DenseNet201 model. The proposed model helped to achieve improvements to implement such systems in various sectors,
higher accuracy of 100%. such as education to teach and assist speech and hearing
Keywords—Transfer Learning, DenseNet201, Deep Learning,
Sign Language Recognition, Multilayer DenseNets, Image recog-
loss students, public places such as railway stations, airports,
nition. and bus stops to guide deaf-mute people and help them
by understanding their quires, providing feedback to them,
and bringing them closer to the hearing world [3]–[5].With
I. I NTRODUCTION
advancements in technologies researchers have adopted many
The language used by deaf and mute people is called approaches for sign language recognition, such as sensory-
sign language (SL). It is used to propagate messages among based, Vision-based and deep learning-based approaches are
them and also used by them to communicate with normal widely used. Traditionally, most sign language recognition and
people. In every country, people with speech and hearing translation systems utilize contact-based approaches, wherein
loss have their own ways of expressing ideas using sign the signer needs some sort of special wearable device (like
language. For example, people living in America use American glove equipped with sensors) to perform gestures, and such
Sign Language (ASL) and the people of India use Indian systems did not grow in large scale for the development of dig-
Sign Language (ISL). Similarly, all other countries have their ital assisting applications for deaf-mute people because of their
own sign language. However, due to the varied nature of obvious limitation with the hard-to-use circuitry of wearable
sign language across countries, people face difficulties in devices. To simplify such systems, vision-based approaches
exchanging their ideas when it is communication between two provide new, easy-to-use, robust, and adequate methods for
deaf-mute persons or between normal and deaf-mute persons. large-scale development of sign language-based assisting sys-
To bridge this communication gap, technology known as Sign tems. Vision-based systems for sign language recognition and
Language Recognition and Translation plays an important translation are easy to use and do not require any hard-to- use
role. The various benefits of such a technology are: (a) wearable devices; instead, they utilize high-quality cameras
removing the communication barrier between the deaf-mute to capture the signs performed by the signer. In recent years,

