Journal of Computing and Communication Vol.2 , No.1 , PP.
9-19 , 2023
MLHandwrittenRecognition:Handwritten Digit Recognition
using Machine Learning Algorithms
Farah Essam a, Hanien Samy a, Judy Wagdy a, Steven Albert a , Diaa Salama AbeElminaamb a,b,*
a Department of Computer Science, Misr International University, Cairo, Egypt
binformation systems department; Faculty of Computers and artificial intelligence; Benha university, Egypt
*
Corresponding Author: Diaa Salama [[Link]@[Link]]
ARTICLE DATA ABSTRACT
Handwritten digit recognition has remained a topic of interest to computer vision scientists. Its
Article history: origination precedes the emergence of the machine as it is a crucial component of the digital
Received 16 Oct 2022
transformation of the majority of institutions in numerous fields. With the uprising of machine
Revised 12 Jan 2023
Accepted 12 Jan 2023 models, choosing a satisfactory and fit algorithm for this multi-class (0-9) classification problem
Available online became challenging. This paper aims to compare seven machine learning algorithms in terms of
their performance metrics in recognizing handwritten digits employing two datasets. The -
Keywords: Nearest Neighbors (kNN), Support Vector Machine (SVM), Logistic Regression, Neural
OCR Network, Random Forest (RF), Naive Bayes, and Decision Tree models are accordingly
Handwritten digit evaluated concerning the Area Under the Curve (AUC), accuracy (ACC), F1-score (F1),
Machine Learning
precision (PREC), and recall (REC). The widely used Modified National Institute of Standards
Computer Vision
and Technology database (MNIST) dataset and the Handwritten Digit Classification dataset
(HDC) have been the providers of the images on which this research is conducted. The results
confirm that the Neural Networks model is a great classifier for this problem; however, it presents
similar results to other machine learning classifiers in several cases. Therefore, this paper does
not provide an absolute choice of a classifier for the handwritten digit recognition problem but
rather explains the reason behind the performance of each model.
1. Introduction
Handwriting recognition, often known as handwriting Optical Character Recognition (OCR) or cursive
OCR, is an OCR sub-field that converts handwritten characters into digital text or commands. Moreover,
handwritten digit recognition is the computer's ability to recognize human handwritten digits from a variety
of sources, such as images, papers, and touch screens, to classify them into ten numerical categories (0-9).
The concept of handwritten digit recognition began with the use of pattern matching. Shelia Guberman,
who lived in Moscow then, built the first applied pattern recognition program in 1962.
OCR is a worldwide used technology that allows people to view, find, and distinguish text in images
and labels in a variety of ways. It is one of the early computer vision challenges, as it does not necessarily
require deep learning in some aspects [1]. As a result, various OCR implementations existed even before
the deep learning hype in 2012. Modern OCR software uses Artificial intelligence (AI) and machine
learning to reach even higher degrees of precision, such as recognizing different languages, reading
handwritten text and writing styles, dealing with typical data restrictions, and much more. It is fair to say
that deep learning has radically changed the field of OCR. In the last few years, a lot of development has
been made in this field; for instance, Convolutional Neural Network (CNN) is currently a powerful tool for
solving computer vision problems in many fields [2].
Handwritten digit recognition is subject to several challenges, including:
• A user's handwriting style varies and is inconsistent from time to time.
• Cursive handwriting makes character separation and recognition complex.
9
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
• As opposed to printed text which is rotation-invariant, handwritten text may be rotated in various
ways.
• Insufficient image quality for the source may result in poor recognition results.
This makes it difficult for various machine-learning techniques to detect the digits.
This paper examines and compares machine learning classifiers in recognizing handwritten digits.
It also addresses the methodology employed to detect digits using the MINST [3] and Handwritten Digit
Classification datasets. To evaluate them, performance metrics were utilized.
