A deep learning approach with weighted voting for Moroccan sign classification

doi:10.21203/rs.3.rs-3596062/v1

Download PDF

Research Article

A deep learning approach with weighted voting for Moroccan sign classification

https://doi.org/10.21203/rs.3.rs-3596062/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

In our Moroccan society, as in the rest of the world, we find a significant proportion of deaf-mutes who represent 5% of the world's population, the equivalent of 466 people suffering from hearing loss. These people use sign language as a means of transmitting their messages, emotions and expressions to other people, which implies little or no hearing.

To facilitate communication between deaf-mutes and normal people who do not know sign language, we have proposed in this article an approach that enables the textual transcription of sign language into natural language. Due to the development of Artificial Intelligence, we will propose an approach that is based on a combination of four deep learning architectures, with the weights of each architecture calculated according to their performance using genetic algorithms. It turns out that using the weighted voting method of deep learning or the so-called ensemble method gives a better performance to the results obtained using each deep learning architecture separately and compared to recent approaches in the literature, enabling us to predict signs and improve the accuracy rate to 99%.

Artificial Intelligence and Machine Learning

deaf-mute

sign language

natural language

Deep Learning

Ensemblist method

weighted voting

Artificial intelligence

When we cannot convey a message orally, we use gestures as a complementary means of communication. There are different gestural communications depending on the need for expressiveness, the simplest are command gestures that do not combine with each other or with another communication channel, mimic gestures are a little more complex and are complementary to speech. At the highest level of complexity are the sign language gestures that combine to form sentences.

In this article and as a special case, we will deal with Moroccan Sign Language (MSL), which is a comprehensive and complex language that uses signs made by hand movements combined with facial expressions and body postures. It is the language of choice for many people who are unable to speak and one of the many alternative communications used by people who are deaf or hard of hearing. Many researchers [1–16] have put considerable effort into developing a tool that can help these people communicate and exchange information with others. In order to facilitate the task of communication between these two categories of people, a tool for the transcription of sign language into natural language will be created, which follows a specific architecture, which is:

The pre-processing step “Resize the images in the training database”
Extract features (Color, Shape, Texture)
Create the models.
Weighted voting (application of the assembly method)
Analysis of the performance of the Classifiers.
The classification and prediction stage.

The following will detail each phase of processing of the sign recognition problem:

The preprocessing stage: is a process applied to images with insufficient quality or containing artifacts that can negatively influence the next stage.
Feature extraction: is a crucial step to obtain a good classification rate. Among the features, we extract the color, shape and texture of the sign.
Feature Engineering: which contains feature normalization, dimensionality reduction and feature selection; this process is used to first normalize all features and then select them to reduce dimensionality and remove redundant features.
Model creation.
Voting system: model that contains four Classifiers and selects the best one among them
Classifier performance analysis: From this step, we calculate the accuracy of each Classifier.
Classification & prediction stage: this is the last stage of such a system, which classifies the signs and their appropriate alphabet.

The next section presents generalities on Sign language and as a particular case the Moroccan sign language, and then the third section represents the state of the art of approaches dealing with the same problem. The fourth section presents in detail the proposed approach. In the fifth section, we will present the results of the application of our approach as well as the dataset used in our project. A synthesis of our application will be presented in the last section.

The problem of sign recognition is posed everywhere because one should not neglect this important number of these deaf-mute people in the whole world because we have more than 450M of deaf-mute people in the world according to statistics of the world health organization (2021), that's why a lot of researches have been done in this theme in order to find the solution for the communication between deaf-mute and normal people, also several hardware and software solutions have been proposed for the automatic sign recognition. A review and analysis of recent advances in this field can be found in (Machine learning methods for sign language recognition)

Most of the proposed solutions are based on connected gloves. For example, there are a multitude of smart gloves using sensors to recognize hand and finger gestures and movements. These embedded solutions do not facilitate the exchange between deaf mutes and normal people and do not recognize the signs made by full body gestures. Other solutions have been proposed that are based on image processing, sensors and artificial intelligence. Thus, image processing is used to capture and identify gestures while artificial intelligence is used for sign recognition and transcription into written language.

