IOP Conference Series: Materials Science and Engineering

PAPER • OPEN ACCESS

Machine learning & image processing for hand written digits and
alphabets recognition from document image through MATLAB simulation
To cite this article: Satrughan Kumar Singh and Jainath Yadav 2021 IOP Conf. Ser.: Mater. Sci. Eng. 1084 012021

View the article online for updates and enhancements.

This content was downloaded from IP address 42.108.245.93 on 24/08/2021 at 08:25

ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

Machine Learning & Image Processing for hand written digits and
alphabets recognition from document image through MATLAB
simulation

Satrughan Kumar Singh1 and Jainath Yadav 2

1
CSIR - Central Institute of Mining & Fuel Research, Dhanbad – 826015, Jharkhand,
India
2
Central University of South Bihar, Gaya-824236, Bihar, India
satrughanksingh@yahoo.com, jainath@cub.ac.in

Abstract. Machine learning is currently playing a vital role in the next generation of the
computer world. Automatic pattern recognition has become an important issue of image
processing and machine learning. Handwritten digits and alphabets are not arranged in the same
size, thickness, position and right direction. Therefore, to determine the issue of handwritten
numerals and alphabet recognition, different classifications and complexity should be analyzed.
The composition styles of different individuals affect mainly the patterns of alphabets and digits.
An effective strategy is to understand the numbers and alphabets transferred from a document
image and to make an orderly pattern. In this research work, a soft computing system has been
developed using the MATLAB programming. This system uses machine learning algorithms that
identify patterns by computerized estimation by identifying handwritten digits and alphabets
from a document image. From the experimental results, we observed 96.24% average
recognition accuracy of our proposed system.

Keywords- Image processing; machine learning; MATLAB; document image; SVM;

digit & alphabet recognition; image segmentation; soft computing;

1. INTRODUCTION

Image processing and computer vision have become a key component of an emerging technology of
machine learning. It has become a unique and innovative integration of aspects of machine learning with
image feature extraction in image processing. Object detection and image segmentation are used in
image processing and computer vision to identify well-defined patterns. The Support Vector Machine
(SVM) algorithm facilitates image classification with segmentation, and it removes noise from the
image and constructs an optimal hyperplane for separating different classes that are processed by the
multi-dimensional phase.

The pattern recognition of handwritten numerals and alphabets is a complex task due to uniqueness
and variation in the writing styles of handwriting. When we scan the handwritten digits and alphabets
in the document image, a complex and large amount of noise is created which complicates the pattern
recognition of the handwritten digits and alphabets. Pattern recognition has been a frontline research
field in the field of human-machine interface for the last few decades. In present times, people are
constantly trying to make computers intelligent so that they can do almost all the work easily like
humans. This intelligent computer can not only reduce human effort but also save time. Classifying
patterns from a documented image is one of the major implementations of image processing and
machine learning. Offline handwritten alphabet or digit recognition is a unique process of
characterization and extraction of alphabets and numbers through segmentation by automated systems.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

2. LITERATURE SURVEY
The recognition accuracy of the digits and alphabets depends on the sensitivity of the selected
features and the classifiers. Hence, several feature extraction and classification methods can be
found in the literature. The recognition accuracy depends on the sensitivity of the extracted features
using the SVM classifier. There are several numbers of feature selection and classification methods
that were already performed on numeral & alphabet analysis and pattern recognition. The following
research works performing the evolution of digit & character recognition for pattern recognition
are shown in Table 1.

Table 1. The recognition accuracy of digits & alphabets using various methods given in the
literature
Accuracy (%)
Ref.
Author(s) Model / Classifier
No. Capital Small
Digit
Alphabet Alphabet
[1] Gupta, A. et al. NN & SVM 62.93 - -
[2] Bellili, A. et al. MLP-SVM 98.01 - -
[3] Nasien, D. et al. SVM - 88.46 86.00
[4] Priya et al. NN & SVM 98.4 - -
[5] Vamvakas, G. et al. SVM - 80.19 -
[6] Gattal, A. et al. SVM 95.21 - -
[7] Neves, R.F.P. et al. SVM 97.94 - -
[8] Kadam, D. et al. LBP-SVM - 96.5 98.00
[9] Khedidja, D. et al. SVM 99.89 - -
[10] Mishra, A. et al. SVM 96.29 - -

3. METHODOLOGY
The handwritten digits and alphabets recognition systems incorporate digitization, pre-processing,
segmentation, attribute selection and tracing, training datasets, validation of datasets, testing
datasets, and attribute validation steps for pattern recognition. The block diagram of the proposed
model for the pattern recognition of the handwritten digit or alphabet is shown in Figure 1.

Figure 1. The block diagram of the proposed approach

2
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

In this work, samples of different handwritten numerals and alphabets of 8 authors have been collected.
Figure 2 shows a handwritten dataset of 8 authors (S1, S2… S8). The proposed method has been
discussed in the following sub-sections.

3.1. Document digitalization

Document digitization is the normal process of converting handwritten and other documents into
electronic and digital forms. Electronic and digital conversion is used to scan a document that is an
electronic representation of the original document in the form of an image of a document or image file.
Digitization is an initial step that executes before the pre-processing step. The digitization phase
produces a digital image of the original document which is processed in the pre-processing phase.