In section 2, similar studies are discussed, and relevant results are presented. In Section 3, the
methodology followed in this study is explained. Section 4 discusses the techniques used and the output
of each one after testing on all data sets. Section 5 evaluates the results presented in Section 4 [4]. Lastly,
the conclusion is conducted in Section 6.
2. Related Work
In [5], the authors compare the performance (accuracy, complexity, execution time, and the number of
epochs) of the kKNN, RF, SVM, and various forms of neural network classifiers in recognizing handwritten
digits. They have based their work on the MNIST dataset while applying simple image preprocessing
(scaling and filtering) techniques. However, throughout their research, they have added extra preprocessing
stages for more satisfactory results. The results show an overall accuracy of 98-100% for the test program
and 96-98% for the industrial images. The proposed algorithms achieved nearly identical accuracy (with a
one percent difference). Despite that, the Convolutional Neural Network (CNN) stands out with the highest
accuracy but with a significant computing time.
The approach proposed in [6] is based on the comparison and evaluation of various machine learning
algorithms, such as Bayesian networks, Multi-layer perceptrons (MLP), Random forest, SVM, and other
Convolutional Neural Network (CNN) models, to train and test a handwritten numbers (0-9) recognition
model. The MLP classifiers either stopped at local minimums or overfit. They have experimentally proven
that the more data was used in training, the longer it took, but the more accurate the model became. They
have also concluded that the SVM presented the best separation margins between a pair of classes. The final
results show that the LeNet 4, a CNN method, achieves the highest accuracy 99.3%, with the drawback of
a long training time (five weeks) but with a recognition time of 0.05 ms.
The authors of [7] experiment with the CNN method using the Rectified Linear Units (ReLU) activation
function on the MNIST dataset of handwritten digits. They also use the Deeplearning4j (DL4J) framework.
They have achieved the highest results with two convolutional layers, where the first consists of 32 filters
and a window size equal to 5×5, and the second layer is composed of 7x7 window 64 filters. Consequently,
they offer a model with an accuracy of 99.21%. In this paper, it is claimed that the high performance of
their model concerns both criteria: accuracy and time.
The strategy pursued by the authors in [8] is to compare several classical machine learning algorithms
in building handwritten digit classification models. The following algorithms were chosen: Logistic
Regression, SVM, Decision Tree, RF, and kNN. The MNIST dataset is employed in the training and testing
process of the above algorithms. The algorithms are compared based on factors such as the learning and
prediction construction speed and the recognition accuracy. Each algorithm underwent 100 iterations of
training and testing phases. According to the performance analysis, the kNN, Decision Tree, Logistic
Regression, and Random Forest were the fastest algorithms during the prediction process for the test dataset,
with respectively a recognition time of 0.01, 0.03, and 0.36 seconds, respectively. The kNN algorithm was
10
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
the least computational power intensive during training, as highlighted with a score of (17.43%) of CPU
load and memory use of 209.8 MiB. The most accurate algorithms in this study are the SVM, kNN, and
RF, with accuracy scores equal to, respectively, 97.93%, 97.21%, and 96.95%.
In this article [9], even though the artificial neural network (ANN) with feature extractors, such as CNN,
is a common choice for image classification, the authors claim that the three boosting algorithms discussed
performed well in correctly predicting classes, with all three algorithms achieving the accuracy of more
than 93%. Algorithms such as Extreme Gradient Boost (XGBoost), AdaBoost, and Gradient Boosting were
used on the MNIST dataset to recognize handwritten digits. Confusion matrix, recall, F1 score, and
precision were used to compare the performance of the algorithms. It can be concluded from the confusion
matrix that the AdaBoost algorithm predicted the classes correctly with an accuracy of 96.86%. With an
accuracy of 94.59%, the gradient boosting algorithm came out on top, followed by XGBoost at 93.6%. It
was proven that the three algorithms could predict the '1' digit most effectively. Throughout the three
algorithms, AdaBoost performed the best in terms of precision at 96.84%, F1 score, and recall at 96.85%.