American researchers have developed connected gloves that can recognize signs and make automatic transcriptions (“The Language of Glove: Wireless gesture decoder with low-power and stretchable hybrid electronics”.). The solution proposed in the form of a leather sports glove on which nine movement sensors have been placed at the level of the joints, its various electrodes allow to translate the gestures made by the fingers and the wrist to designate the 26 letters of the American alphabet. To date, this innovation can only spell words. Unfortunately and due to lack of funding, this solution remains expensive and unreliable. Boon Giin Lee et al. in ("Smart Wearable Hand Device for Sign Language Interpretation System With Sensors Fusion,", Feb.1, 2018) proposed a wearable smart device for sign language recognition. This proposed system uses 10 sensors: five flexible sensors, two pressure sensors, and a three-axis inertial motion sensor. The system can recognize sign language with an accuracy rate of 65.7%. In ("A new data glove approach for Malaysian sign language detection", Decembre 2015), Shukor et al. proposed a glove for Malay sign language recognition. The proposed system is based on 10 click sensors, an accelerometer, a microcontroller, and a Bluetooth module. The proposed system has not been tested extensively to approve its effectiveness but over four uses, they were able to achieve a sign recognition accuracy equal to 89%.

Solutions based on gloves and sensors are very limited in their use and do not manage all the gestures that deaf and dumb people can make. On the other hand, the use of image processing and artificial intelligence remains an effective solution for the recognition of signs and gestures of deaf and dumb people. In sign language recognition approaches that rely on image processing and artificial intelligence, image processing is used to capture and identify gestures while artificial intelligence is used for sign recognition and transcription into written language. Hand detection is the first step in such a system. In addition, it is essential to detect it correctly in order to be able to effectively recognize the sign made. This step is considered a challenge. In (Hand landmarks detection and localization in color images., 2016), the authors proposed an approach for the detection and localization of hand landmarks using color images. The proposed approach is based on the use of a mask for skin identification and the use of a distance to detect hand landmarks. Ravikiran et al. in (Finger detection for sign language recognition, March 2009) proposed an approach for finger detection. Fingers are detected based on boundary tracing for tip detection. For sign recognition, we use machine-learning methods for their classification. Thus, the objective is to design a classification model using supervised classification algorithms able to recognize the sign of a gesture image. In this step, a database of labeled images is used to train a classification model.

In paper (CNN based feature extraction and classification for sign language., 2021), the authors used two deep learning architectures, AlexNet and VGG16 for feature extraction. Classification is performed using the SVM algorithm. The authors obtained a very good classification rate, but using only the training data. Deriche et al. in ( "An Intelligent Arabic Sign Language Recognition System Using a Pair of LMCs With GMM Based Classification,", Sept.15, 2019) proposed to use a classification based on a combination of a Bayesian approach and a mixture of Gaussian models (GMM) with the use of linear discriminant analysis (LDA) for dimensionality reduction. The simulation results showed that the proposed approach performs moderately well with a classification rate equal to 92%.

There are many works in the literature for American Sign Language, Indian, and Indonesian...

However, for Moroccan Sign Language (MSL), there are not yet enough research works except for those shown below which are based on an approach to classify Moroccan signs using different machine learning techniques by a group of researchers(Moroccan sign language recognition based on machine learning, 2022) who tried to create a new approach to have the alphabet suitable for the captured signs following a set of procedures shown in the figure below:

Moreover, they found that the system with a fully recurrent architecture performed best with 98.33% accuracy for static sign recognition.

The gesture recognition techniques proposed in the literature are not always adequate for RLS. As a result, the field of RLS has developed its own literature over the years. (RLSAr) systems, in particular, have not received much attention only in recent years. In (Automatic recognition of Arabic sign Language finger spelling ), the authors used colored gloves to collect alphabet data (RLSAr) from multiple users, where an adaptive neuro-fuzzy inference system was the recognition approach.