979-8-3503-9874-8/23/$31.00 © 2023 IEEE

Authorized licensed use limited to: M S RAMAIAH INSTITUTE OF TECHNOLOGY. Downloaded on March 10,2025 at 08:23:25 UTC from IEEE Xplore. Restrictions apply.
 computer vision-based sign language recognition systems have the use of deaf communities in their work environment. This
achieved major advancements with the help of deep learning, system is able to recognize 25 different signs of Croatian sign
and many companies and researchers have developed various language commonly used in home environments. With recent
applications and digit-assisting devices for deaf-mute people. advancements, Wi-Fi based technologies have been adopted to
With recent advancements of deep learning in machine learn- enable pervasive and device-free gesture recognition [22]; such
ing, different convolutional neural network architectures have systems use off-shelf Wi-Fi devices for gesture recognition and
been employed for sign language recognition to achieve higher support assisted living. Recently, deep learning-based methods
recognition accuracies [6].In this research, we propose a deep- [23]–[26] have been incorporated with Wi-Fi based technology
learning-based method for the recognition of static gestures to improve the performance of sign language-based assistive
of numbers and letters in the ISL dataset. The proposed systems.
method uses DenseNet201 and transfer learning techniques
to achieve higher recognition accuracies than most studies in III. P ROPOSED W ORK
the literature. The remainder of this paper is organized as in this work, we applied transfer learning to DenseNet201
follows. Section II presents a brief summary of various related for the recognition of ISL hand signs (as described in Section
works. Section III describes the proposed work, DenseNet B). Transfer Learning is one of the most useful and widely
architecture, and transfer learning using DenseNet201. Section used techniques in deep learning, which helps transfer the
IV presents the dataset and data processing technique used. knowledge of the model already acquired while solving any
Section V provides experimental results and evaluation metrics other relevant problem to solve a new problem. Transfer
used for the performance analysis of the prosed model. Finally, learning not only saves time and computational resources but
section VI presents the conclusion and future works. also achieves higher accuracy when solving new but relevant
problems.
II. R ELATED W ORK
A. DenseNets
For sign language recognition and sign language based
assistive technology for deaf and mute people, many re- DenseNet is a convolutional neural network with dense
searchers worldwide have worked on different sign languages connectivity between layers; that is, there are direct feed-
such as Indian [7]–[9], American [10], [11], Mexican [12], forward connections from each layer to every other layer (as
Chinese [13], [14], Brazilian [15], and Arabic [16]. And have shown in Fig. 1). Therefore, the number of connections in
developed various deep learning-based methods to support a densenet of n layers is n(n + 1)/2 [27]. The key benefits
efficient and effective gesture recognition. Tang et al. [17] of DenseNets are: the resolution of vanishing gradients, deep
proposed a two-stage hand-gesture recognition system using supervision, and regularization to minimize overfitting that
a Kinect sensor and deep neural networks. The system used occurs due to the availability of small training datasets [28].
DNNs for automatic feature extraction and recognition of The dense connectivity in DenseNet helps to extract the
hand postures and achieved a recognition accuracy of 98.12%. deep features from an input image with the help of feature
Microsoft Kinect and a convolution neural network were used concatenation, that is, each layer receives the feature maps
in [18] to develop a system for the recognition of 20 Italian from its preceding layers. Mathematically, it is defined as:
gestures. This system is able to generalize different users
Ai = Hi ([A0 , A1 , . . . , Ai−1 ]) (1)
and backgrounds data irrespective of the data that occurred
during training. A validation accuracy of 91.7% was achieved where, [A0 , A1 , . . . , Ai−1 ] represents the concatenation of the
by using this system. To deal with long-term motions (i.e., feature maps of all preceding layers of the ith layer, and
motions or actions that can span over a relatively long period Hi is a composite function of BN (branch normalization),
of time), Duan et al. [19] presented a two-stream framework Relu (rectified linear unit) and conv (a 3 × 3 convolution).
for gesture recognition using a convolution neural network The feature map concatenation in DenseNet enhances the
with a consensus voting mechanism. The framework achieved information flow between layers and consequently improves
better results than those its closest competitors. Chong et al. [6] its generalization abilities. Different variants of DenseNet have
presented a deep-learning-based approach for the recognition been used in several studies [29], [30]. In this study we used
and interpretation of American signs. A wearable IMU sensory one of its architectures called DenseNet201 for the recognition
device was used to capture gestures, which were classified of ISL hand signs.
using the LSTM model. This model was used to recognition
28 commonly used words of the American sign language with B. Transfer Learning using DenseNet201
an accuracy of up to 99.89%. To reduce the response time As previously mentioned, in this study, DenseNet201 was
while recognizing hand gestures and converting them into used for the recognition of ISL gestures using transfer learn-
audible sound using WebRTC, Gupta et al. [20] presented a ing. The DenseNet201 architecture (shown in Fig. 2 con-
hybrid CNN-LSTM model. This web-based model acts as a sists of 708 layers) is a deep, condensed, and parametrically
media for deaf-mute people to share their ideas with normal compressed model because it contains transition layers that
people. Another sign language recognition system based on the are used to reduce the spatial size of the input images and
3DCNN architecture was presented by Kraljevic et al. [21] for reduces their feature maps to a smaller number. Another

SCEECS 2023
Authorized licensed use limited to: M S RAMAIAH INSTITUTE OF TECHNOLOGY. Downloaded on March 10,2025 at 08:23:25 UTC from IEEE Xplore. Restrictions apply.
 Fig. 1: A three-layer DenseNet architecture.

Fig. 3: Architecture of the proposed model.