In this paper [10], a framework for handwritten digit recognition is developed. It is based on feature
extraction and algebraic fusion of various classifiers. The feature extraction is based on a CNN model using
the MNIST dataset. The results demonstrate that the model's fusion achieves at least 98% accuracy. The
authors indicate that the same classifier can perform differently depending on the test set. This remark is
due to the various ways of shaping a single digit based on the user's handwriting style. The ensemble
learning techniques construct the solution provided by the authors to improve and stabilize the performance
of the classifiers. For training base classifiers and primary ensembles, the kNN and RF have been used. The
findings show that the KNN method's instance-based learning resulted in over 95.8% accuracy on the two
feature sets. While the RF approach may create highly diverse decision tree classifiers using a range of
training samples and feature subsets, the accuracy on the two feature sets is above 95.7%. The results also
reveal that by utilizing the MNIST data set, their suggested ensemble technique may reach a classification
accuracy of more than 98%, indicating that employing ensemble learning to train various classifiers is
particularly advantageous to improving overall classification performance.
Several machine learning techniques, such as unsupervised learning algorithms, are investigated in
[11]. K-means clustering is used to minimize the size of the data. Neural networks, SVMs, and closest-
neighbor methods then follow a linear classifier. The author used the MNIST dataset to compare
performance using prediction accuracy. Handwritten digit recognition performance is determined by two
factors: (1) the method used to extract features and (2) the algorithm used for classification. Users should
choose configuration parameters to improve performance and some computational constraints associated
with implementing the algorithms. SVM is a binary classifier that determines the optimal separating
hyperplane between two classes and can be utilized to split the feature space into classification regions for
multi-class problems. Even with established algorithms, further study is needed to discover the best tuning
parameters. Neural networks can be investigated using a variety of hidden layers and nodes. The softmax
function is used in the neural network's output layer. Other output functions for those layers that could have
an important impact on the performance may be investigated.
In [12], the author uses Convolutional neural networks (CNN) as an example of image classification
and, more precisely, Keras sequential models as a classifier. The dataset used for training and testing is the
MNIST dataset. The most important phase was picture preprocessing, which was accomplished with
OpenCV and Scipy. This study's classifier is a sequential model with a four-layer CNN. Different CNN
layers, such as two, three, and four-layered CNN, were tested for accuracy on the test data. The four-layered
CNN exhibited the best accuracy of 99.25%, with the least loss in the case of 15 epochs among all the layers
evaluated. Comparative studies have been conducted over five epochs. 4-layer CNN had the highest
accuracy 98.89%, but the accuracy was lower than the one in the 15 epochs. This is because the algorithm
11
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
can train itself more efficiently when 15 epochs are employed, resulting in improved accuracy and smaller
losses. Also, API can show the outcome for every digit the user enters. This allows the online mechanism
to recognize numbers with high accuracy accurately. Users can insert a number into the terminal window
by hand, and the recognized digit will be displayed in the next window. The entire process may be enhanced
by using other methods, such as a multi-layer perceptron (MLP).
In [13], the paper compares three classification algorithms using the National Institute of Standards &
Technology (NIST) handwritten dataset: K Star, MLP, and Naive Bayes (NB) algorithms based on
correlation features selection (CFS). This comparison aims to determine which of the three classifiers can
provide a suitable accuracy rate while using the fewest number of features possible. Precision, recall, and
F-measure is the accuracy measurement measures used to evaluate each classifier's performance
independently. The results reveal that the K star algorithm performs better than NB and MLP algorithms,
with an accuracy of (82.36%). A comparison of three classification techniques was performed to recognize
NIST handwritten digits. To get a suitable recognition rate, only 37 features out of 256 were picked using
(CFS). After evaluating each classifier on 46080 samples using 10-fold cross-validation, K Star performed
better than the other classifiers with an accuracy of 82.36%, compared to NB 67.04% and MLP 78.35%. In
this comparative approach, the K Star algorithm outperforms the two other classifiers (NB and MLP) in
handwritten digit recognition.