Below is the proposed architecture of the RLSAr system

Fasihuddin et al. (Smart Tutoring System for Arabic Sign Language Using Leap Motion Controler, 2018) proposed an interactive system that allows able-bodied learners who wish to learn Arabic sign language to practice (RLSAr) at different levels and to self-assess. This system uses the K-nearest neighbor algorithm for sign classification.

In (Automatic translation of Arabic text-to-Arabic sign language, 2018), the authors proposed an automatic visual system that translates isolated Arabic word signs into text. This system has four main steps: segmentation and tracking of the hand by a skin detector, feature extraction and finally classification are done by Euclidean distance since the data is not large enough.

Luqman et al. (Automatic translation of Arabic text-to-Arabic sign language, 2018) have gone further by proposing an automatic system that simultaneously performs morphological, syntactic and semantic analysis on an Arabic sentence to translate it into a sentence with the grammar and structure of (LSAr).

Maraqa et al. (Recognition of Arabic Sign Language (ArSL) using recurrent neural network, 2008) Compared two neural network-based systems: feedforward (Svozil, 1997) and recurrent (Recurrent neural network based language model, 2010)and found that the system with a fully recurrent architecture provided the best performance with 95% accuracy for static sign recognition.

In this next section we will talk about sign language and especially Moroccan sign language

Before starting the section about the search, methodology we will give a general idea about Sign language can be defined as a gestural language produced by the movements of the hands, face and body as a whole. This language is the language used by deaf-mute people to communicate with others.

In the following we will work specifically on the Moroccan sign language

3.1 Moroccan Sign Language (MSL): the language of deaf-mutes in Morocco

Deaf and dumb people in Morocco use Moroccan sign language to communicate which has developed through local signs, which is similar to about 50% of the French language signs; this similarity is due to the closeness of the two countries.

MSL consists of 28 signs for the Arabic alphabet, plus 10 signs for numbers from 0 to 10.

Figure 1 shows the Moroccan Sign Language (MSL)

One forgets about the great diversity of Moroccan signs in the different regions of Morocco because there are small differences between the signs and the way of expression.

3.2 Methods Used

3.3.1 Ensemblist method

In this proposed approach, we propose a new and powerful method for Moroccan sign classification based on ensemble learning, which is a machine-learning paradigm where multiple learners are trained to solve the same problem. It has been shown that the combination of several classifiers can be more effective than a single one. From a technical point of view, ensemble learning is mainly implemented in two steps: The training of weak classifiers and the selective combination of member classifiers into a stronger classifier.

According to this source (A deep learning-based multi-model ensemble method for cancer prediction. Computer methods and programs in biomedicine, 2018), it is feasible to combine the performance of a set of classification architectures to obtain a single robust classification model; each architecture has its strengths and weaknesses. Typically, the members of an ensemble are constructed in two ways. One is to apply a single learning algorithm and the other is to use different learning algorithms on a data set (Applying Ant Colony Optimization to configuring stacking ensembles for data mining.).

Then, the basic classifiers are combined to form a decision classifier. In general, to obtain a good ensemble, the base learners should be as accurate as possible and as diverse as possible. How to choose an accurate and diverse set of base learners is therefore a concern for many researchers. In recent years, more and more researchers have become interested in ensemble learning (Ensemble learning, Encyclopedia of Biometrics., 2009). The major principles of ensemble learning for analysis can be classified into the following methods: bagging, boosting, stacking and voting detailed below