advantage of DenseNet201 is that it is easy to train and
achieve efficient parameter optimization through feature reuse.
We used DenseNet201 to extract the features of the gesture IV. DATA AND DATA P ROCESSING
images using its learned weights on the ImageNet dataset by The ISL hand-gesture image dataset available on Kaggle
freezing its layers. Finally, five custom layers, including two (https://www.kaggle.com/datasets/prathumarikeri/indian-sign-
hidden layers with 256 and 128 neurons, two dropout layers language-isl) was used in this experiment. This dataset is
with a value of 0.5, and one output layer with 35 neurons publicly available and consists of 42,000 RGB images of 35
(equivalent to the number of classes in ISL dataset), were hand signs including 9 numbers (1-9) and 26 letters (A-Z).
added, followed by softmax activation to classify the gestures. Each image had a size of 128×128×3, and we rescaled these
The overall architecture of our proposed model is shown in images to a higher resolution of 224 × 224 × 3 to increasing
Fig 3. the performance of the model with higher resolution inspired
by Thambawita et al. [31]. To validate the model, we divided
the dataset into three sets: training, validation, and testing at
a ratio 80:15:05 (3,360 images were used for training, 6,300
for validation, and 2,100 for testing). After rescaling the
image to a higher resolution, data augmentation techniques
were applied to the training set, including a shear range
of 0.2, zoom up to 0.2, rotation up to 20° and horizontal
filliping. A sample of the training data after the augmentation
is shown in Fig. 4. All images were normalized to rescale the
pixel values from the range of 0-255 to the range of 0-1 and
were converted to numpy arrays for feature extraction. The
description of the dataset used in this study is given in Table I.

Fig. 4: ISL hand gesture image dataset sample after augmen-

Fig. 2: DenseNet201 layered architecture. tation.

SCEECS 2023
Authorized licensed use limited to: M S RAMAIAH INSTITUTE OF TECHNOLOGY. Downloaded on March 10,2025 at 08:23:25 UTC from IEEE Xplore. Restrictions apply.
 TABLE I: Description of the data set used in this study
Dataset Total No. of classes No. of images in train set No. of images in val. set No. of images in test set Total no. of images in data set
ISL 35 33,600 6,300 21,00 42,000

V. R ESULTS

In this research, the proposed DenseNet201 based transfer

learning model was trained to recognize and classify ISL
hand gestures. A pretrained DenseNet201 was used to extract
the features of the gesture images. For transfer learning,
we set the trainable property of the DenseNet201 layers to
false, which prevented the weights (obtained by DenseNet201
from ImageNet) from being updated during training with the Fig. 5: Training results of the proposed model
ISL dataset. The proposed model was trained for 10 epochs,
and the validation loss was calculated for each epoch. The
model is saved to the best with minimum validation loss,
or the validation curve reaches saturation or a steady-state
condition. Training data were provided to the model in batches
with a batch size of 64. The optimization algorithm for our
proposed model was Adam, with a learning rate of 0.001. The
Adam optimization algorithm is used to optimize the ‘gradient
descent’ by combining the two gradient decent methodologies
namely ‘momentum’ and ‘RMSP’ algorithms. Mathematically,
it is defined as:
 
α
Wt+1 = Wt − m̂t √ (2)
vˆt + 

Where, Wt is the weights at time t, Wt+1 is the weights at

time t+1, m̂t & vˆt are bias- corrected weight parameters, α is
the Learning rate parameter, and  is a small positive constant
(10−8 ) to avoid ‘division by zero’ error.
The classification performance of the proposed model given
in Fig.6 was accessed using the four metrics defined in Eqs.
3, 4, 5, and 6 and visualized using the confusion matrix given
in Fig. 7. The training and validation performances of the
proposed model were visualized by plotting their graphs, as
shown in Fig. 5. The model achieved higher accuracies of
100% for both validation and test. Good values of precision,
recall, and F1-score indicate that the model performance in
classifying all classes accurately is satisfactory. A comparison
of the proposed and existing deep learning-based sign lan-
guage recognition models is presented in Table II.

TP + TN
Accuracy = (3)
TP + TN + FP + FN

TP
P recision = (4) Fig. 6: Classification report of the proposed model.
TP + FP

TP
Recall = (5)
TP + FN

2 × P recision × Recall
F 1 − score = (6)
P recision + Recall

SCEECS 2023
Authorized licensed use limited to: M S RAMAIAH INSTITUTE OF TECHNOLOGY. Downloaded on March 10,2025 at 08:23:25 UTC from IEEE Xplore. Restrictions apply.
 TABLE II: Comparison of this study with existing related studies.
Study Model Dataset Accuracy (%)
Abdul et al. [32] Inception- BiLSTM Arabic sign language (KSU-ArSL) 84.20
Molchanov et at. [33] 3DCNN VIVA challenge dataset 77.05
Islam et al. [34] VGG16 BdSL 99.92
Mistree et al. [35] MobileNet ISL 97.26
Wangchuk et al. [36] CNN BSL 97.62
Suri et al. [37] CapsNet Constructed 94
Katoch et al. [8] CNN ISL 99
Singh et al. [38] 3DCNN ISL 88.24
Dhulipala et al. [39] CNN British Sign Language 97.40
Patil et al. [40] CNN ISL 95
Sharma et al. [9] CNN ISL 99.52
Nandi et al. [41] CNN ISL 99.76
This study DenseNet201 ISL 100