In [14], the authors tested five classical machine learning classifiers and compared them using
performance metrics such as recall and F1-Score. They used only 1 dataset: the MNIST dataset. The paper
discusses the preprocessing steps the data went through before training, including line localization and
thresholding techniques. Feature extraction then took place, and the data was ready for training. The five
classifiers were SVM, Decision Tree, RF, ANN, kNN, and K-Means Algorithm. The accuracies achieved
were 90%, 87%, 97%, 98%, and 98%. The paper tested four different kernels for the SVM, and the best
accuracy was achieved with a polynomial kernel of 90%. However, ANN achieved the highest precision,
0.97. The SVM took the longest time in training, while KNN took the longest time in testing.
The authors in [15] propose different feature extraction methods to be applied to the MNIST dataset
before considering which classifiers to use. They tried four feature extraction methods: Cavity, Hu
moments, Zernike Moments, and Hog transformed and combined them with different classifiers. The
machine learning techniques were KNN, SVM one versus one, SVM one versus all, Decision Tree, and
MLP. The authors tested 17 combinations, and the best results were by training a KNN model on HOG
Features, resulting in a recognition rate of 96.57%. However, they faced a limitation due to the size of the
feature vector (1296) generated by the HOG extraction method. The paper couldn't showcase the
performance of MLP, SVOMO, and SMOVA models with HOG Features. Also, other related work in [16-
25] has been proposed in recent years to address machine learning and its application in different fields.
12
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
3. Proposed Approach
FIGURE 1. Methodology Diagram
3.1. Datasets
1) MNIST (Modified National Institute of Standards and Technology database)[3] is a common dataset
of handwritten digits that carries 60,000 handwritten digits for training and 10,000 handwritten
digits for a machine learning model to be tested and verified that they work. A sample of this dataset
is shown in figure 2. This is a subset of the larger set available from NIST. NIST's black-and-white
images were also normalized in size, centered and smoothed to fit within a 28 x 28-pixel border
box, and introduced with gray shading. The class distribution is illustrated in figure 3.
FIGURE 2. MNIST Dataset FIGURE 3. MNIST Class Distribution
2) MHDC (Handwritten Digit Classification dataset) [26] is a dataset of 30,010 binary images of
handwritten digits, as shown in the representative piece below (figure 4). Each digit (0-9) is assigned
a folder that holds 3001 enhanced and binarized images of that digit. All of the images are 128 x
128 pixels.
13
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
FIGURE 4. HDC Dataset
3.2. Training
The datasets described above are used to train the chosen classifiers. The chosen classifiers are the following:
1) KNN is a simple and straightforward algorithm that maintains all existing cases and classifies new
ones based on a majority vote of its k neighbors [27]. KNN (k- Nearest Neighbors) applies to both
classification and regression problems. However, in practice, it is more commonly used in
classification problems.
2) SVM (Support Vector Machine) is a supervised machine learning model. Each data point is plotted
in n-dimensional space (n being the number of features), with every feature being the value of a
specific coordinate. The kernel trick allows the performance of nonlinear classification [28].
3) Logistic Regression is a classification algorithm. It is generally used to calculate discrete values
depending on variables [29]. In other words, it fits data to a logistic function to forecast the likelihood
of an event occurring. Hence, it's also called logistic Regression as it predicts probabilities. Its output
values range from 0 to 1.
4) Neural Network: In deep learning techniques, artificial neural networks (ANNs) and simulated neural
networks (SNNs) are two forms of neural networks. They get their name and structure from the
human brain and work similarly to how neurons communicate. Neural networks can support
computers in making accurate decisions while requiring little human involvement [30]. This depends
on its ability to learn and model nonlinear and complex relationships between input and output data.