Bagging: is the most basic and simple method; it has two steps: bootstrapping and aggregation, in which a single learning algorithm is applied to different subsets of training data, with sampling of the subsets being done with replacement (bootstrapping). Once the algorithm has been trained on all subsets, the bagging predicts by aggregating all predictions made by the algorithm on all subsets. Since the sampling is done by bootstrapping, the bagging method works very well due to the diversity of the training data. It reduces variance while not increasing bias (Bagging predictors. Machine learning., 1996 ); (Exploring the sources of uncertainty: Why does bagging for time series forecasting work?, 2018).
Boosting: This is a machine learning algorithm in which the weights of the classifiers are iteratively adjusted. The principle behind the boosting approach is based on the importance of more complex models. Classification has attracted interest in the amount of weights set in each model in the training set (Study of corporate credit risk prediction based on integrating boosting and random subspace., 2011). By decreasing the distribution and combination of weak classifiers, the AdaBoost algorithm expanded the weight of each case by forming a set of identical and different weak classifiers. Finally, the AdaBoost algorithm was used to construct the final decision using weighted majority voting.
Stacking: is another well known ensemble learning technique to obtain the best possible result. It often takes into account heterogeneous weak learners, learns them in parallel and combines them by forming a meta-model to produce a prediction based on the predictions of the different weak models. The stacking architecture set (Stacked generalization, Neural Networks, 2011) was a two-tiered structure consisting of level zero (base level) classifiers and level one (meta-level) classifiers. The training dataset was used to train the base-level classifiers, which were then used to build the prediction. The metadata was then used to train a meta-classifier to recognize the output of the base classifier as the class label. Stacking was generally used to combine the developed model with the different classifiers. These classifiers were combined in various predictions to form the final resolution.

In addition, in order to have a good performance of the classifiers we will use Bagging which consists in sub-sampling the data, producing a data set for each model and in order to identify the final result, we proceed by a vote of the results of each model for a classification, or by an average for a regression

3.3.2 Voting method

The voting method that is used to make decisions by selecting only one alternative among several. Voting depends on the class predicted with majority voting. In addition, voting is the most commonly used method in ensemble methods. In general, the voting method consists of an unweighted vote and a weighted vote.

Unweighted voting includes simple and majority voting, while weighted voting includes simple weighted majority voting, while weighted voting includes simple weighted voting (A weighted voting framework for classifiers ensembles, 2014). It supports these types of voting.
The hard voting or majority-based voting mechanism includes the sum of the predictions for each class label and prediction of the class label with the most votes, which means that the final class prediction is made by the majority vote that the estimator the estimator selects the class that occurs most frequently among the baseline models (Memory-based weighted majority prediction, (1999, August)) ; (Robust and secure zero watermarking algorithm for color images based on majority voting pattern and hyper-chaotic encryption, 2020).

Equation 1: the hard voting

and to improve our results we will use the majority weighted voting which was a simple voting considered as the appropriate method for all classifiers having equivalent performances. However, the basic classifier was practically realized with different weights to define the weight. Thus, weighted voting was designed to define the weight to build a strong classifier (E-Rickshaws Management for Small Scale Farmers using Big Data-Apache Spark., 2021). We can compute a weighted majority vote by associating a weight Wj with the classifier Cj:

Equation 2: a weighted majority vote

As well as Weighted voting has the same principle as above but is expressed in terms of probabilities. This means that the final output is predicted based on the predicted probabilities p for the classifiers (Du and Swamy, 2019). For each class label and predict the class label with the highest probability. In each case below, the probability of class labels assigned by classifier C to input x is defined as (Class-specific soft voting based multiple extreme learning machines ensemble, 2015):

Average of Probabilities Voting

Equation 3: Average of Probabilities Voting

Product of probabilities voting:

Equation 4: Product of probabilities voting

Minimum of probabilities voting:

Equation 5: Minimum of probabilities voting

Maximum of probabilities voting:

Equation6: Maximum of probabilities voting

3.3.3 The genetic algorithm

After the phase of application of the voting method we will apply the genetic algorithm for the calculation of the weight we will select our weighting parameters. Here, we will use the genetic algorithm approach to find the optimal weights for each classifier to increase the prediction accuracy. The main steps of the genetic algorithm used are:

Initialization by producing a collection of many individuals and each individual denotes the weight of each classifier.
Fitness: The GA calculates the fitness of individuals based on some objective evaluation function, learning the survival probability of individuals in the next evolution.
§Selection: Using random or specific population rules to participate in crossover and mutation. A number of excellent individuals with "more fitness" are selected.
Frequently, the great individuals are generally of high fitness, that is, better And the last phase is the prediction of the signs to obtain the suitable alphabet for each class.