Fig. 7: Confusion Matrix of the proposed model.

and the custom CNN layers for classifying these gesture

images into their appropriate classes. This model achieved
100% accuracy for the ISL dataset. The accuracy achieved by
this model is significantly higher than that of some existing
sign language recognition studies. This model proved helpful
in reducing loss during epochs without overfitting. Efficient
and highly accurate automatic sign language recognition and
VI. C ONCLUSION AND F UTURE W ORK translation systems are required at various places for deaf and
mute communities to bridge their communication gap. This
In this paper, we are dedicated to the recognition of static
model can improve the recognition rate of such a system. In
hand gestures in Indian sign language with higher accuracy.
future work, we intend to extend this model to develop an
We propose a novel transfer learning-based model using
efficient sign language-based learning assistive system for deaf
pre-trained DenseNet201 and a convolutional neural network
and mute students.
(CNN). This model utilized the pre-trained DenseNet201
model for the extraction of features from the gesture images

SCEECS 2023
Authorized licensed use limited to: M S RAMAIAH INSTITUTE OF TECHNOLOGY. Downloaded on March 10,2025 at 08:23:25 UTC from IEEE Xplore. Restrictions apply.
 R EFERENCES [22] H. Farhana, T. Ahmed, H. Ahmad, K. Narasingamurthi , H. Harkat and
S. K. Phang, ”DF-WiSLR: Device-Free Wi-Fi-based Sign Language
[1] M. Sobhan, M. Z. Chowdhury, I. Ahsan, H. Mahmud and M. K. Hasan, Recognition,” Pervasive and Mobile Computing, vol. 69, p. 101289,
”A Communication Aid System for Deaf and Mute using Vibrotactile 2020.
and Visual Feedback,” 2019 International Seminar on Application for [23] M. G. Martı́n, R. S. Segundo, F. F. Martı́nez, and J. F. López, ”Improving
Technology of Information and Communication (iSemantic), 2019, pp. physical activity recognition using a new deep learning architecture
184-190, doi: 10.1109/ISEMANTIC.2019.8884323. and post-processing techniques,” Engineering Applications of Artificial
[2] K. Yousaf et al, ”A Novel Technique for Speech Recognition and Visual- Intelligence, vol. 92, p. 103679, 2020.
ization Based Mobile Application to Support Two-Way Communication [24] Y. Ma, G. Zhou, S. Wang, H. Zhao, and W. Jung, ”SignFi: Sign Lan-
between Deaf-Mute and Normal Peoples,” Wireless Communications & guage Recognition,” in Proc. ACM Interact. Mob. Wearable Ubiquitous
Mobile Computing (Online), p. 12, 2018. Technol, 2018.
[3] E.E. Abdallah and E. Fayyoumi, ”Assistive Technology for Deaf People [25] Y. Xu, W. Yang, M. Chen, S. Chen and L. Huang, ”Attention-Based
Based on Android Platform,” Procedia Computer Science, vol. 94, pp. Gait Recognition and Walking Direction Estimation in Wi-Fi Networks,”
295-301, 2016. IEEE Transactions on Mobile Computing, vol. 21, no. 2, pp. 465-479,
[4] M.A. Azizan, I. Zulkiflee and N. Norhashim, ”Development of Sign Lan- 2022.
guage Translator for Speech Impairment Person,” in Human-Centered [26] H. Xue, W. Jiang, C. Miao, F. Ma, S. Wang, Y. Yuan, S. Yao, A. Zhang,
Technology for a Better Tomorrow, Springer Singapore, 2022, pp. 115– and L. Su, ”DeepMV: Multi-View Deep Learning for Device-Free