5) Random Forest: is a machine learning ensemble technique used in classification and Regression. Its
algorithm is responsible for determining the output based on the assumptions of the decision trees
[31]. It makes predictions by averaging the results of several trees. As the number of trees grows, so
does the precision of the outcome. Decision nodes, root nodes, and leaf nodes are the three parts of
a decision tree. A decision tree approach divides a dataset into branches, which are then divided
further into branches. This procedure is repeated until a leaf node is reached and cannot isolate the
leaf node anymore. To forecast the outcome, the qualities represented by the nodes in the decision
tree are utilized.
14
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
6) Naive Bayes is a probabilistic classifier that uses Bayes' theorem and assumes features independence.
According to a Naive Bayes classifier, the presence of one feature in a class is unaffected by the
presence of any other feature. It is simple to implement and performs well when a class's probability
depends on casual factors' probability [32].
7) Decision Tree: is a supervised learning algorithm commonly used for classification issues. It works
with both continuous and categorical dependent variables. The population is divided into two or more
homogeneous groups using this algorithm. To make as many unique groups as possible, this is done
using the most significant attributes and independent variables [33].
3.3. Test and Score
The proposed classifiers are compared against each other in each dataset using various performance metrics.
4. Experimental Results
In this section, the Orange data mining platform is used to evaluate the mentioned algorithms'
performance in recognizing handwritten digits with the MNIST and HDC datasets.
4.1. Performance Metrics
The proposed algorithms are evaluated in terms of Area Under the Curve (AUC), accuracy (ACC), F1-
score (F1), precision (PREC), and recall (REC). Four variables contribute to calculating most of these
metrics: true positive (TP), which represents the number of instances that have been correctly classified to
the selected class; false positive (FP), true negative (TN), and false negative (FN). The calculation formulas
are presented in table 1.
TABLE 1: Evaluation measures
Measures Formulas
Accuracy 𝑇𝑃 + 𝑇𝑁
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁
Precision 𝑇𝑃
𝑇𝑃 + 𝐹𝑃
Recall 𝑇𝑃
𝑇𝑃 + 𝐹𝑁
F-measure 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙
2×
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙
4.2. Algorithms parameters
The recognition of digits (0-9) is a multi-class classification problem; therefore, the algorithm's parameter tuning
was adapted accordingly. Orange data mining tool provides several customizations to each model before training and
testing it.
In the kNN model, it was set to 5 nearest neighbors with the Euclidean metric (distance parameter) and uniform
weights for all points in the same neighborhood.
The SVM's kernel function of choice was the Radial Basis Function (RBF). The SVM Linear is the traditional
SVM with a linear kernel function.
The Logistic Regression model used a Tikhonov regularization (L2) type with a cost strength (C) equal to 1.
15
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
The Neural Network model was accorded the Rectified Linear Unit (ReLu) activation function for the hidden
layer along with the Adam stochastic gradient-based optimizer and a maximal number of iterations equal to 100
iterations. The hidden layers were also adjusted to 100 neurons per hidden layer.
The Random Forest model was calibrated to include ten decision trees in the forest. The growth control attributes
were regulated so that the smallest subset that can be split has to have five instances.
No customization was needed for the Naive Bayes model.
Regarding the tree model, the limit of the maximal tree depth was set to 100 node levels, nodes with less than five
instances cannot be split, and splits would never produce branches with less than two training examples.
Most of Orange's models apply a few default preprocessing techniques, such as one-hot encoding for categorical
variables, removing empty columns, or substituting missing values with mean values. They also eliminate instances
with missing target values.
4.3. MNIST dataset results
As shown in table 2, the Neural Network model outperformed the other experimented models in all the
chosen performance metrics with an AUC of 1.0, ACC of 0.978, an F1-score of 0.978, and a recall of 0.978.
This experiment suggests that the Neural Network offers more satisfactory results than Trees, SVM, kNN,
and Logistic regression models in recognizing handwritten digits, as illustrated by the confusion matrix in
figure 5.