4.1 Dataset

The dataset is a collection of Moroccan sign language alphabet images, separated between training and test images. The training dataset contains 5839 images of 256 x 256 pixels. There are 28 classes, 28 of which are for ا-ي letters. The test dataset contains only 28 images, to encourage the use of real test images. Some of them are displayed in this Figure.

It encompasses the signs and their suitable alphabets. To make it appropriate for the initial values of the trained models. The size of the images was changedto 32x32 and the normalization was measured by dividing the pixel values by 255. To avoid the overfitting problem, we divide our dataset into 2 categories: 80% for training, 20% for testing

4.2 Measures de performance

In this study, we evaluate the performance of our proposed approach by computing the accuracy of each architecture this measure is the precision. For accuracy, it is the ratio of correctly predicted positive observations to the total number of predicted positive observations;

Equation 7:calcul of accuracy

Where: n = the number of classes, i the current class.

4.3 Simulations results

The overall workflow of the proposed method is shown in the figure below.

As we can clearly see, after applying the four architectures mentioned above, using the different voting mechanisms to improve the classification results. Each of the classification results of these architectures is calculated for each test instance and the final results are predicted based on the weighted majority results. The details of the voting mechanism used are given in the previous subsections. We summarize the main steps as the following descriptions:

Step 1: Before applying ensemble learning to the dataset, we perform a fine-tuning of each of the four deep learning algorithms to be used to define the hyper parameters to be used.
Step 2: Application of Global Average Pooling (GAP).

The GAP layer reduces each feature map to a single number by simply taking the average of all values and minimizes overfitting by reducing the total number of parameters in all models.

Step 3: We combine the 4 DL algorithms. The weights of each of the 4 algorithms are optimized using the genetic algorithms.
Step 4: The final prediction is obtained by a weighted vote using the optimal weights obtained in step three

Now we select our weighting parameter. Here, we proposed to use the voting system approach to increase the prediction accuracy.

We used the python language to program ensemble learning architectures based on the various voting mechanisms, due to the simplicity and availability of very useful DL frameworks and libraries.

Our proposed approach has been implemented on a Windows 10 computer with an Intel Core i7 with 3.6 GHz processor and 8GB of RAM.

In our implementation, each of the four classification algorithms is tested using the scaled extracted feature matrix. To compare the performance of each of the classification algorithms, Table 1 shows the accuracy values:

Table 1

Performance measures for every model
Model	Accuracy
CNN	76%
AlexNet	86%
VCG16	95%
Inception ResNetV2	96%
méthode de vote	99%

Here is the confusion matrix of our model

After examining these results, we notice that all architectures showed significant performance, starting with accuracy; CNN is obtained one of the lowest values with 76% followed by the result obtained by AlexNet with 86%, after the common result obtained by VCG16 and InceptionRecNetV2 with 95% and 96% respectively. The best recall represented by InceptionRecNetV2 provides almost excellent classification results which reached 99%.