Figure 2. Samples of datasets for handwritten digits and alphabets

3.2. Pre-processing
In the pre-processing phase, the unimportant data information is removed from the digital image of the
original document which can adversely affect the recognition accuracy. This step can include steps such
as binarization, noise removal, skew detection, and skeletalization that can help in completing correctly
the appropriate algorithms for the decimation system and devising precise strategies. The major role of
the pre-processing is to filter out the impurities from the image and also to perform smoothing and
normalization[11]. Pre-processing can produce a noise-free feature as well as an appropriate image set
of alphabets and numbers for effective selection. The following are the various steps of pre-processing.

3.2.1. Binarization: The binarization is the stepwise process of converting a gray-scale image into a
binary image which is shown as 0 and 1. In the binary image, the visible part of the image can be
represented by digit 1 and the invisible part by digit 0 as shown in Figure 3. Generally, the scanned

3
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

image of the document does not align correctly in the horizontal direction, so we need to adjust it by
skew angle correction. This can reduce its ability to convert gray-scale images into binary images that
are necessary for reducing dimensions to increase processing speed.

Figure 3. The binarization process for the handwritten symbol recognition

3.2.2. Noise elimination: Noise elimination is used to remove any undesirable or meaningless bit
patterns and noise from binary images. Noise is an undesired disturbance that creates unwanted errors.
Such interference would have to be eliminated for maintaining sufficient accuracy.

3.2.3. Skew detection: Generally, the digital image of the document does not fit directly, so the skew
detection process is required for the recognition of the document's digital image.

3.2.4. Skeleton: Typically there is opposition to the width of the line from several pixels to one pixel
wide. In this case, the skeletalization process can refer to reducing line width by removing many
discrepancies, to simplify the classification algorithm and at the same time reduce processing time.

3.3. Segmentation
Segmentation is an impartment process of pattern recognition of any digit and alphabet symbol that
determines the components of the document image. Image segmentation is a very important problem
for estimating and acquiring information from an image[12]. It is essential to locate the areas of a
document image where the data is printed and their figures and graphics are different. In image
segmentation, the image is divided into several square pixels[12]. The segmentation technique identifies
the area of interest of a picture by using criteria such as color, composition, and excess[12]. It identifies
the end and appearance of the image as it appears on the computer screen[12].

3.4. Feature selection or extraction

The feature selection method can be used to choose important features and to remove inconsistencies
and redundancies from these pre-processed image datasets. Diagonal features play a vital role to achieve
higher accuracy of the pattern recognition system. The attribute linking shows the applied information
of the shape enclosed in the pattern to classify the images to simplify the validation technique.
Handwritten characters include many specific elements such as diagonal component extraction, chain
code, scale-invariant, etc.

3.5. Training dataset

In this phase, the training algorithm is used to train the classifier using the input dataset to fit the model
parameters such as model weights. The model is trained on training supervisors, who use a supervised
learning method. The training dataset usually consists of a pair of input and related output vectors, where

4
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

the answer key is shown as the target. The current model has to extract a result by running it with the
training dataset, and the model's parameters can be adjusted based on the result.

3.6. Validation of dataset

In dataset validation, the model that fits on the training dataset can be objectively evaluated by tuning
the model. Validation of datasets can be used to routinely stop training when errors escalate due to
overfitting of the training dataset. This process becomes complicated by the error fluctuations of the
validation of the dataset during training that generates multiple local minima.

3.7. Testing dataset

It evaluates the model developed from training dataset. If the data in the test dataset has never been
used in training, it can be holdout under cross-validation to be re-trained. If a model fits into the training
dataset, it can be fully fitted to the test dataset with minimal overfitting. Generally, better fitting of
training datasets indicates overfitting as opposed to testing datasets. The test set is therefore an example
of making a proper assessment based on the performance of a fully specified classifier.

4. SUPPORT VECTOR MACHINE

Support vector machine (SVM) constructs an optimal hyperplane (multiplane) in multi-
dimensional space separating different classes, which can be used to minimize an error. The main
task of SVM during the model learning phase is to find the optimal hyperplane with maximum
margin that can divide the dataset into classes and also estimate the width between classes. The
data points which are closest to the hyperplane is called support vectors. These support vectors can
better define a separate line by calculating the point margin that may be most relevant for
constructing digit and alphabet classifiers. Figure 4 shows the maximum margin and optimal
hyperplane between classes.