TABLE 2: Performance measures using the mentioned algorithms on the MNIST dataset
Algorithms AUC ACC F1 PREC REC
kNN 0.988 0.928 0.928 0.929 0.928
SVM w/RBF Kernel 0.997 0.924 0.924 0.928 0.924
SVM w/Linear Kernel 0.995 0.906 0.905 0.907 0.906
Logistic Regression 0.999 0.976 0.976 0.976 0.976
Neural Network 1.000 0.978 0.978 0.978 0.978
Random Forest 0.978 0.852 0.852 0.853 0.852
Naïve Bayes 0.767 0.769 0.774 0.767
Tree 0.832 0.748 0.747 0.747 0.748
FIGURE 5. NN Confusion Matrix on MNIST Dataset
4.4. HDC dataset results
Table 3 demonstrates the result of applying the studied algorithms to the HDC dataset. The metrics
values of the SVM, Logistic Regression, and Neural Network models are similar and represent the highest
16
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
values among the other algorithms as they have values not less than 0.954 that reach 0.999. The Tree model,
however, presents significantly low-performance measures compared to the rest of the models.
TABLE 3: Performance measures using the mentioned algorithms on the HDC dataset
Algorithms AUC ACC F1 PREC REC
kNN 0.985 0.913 0.912 0.914 0.913
SVM w/RBF Kernel 0.998 0.954 0.954 0.954 0.954
SVM w/Linear Kernel 0.997 0.935 0.935 0.936 0.935
Logistic Regression 0.999 0.973 0.973 0.973 0.973
Neural Network 0.999 0.974 0.974 0.974 0.978
Random Forest 0.977 0.850 0.850 0.850 0.850
Naïve Bayes 0.789 0.789 0.790 0.789
Tree 0.837 0.730 0.730 0.730 0.730
5. Discussion
All classifiers performed comparably well, as shown in Table 2. However, both ANN and SVM and
Logistic Regression performed exceptionally well. The first two models mentioned convert data into higher
dimensional spaces, which is the responsibility of the kernel in SVM, and the hidden layers in neural
networks. This allows cleaner separation between classes. However, increasing the dimensional space of
the dataset also increases the risk of overfitting the model, which can be noticed slightly in the recall score
of the SVM model, which used an RBF Kernel. This is where algorithms such as Logistic Regression may
outperform others. Logistic Regression does not change the dimensional spaces of the data but rather by
predicting the likelihood of classes using the sigmoid function based on some input parameters. Other
models, such as the Decision Tree and Naive Bayes, performed poorly compared to the abovementioned
models. kNN is a rather simple model; it does not require training, nor does it have a loss function that
requires minimization by training. When making a prediction, it searches the data and computes to find the
nearest k data points during runtime. This makes kNN vulnerable to outliers and noise in datasets. Both
kNN and decision trees are non-parametric. Decision trees have a high chance of growing complexity and
overfitting when applied to large datasets, especially image datasets. Both datasets were used to indicate
very similar conclusions. Overall, the difference in performance between the top classifiers was fairly small
to make definitive conclusions about the preferred model.
6. Conclusion
The field of OCR is very broad; even though significant work has been accomplished throughout the
years, there still is room for more to be done. The main purpose of this paper was to showcase the
performance of various machine learning models in handwritten digit recognition. The models, in
comparison, were trained on two fairly medium-sized datasets, 1 of them being the popular MNIST dataset.
After thoroughly analyzing the performance metrics, it was proven that neural networks tend to outperform
classical Machine Learning algorithms. However, the performance margin between them is insufficient to
conclude that classical machine learning algorithms are not ideal.
References
[1] B. V. Dasarathy, "Nearest neighbor (NN) norms: NN pattern classification techniques," IEEE Computer
Society Tutorial, 1991.