131. Human Activity Recognition,” in Proc. ACM Interact. Mob. Wearable,
[5] S. Sundaramoorthy and B. Muthazhagan, ”Human-Computer Interaction 2020.
System for Improving Digital Literacy Among Speech- and Hearing- [27] G. Huang, Z. Liu, L. Van Der Maaten and K. Q Weinberger, ”Densely
Impaired in Developing Countries,” in Digital Literacy and Socio- Connected Convolutional Networks,” in 2017 IEEE Conference on
Cultural Acceptance of ICT in Developing Countries, Cham, Springer Computer Vision and Pattern Recognition (CVPR), 2017.
International Publishing, 2021, pp. 139-156. [28] M. Z. Hossain, F. Sohel, M. F. Shiratuddin, and H. Laga, ”Attention-
[6] T.W. Chong and B.J. Kim, ”American Sign Language Recognition Based Image Captioning Using DenseNet Features,” in Neural Informa-
System Using Wearable Sensors with Deep Learning Approach,” The tion Processing, Cham, 2019.
Journal of the Korea institute of electronic communication sciences, vol. [29] B. Aksoy , O. K. M. Salman , and Ö. Ekrem, ”Detection of Turkish
15, no. 2, p. 291–298, 2020. Sign Language Using Deep,” Applid Artifial Intelligence, vol. 35, no.
[7] N. Musthafa and C.G. Raji, ”Real time Indian sign language recognition 12, p. 952–981, 2021.
system,” Materials Today: Proceedings, vol. 58, pp. 504-508, 2022. [30] M. Kołodziej, E. Szypuła, A. Majkowski and R. Rak, ”Using deep
[8] S. Katoch, V. Singh and U. S. Tiwary, ”Indian Sign Language recog- learning to recognize the sign alphabet,” Przeglad Elektrotechniczny,
nition system using SURF with SVM and CNN,” Array, vol. 14, p. vol. 6, p. 32, 2022.
100141, 2022. [31] V. Thambawita, I. Strümke, S.A. Hicks, P. Halvorsen, S. Parasa and M.
[9] S. Sharma, and S. Singh, ”Recognition of indian Sign Language (ISL) A. Riegler, ”Impact of Image Resolution on Deep Learning Performance
using Deep Learning Model.,” Wireless personal communications, vol. in Endoscopy Image Classification: An Experimental Study Using a
123, pp. 671-692, 2021. Large Dataset of Endoscopic Images,” Diagnostics-Basel, vol. 11, no.
[10] C.K.M. Lee, Kam K.H. Ng, Chun-Hsien Chen, H.C.W. Lau, S.Y. Chung, 12, p. 2183, 2021.
and T. Tsoi, ”American sign language recognition and training method [32] W. Abdul, M. Alsulaiman, S.U. Amin, M. Faisal, G. Muhammad, F.
with recurrent neural network,” Expert Systems with Applications, vol. R. Albogamy, M. A. Bencherif and H. Ghaleb, ”Intelligent real-time
167, p. 114403, 2021. Arabic sign language classification using attention-based inception and
[11] S. Sharma, and K. Kumar, ”ASL-3DCNN: American sign language BiLSTM,” Computers and Electrical Engineering, vol. 95, p. 107395,
recognition technique using 3-D convolutional neural networks,” Mul- 2021.
timed Tools Appl, vol. 80, p. 26319–26331, 2021. [33] P. Molchanov, S. Gupta, K. Kim and J. Kautz, ”Hand gesture recogni-
[12] K.M. Peréz, D.M.C. Esparza, J. Terven, A.M.H. Navarro, T.G. Ramı́rez, tion with 3D convolutional neural networks,” in IEEE Conference on
and A.R. Pedraza , ”Automatic Recognition of Mexican Sign Language Computer Vision and Pattern Recognition Workshops (CVPRW), 2015.
Using,” Appl. Sci., vol. 12, p. 5523, 2022. [34] M. M. Islam, M.R. Uddin, M. N. AKhtar and K.M. R. Alam, ”Recogniz-
[13] Q. Xiao, M. Qin and Y. Yin, ”Skeleton-based Chinese sign language ing multiclass Static Sign Language words for deaf and dumb people