Attached is the graph comparing the accuracy of each model and the weighted vote:

To verify the performance of the proposed weighted voting technique we will compare this approach with what has already been done by previous researchers: In (Automatic recognition of Arabic sign Language finger spelling ) have used deep learning to create a system named RLSar which uses gloves to collect signs from users and have the suitable alphabet .In addition, the authors (Smart Tutoring System for Arabic Sign Language Using Leap Motion Controler, 2018) continue the work on RLSar system to develop it but this time with the use of the nearest neighbor algorithm (KNN), as a result (Automatic translation of Arabic text-to-Arabic sign language, 2018) propose a virtual system for translation of Arabic sign words using the following scheme: Segmentation -> hand tracking->feature extraction-> Classification, another research done on this topic by (Moroccan sign language recognition based on machine learning, 2022)with the use of machine Learning in order to create a tool capable of doing the transcription of signs into natural language and they had an accuracy of 98.33% using this approach. In comparison with these existing works and as you can see, the new method we propose, which consists on the use of weighted voting architectures, has given us an accuracy of 99% and approved its feasibility. As far as we can tell, the method we propose is more powerful and efficient than any other method in the literature to date. Here is a graph that presents the different researches with their accuracy and its comparison with the new proposed approach:

The problem of information exchange between deaf mutes and normal people leads us to find a solution for the transcription of these signs into natural language.

From this study, we tried to propose a weighted voting technique for the classification and prediction of Moroccan signs based on an ensemble learning approach. The main task is to improve the accuracy of classification using the weighted voting mechanism; the results obtained prove the effectiveness of our system by giving an average accuracy of 99% on the signs dataset. In future work, we intend to deploy this technique on other signs (ASL, LSF ...)

In order to automatically detect, classify and predict signs and then transcribe these signs into natural language as well as build our own dataset using pieces of video and transcribe words and sentences into natural language

Acknowledgment

We would like to express our gratitude and thank you to the LISAC laboratory, the Faculty of Sciences, the University Sidi Mohamed Ben Abdellah, Fez, Morocco for the funding and those who are involved directly or indirectly in this research.

Author’s Contributions

Meryem cherrate: Developpement, experimentations testing, validation and writing the manuscript.

My Abdelouahed Sabri: Concept developpement, experimentations testing, validation and proof reading.

Ali yahyaouy: Concept developpement, experimentations validation and proof reading.

Abdellah Aarab: Concept developpement, experimentations validation and proof reading.

Ethics

All information provided in this study is confidential and unique. This paper has neither been published nor is under review elsewhere. There are no ethical issues associated with this research.

Conflict of interests

The authors have no conflicts of interest to declare.

We certify that the submission is original work and is not under review at any other publication.