[2] E. Hancer, B. Xue, and M. Zhang, "Differential evolution for filter feature selection based on information
theory and feature ranking," Knowledge-Based Systems, vol. 140, pp. 103-119, 2018.
[3] Y. LeCun, "The MNIST database of handwritten digits," [Link] , 1998.
[4] R. Kohavi and G. H. John, "Wrappers for feature subset selection," Artificial intelligence, vol. 97, no. 1-2, pp.
273-324, 1997.
17
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
[5] Y. Chychkarov, A. Serhiienko, I. Syrmamiikh, and A. Kargin, "Handwritten Digits Recognition Using SVM,
KNN, RF and Deep Learning Neural Networks," in CMIS, 2021, pp. 496-509.
[6] O. M. Khanday and S. Dadvandipour, "Analysis of machine learning algorithms for character recognition: a
case study on handwritten digit recognition," Indonesian Journal of Electrical Engineering and Computer
Science, vol. 21, no. 1, pp. 574-581, 2021.
[7] S. Ali, Z. Shaukat, M. Azeem, Z. Sakhawat, and T. Mahmood, "An efficient and improved scheme for
handwritten digit recognition based on convolutional neural network," SN Applied Sciences, vol. 1, no. 9, pp.
1-9, 2019.
[8] O. Voloshchenko and M. Plechawska-Wójcik, "Comparison of classical machine learning algorithms in the
task of handwritten digits classification," Journal of Computer Sciences Institute, vol. 21, 2021.
[9] S. Ghosh, "Comparative Analysis of Boosting Algorithms Over MNIST Handwritten Digit Dataset," in
Evolutionary Computing and Mobile Sustainable Networks: Springer, 2022, pp. 985-995.
[10] H.-h. Zhao and H. Liu, "Multiple classifiers fusion and CNN feature extraction for handwritten digits
recognition," Granular Computing, vol. 5, no. 3, pp. 411-418, 2020.
[11] A. O. Otiko, J. A. Odey, and G. A. Inyang, "HANDWRITTEN DIGIT RECOGNITION: A PERFORMANCE
STUDY OF MACHINE LEARNING TOOLS."
[12] K. Senthil Kumar, S. Kumar, and A. Tiwari, "Realtime Handwritten Digit Recognition Using Keras Sequential
Model and Pygame," in Proceedings of 6th International Conference on Recent Trends in Computing, 2021:
Springer, pp. 251-260.
[13] M. B. Abdulrazzaq and J. N. Saeed, "A comparison of three classification algorithms for handwritten digit
recognition," in 2019 International Conference on Advanced Science and Engineering (ICOASE), 2019:
IEEE, pp. 58-63.
[14] R. KARAKAYA and S. KAZAN, "Handwritten digit recognition using machine learning," Sakarya
University Journal of Science, vol. 25, no. 1, pp. 65-71, 2021.
[15] K. Derdour, H. Mouss, and R. Bensaadi, "Multiple Features Extraction and Classifiers Combination Based
Handwriting Digit Recognition," International Journal on Electrical Engineering and Informatics, vol. 13,
no. 1, pp. 163-178, 2021.
[16] Abd Elminaam, D.S., El Tanany, A., Salam, M.A. and Abd El Fattah, M., 2022, May. CPSMP_ML: Closing
price Prediction of Stock Market using Machine Learning Models. In 2022 2nd International Mobile,
Intelligent, and Ubiquitous Computing Conference (MIUCC) (pp. 251-255). IEEE.
[17] Ai, M.A., Shanmugam, A., Muthusamy, S., Viswanathan, C., Panchal, H., Krishnamoorthy, M., Elminaam,
D.S.A. and Orban, R., 2022. Real-time facemask detection for preventing COVID-19 spread using transfer
learning based deep neural network. Electronics, 11(14), p.2250.