recognition and generation for bidirectional communication between of Bangladesh based on transfer learning techniques,” Informatics in
deaf and hearing people,” Neural Networks, vol. 125, pp. 41-55, 2020. Medicine Unlocked, vol. 33, 2022.
[14] J. Zhang, W. Zhou, C. Xie, J. Pu and H. Li, ”Chinese sign language [35] K. Mistree, D. Thakor and B. Bhatt, ”Indian Alphabets and Digits Sign
recognition with adaptive HMM,” in IEEE International Conference on Recognition Using Pretrained Model,” Smart Innovation, Systems and
Multimedia and Expo (ICME), 2016. Technologies, vol. 283, 2022.
[15] S.G.M. Almeida, F.G. Guimarães, and J.A. Ramı́rez, ”Feature extraction [36] K. Wangchuk, P. Riyamongkol and R. Waranusast, ”Real-time Bhutanese
in Brazilian Sign Language Recognition based on phonological structure Sign Language digits recognition system using Convolutional Neural
and using RGB-D sensors,” Expert Systems with Applications, vol. 41, Network,” ICT Express, vol. 7, no. 2, pp. 215-220, 2021.
no. 16, pp. 7259-7271, 2014. [37] K. Suri and R. Gupta, ”Continuous sign language recognition from wear-
[16] M. Elpeltagy,M. Abdelwahab,M. E. Hussein,A. Shoukry,A. Shoala, and able IMUs using deep capsule networks and game theory,” Computers
M. Galal, ”Multi-modality-based Arabic sign language recognition,” IET & Electrical Engineering, vol. 78, pp. 493-503, 2019.
Comput. Vis., vol. 12, pp. 1031-1039, 2018. [38] D. Singh, ”3D-CNN based Dynamic Gesture Recognition for Indian
[17] A. Tang, K. Lu, Y. Wang, J. Huang and H. Li, ”A Real-time Hand Sign Language Modeling,” Procedia Computer Science, vol. 189, pp.
Posture Recognition System Using Deep Neural,” ACM Trans. Intell. 76-83, 2021.
Syst. Techno, vol. 9, no. 4, p. 23, 2013. [39] S. Dhulipala, F. F. Adedoyin and A. Bruno , ”Sign and Human Action
[18] L. Pigou, S. Dieleman, P. J. Kindermans and B. Schrauwen, ”Sign Lan- Detection Using Deep Learning,” J. Imaging, vol. 8, no. 192, pp. 1-34,
guage Recognition Using Convolutional Neural Networks,” in Computer 2022.
Vision - ECCV 2014 Workshops, 2015. [40] R. Pati, V. Pati, A. Bahuguna and G. Datkhile, ”Indian Sign Language
[19] J. Duan, J. Wan, S. Zhou, X. Guo, and S. Z. Li, ”A Unified Framework Recognition using Convolutional Neural Network,” in ITM Web of
for Multi-Modal Isolated Gesture Recognition,” ACM Trans. Multimedia Conferences (ICACC), 2021.
Comput. Commun. Appl, vol. 9, no. 4, p. 17, 2017. [41] U. Nandi, A. Ghorai, M. M. Singh, C. Changdar, S. Bhakta and R.
[20] Gupta M, Thakur N, Bansal D, Chaudhary G, Davaasambuu B, and Hua K. Pal, ”Indian sign language alphabet recognition system using CNN
Q, ”CNN-LSTM Hybrid Real-Time IoT-Based Cognitive Approaches with diffGrad optimizer and stochastic pooling,” Multimedia Tools and
for ISLR with WebRTC: Auditory Impaired Assistive Technology,” J Applications, 2022.
Healthc Eng, 2022.
[21] L. Kraljević, M. Russo, M. Pauković, and M. Šarić, ”A Dynamic Gesture
Recognition Interface for Smart Home Control based on Croatian Sign
Language,” Applied Sciences, vol. 10, no. 7, 2020.

SCEECS 2023
Authorized licensed use limited to: M S RAMAIAH INSTITUTE OF TECHNOLOGY. Downloaded on March 10,2025 at 08:23:25 UTC from IEEE Xplore. Restrictions apply.