"World Health Organization (WHO)," 1 April 2021. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/deafness-and-hearing-loss.
O. B. M. A. I.A. Adeyanju, "Machine learning methods for sign language recognition," vol. 12, no. https://doi.org/10.1016/j.iswa.2021.200056, p. 2021.
T. &. F. M. &. M. R. &. R. &. M. P. &. L. D. O’Connor, "“The Language of Glove: Wireless gesture decoder with low-power and stretchable hybrid electronics”.," PLoS ONE. 12. 10.1371/journal.pone.0179766.
B. G. L. a. S. M. Lee, ""Smart Wearable Hand Device for Sign Language Interpretation System With Sensors Fusion,"," IEEE Sensors Journal, vol. 18, pp. 1224-1232, Feb.1, 2018.
A. Z. Shukor, M. F. (Decembre 2015). "A new data glove approach for Malaysian sign language detection". Procedia Comput. Sci., 76, 60-67.
T. K. M. &. G. A. Grzejszczak, "Hand landmarks detection and localization in color images.," Multimed Tools, pp. 75-16363–16387 , 2016.
J. M. K. M. S. D. R. S. S. &. P. N. V. Ravikiran, "Finger detection for sign language recognition," in In Proceedings of the international MultiConference of Engineers and Computer Scientists, March 2009.
Barbhuiya, A. K. (2021). CNN based feature extraction and classification for sign language. Multimed Tools.
S. O. A. a. M. M. M. Deriche, " "An Intelligent Arabic Sign Language Recognition System Using a Pair of LMCs With GMM Based Classification,"," in IEEE Sensors Journal, Sept.15, 2019.
C. M. Y. A. A. A. Sabri abdelouahed, "Moroccan sign language recognition based on machine learning," in ICSV22, 2022.
Al-Rousan, M. &. (n.d.). Automatic recognition of Arabic sign Language finger spelling . International Journal of Computers and Their Applications, 80-88.
H. A. S. A. S. e. a. FASIHUDDIN, "Smart Tutoring System for Arabic Sign Language Using Leap Motion Controler," in International Conference on Smart Computing and Electronic Enterprise (ICSCEE), 2018.
H. &. M. S. A. Luqman, "Automatic translation of Arabic text-to-Arabic sign language," Universal Access in the Information Society, pp. 1-13, 2018.
M. Maraqa and R. Abu-Zaiter, "Recognition of Arabic Sign Language (ArSL) using recurrent neural network," in First International Conference on the Applications of Digital Information and Web Technologies (ICADIWT), Ostrava, 2008.
D. V. K. a. J. P. Svozil, Chemometrics and intelligent laboratorysystems, pp. 43-62, 1997.
T. e. a. Mikolov, "Recurrent neural network based language model," in Eleventh annual conference of the international speech communication association., 2010.
Y. W. J. L. Z. &. Z. X. Xiao, "A deep learning-based multi-model ensemble method for cancer prediction. Computer methods and programs in biomedicine," no. 153, pp. 1-9, 2018.
Y. W. M. L. &. L. H. Chen, "Applying Ant Colony Optimization to configuring stacking ensembles for data mining.," Expert systems with applications, pp. 2688-2702.
Z. H. Zhou, "Ensemble learning, Encyclopedia of Biometrics.," 2009.
L. Breiman, "Bagging predictors. Machine learning.," 1996 .
F. H. R. J. &. B. C. Petropoulos, "Exploring the sources of uncertainty: Why does bagging for time series forecasting work?," European Journal of Operational Research,, no. 268(2), pp. 545-554, 2018.
G. &. M. J. Wang, "Study of corporate credit risk prediction based on integrating boosting and random subspace.," Expert Systems with Applications, pp. 13871-13878, 2011.
D. H. Wolpert, "Stacked generalization, Neural Networks," vol. 5, no. 2, pp. 241-259, 2011.
L. I. &. R. J. J. Kuncheva, "A weighted voting framework for classifiers ensembles," Knowledge and Information Systems , no. 38(2), pp. 259-275, 2014.
J. &. I. N. Delgado, "Memory-based weighted majority prediction," SIGIR Workshop Recomm. Syst. Citeseer, p. 85, (1999, August).
X. B. L. G. F. C. Y. J. Z. F. Z. E. H. &. J. C. N. Kang, "Robust and secure zero watermarking algorithm for color images based on majority voting pattern and hyper-chaotic encryption," Multimedia Tools and Applications, pp. 1169-1202, 2020.
N. K. R. C. N. S. A. &. O. J. Tyagi, "E-Rickshaws Management for Small Scale Farmers using Big Data-Apache Spark.," in IOP Conference Series: Materials Science and Engineering, 2021.
J. K. S. W. R. L. X. L. K. &. K. X. Cao, "Class-specific soft voting based multiple extreme learning machines ensemble," Neurocomputing, pp. 275-284, 2015.

Download PDF

Version 1

posted

You are reading this latest preprint version

A deep learning approach with weighted voting for Moroccan sign classification

Status:

Version 1

Abstract

Figures

1. Introduction

2. State of the art

3. Sign language & methodology of research

3.1 Moroccan Sign Language (MSL): the language of deaf-mutes in Morocco

3.2 Methods Used

3.3.1 Ensemblist method

3.3.2 Voting method

3.3.3 The genetic algorithm

4. Experimental Results

4.1 Dataset

4.2 Measures de performance

4.3 Simulations results

5. Comparative study

Conclusion

Declarations

Acknowledgment

Author’s Contributions

Ethics

Conflict of interests

References

Status:

Version 1