[18] Neggaz, N. and AbdElminaam, D.S., 2021, May. Automatic sport video mining using a novel fusion of
handcrafted descriptors. In 2021 International Mobile, Intelligent, and Ubiquitous Computing Conference
(MIUCC) (pp. 387-394). IEEE.
[19] Salam, M.A., Ibrahim, L. and Abdelminaam, D.S., 2021. Earthquake Prediction using Hybrid Machine
Learning Techniques. International Journal of Advanced Computer Science and Applications, 12(5), pp.654-
6652021.
[20] Mahmoud, E., Kader, H.A. and Minaam, D.A., 2019, October. Fuzzy knowledge base system for floating car
data on SUMO. In 2019 29th International Conference on Computer Theory and Applications (ICCTA) (pp.
38-42). IEEE.
[21] AbdElminaam, D.S., ElMasry, N., Talaat, Y., Adel, M., Hisham, A., Atef, K., Mohamed, A. and Akram, M.,
2021, May. HR-chat bot: Designing and building effective interview chat-bots for fake CV detection. In 2021
International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC) (pp. 403-408). IEEE.
[22] AbdElminaam, D.S., Fahmy, A.G., Ali, Y.M., El-Din, O.A.D. and Heidar, M., 2022, May. DeepECG:
Building an Efficient Framework for Automatic Arrhythmia classification model. In 2022 2nd International
Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC) (pp. 203-209). IEEE.
[23] AbdElminaam, D.S., Fahmy, A.G., Ali, Y.M., El-Din, O.A.D., Aly, A.R. and Heidar, M., 2022, May. ESEEG:
An Efficient Epileptic Seizure Detection using EEG signals based on Machine Learning Algorithms. In 2022
2nd International Mobile, Intelligent, and Ubiquitous Computing Conference (MIUCC) (pp. 210-215). IEEE.
[24] AbdElminaam, D.S., Ahmed, N., Yasser, M., Ahmed, R., George, P. and Sahhar, M., 2022, May.
DeepCorrect: Building an Efficient Framework for Auto Correction for Subjective Questions Using
GRU_LSTM Deep Learning. In 2022 2nd International Mobile, Intelligent, and Ubiquitous Computing
Conference (MIUCC) (pp. 33-40). IEEE.
18
Diaa Salam Abd Elminaam et al. Journal of Computing and Communication Vol.2 , No.1 , PP. 9-19 , 2023
[25] Ali, M.A., Orban, R., Rajammal Ramasamy, R., Muthusamy, S., Subramani, S., Sekar, K., Rajeena PP, F.,
Gomaa, I.A.E., Abulaigh, L. and Elminaam, D.S.A., 2022. A Novel Method for Survival Prediction of
Hepatocellular Carcinoma Using Feature-Selection Techniques. Applied Sciences, 12(13), p.6427.
[26] Pradheep. Handwritten Digit Classification dataset.
[27] M. Mohammed, M. B. Khan, and E. B. M. Bashier, Machine learning: algorithms and applications. Crc Press,
2016.
[28] B. Mahesh, "Machine learning algorithms-a review," International Journal of Science and Research
(IJSR).[Internet], vol. 9, pp. 381-386, 2020.
[29] T. G. Nick and K. M. Campbell, "Logistic regression," Topics in biostatistics, pp. 273-301, 2007.
[30] R. Y. Choi, A. S. Coyner, J. Kalpathy-Cramer, M. F. Chiang, and J. P. Campbell, "Introduction to machine
learning, neural networks, and deep learning," Translational Vision Science & Technology, vol. 9, no. 2, pp.
14-14, 2020.
[31] Y. Liu, Y. Wang, and J. Zhang, "New machine learning algorithm: Random forest," in International
Conference on Information Computing and Applications, 2012: Springer, pp. 246-252.
[32] G. Bonaccorso, Machine learning algorithms. Packt Publishing Ltd, 2017.
[33] V. Nasteski, "An overview of the supervised machine learning methods," Horizons. b, vol. 4, pp. 51-62, 2017.
19