Figure 4. Maximum margin and optimal hyperplane between classes

5. RESULT AND DISCUSSION

In this work, we have used the confusion matrix terminology to evaluate the performance of the
proposed method. The confusion matrix is sometimes called the error matrix, which can be used as
a matrix representation on a set of test data in a problem of machine learning and statistical
classification. Usually, one class is confused with another which can allow easy identification of
confusion between classes. In collimation, the confusion matrix shows how the classification model
is confounded when it makes predictions and then provides insight into the errors made by the
classifier based on these predictions.
In our study, we have collected the scanned symbols written by eight persons. For each person, we
have collected 10 instances of each symbol. In our database, the total numbers of symbols are 4960
(62 symbols × 8 persons × 10 instances). The total numbers of symbol instances used for training

5
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

and testing are 3472 and 1488 respectively. The recognition accuracy of the proposed method is
shown in Table 2. From the table, we observe that the average recognition accuracy of capital
alphabets, small alphabets and digits are 95.37%, 96.27% and 97.08% respectively. Among the
capital alphabets, ‘U' symbol has highest recognition accuracy and ‘B’ symbol has lowest
recognition accuracy. Among small alphabet, ‘k’ and ‘p’ have highest and lowest recognition
accuracy respectively. Similarly, we observed highest and lowest recognition accuracy for digits
‘7’ and ‘4’ among the digits respectively. It can be observed that for the specific pattern, the
proposed method recognized all the alphabets & numerals. We have developed a technique for
recognition of scanned or handwritten alphabets and digits by density method and potential
application. This experimental work on handwritten pattern recognition uses sequential rules to
segment alphabets and digits from document images for classification using the SVM classifier
model. The classifier also describes the pattern recognition results to improve classification
performance and error evaluation after the pre-processing stage. The graphical interface of the
handwritten pattern recognition has been illustrated through Figure 5.

Figure 5. GUI Screenshot of handwritten pattern recognition

Table 2. Accuracy of handwritten digits & alphabets of authors (S1-S8) using proposed method
Accuracy (in %) Average
Symbol Accurac
S1 S2 S3 S4 S5 S6 S7 S8 y (in %)
A 99.50 98.57 98.57 98.07 96.82 96.57 96.82 99.07 98.00
B 87.07 89.70 89.70 89.20 87.95 87.70 87.95 90.20 88.68
C 92.26 92.79 92.79 92.29 91.04 90.79 91.04 93.29 92.04
D 97.12 97.78 97.78 97.28 96.03 95.78 96.03 98.28 97.01
E 98.84 93.17 93.17 92.67 91.42 91.17 91.67 93.67 93.22
F 98.46 97.79 97.79 97.29 96.04 95.79 96.29 98.29 97.22
G 98.86 95.51 95.51 95.01 93.76 93.51 94.01 96.01 95.27
H 87.43 93.95 93.95 93.45 92.20 91.95 92.70 94.45 92.51
I 99.29 98.98 98.98 98.48 97.23 96.98 97.23 99.48 98.33
J 98.66 97.36 97.36 96.86 95.61 95.36 95.86 97.86 96.87
K 94.65 96.21 96.21 95.71 94.46 94.21 94.71 96.71 95.36
L 96.90 96.40 96.40 95.90 94.65 94.40 94.90 96.90 95.81
M 91.20 90.70 90.70 90.20 88.95 88.70 89.45 91.20 90.14
N 97.76 97.26 97.26 96.76 95.51 95.26 95.76 97.76 96.67
O 98.93 98.43 98.43 97.93 96.68 96.43 96.93 98.93 97.84
P 93.20 92.70 92.70 92.20 90.95 90.70 90.95 93.20 92.08
Q 94.26 93.76 93.76 93.26 92.01 91.76 92.51 94.26 93.20
R 98.10 97.60 97.60 97.10 95.85 95.60 95.85 98.10 96.98
S 97.73 97.23 97.23 96.73 95.48 95.23 95.73 97.73 96.64

6
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

T 98.85 98.35 98.35 97.85 96.60 96.35 96.85 98.85 97.76

U 99.47 98.97 98.97 98.47 97.22 96.97 97.47 99.47 98.38
V 98.74 98.24 98.24 97.74 96.49 96.24 96.74 98.74 97.65
W 95.08 94.58 94.58 94.08 92.83 92.58 93.08 95.08 93.99
X 95.97 95.47 95.47 94.97 93.72 93.47 93.97 95.97 94.88
Y 98.00 97.50 97.50 97.00 95.75 95.50 96.00 98.00 96.91
Z 97.38 96.88 96.88 96.38 95.13 94.88 95.38 97.38 96.29
a 98.44 97.94 97.94 97.44 96.19 95.94 96.44 98.44 97.35
b 99.16 98.66 98.66 98.16 96.91 96.66 97.41 99.16 98.10
c 99.80 94.78 94.78 94.28 93.03 92.78 93.53 95.28 94.78
d 96.40 95.90 95.90 95.40 94.15 93.90 94.65 96.40 95.34
e 98.92 98.42 98.42 97.92 96.67 96.42 97.17 98.92 97.86
f 96.94 96.44 96.44 95.94 94.69 94.44 94.94 96.94 95.85
g 92.95 92.45 92.45 91.95 90.70 90.45 90.95 92.95 91.86
h 96.96 96.46 96.46 95.96 94.71 94.46 94.96 96.96 95.87
i 94.46 93.96 93.96 93.46 92.21 91.96 92.46 94.46 93.37
j 95.46 94.96 94.96 94.46 93.21 92.96 93.46 95.46 94.37
k 99.97 99.47 99.47 98.97 97.72 97.47 98.22 99.97 98.91
l 99.47 98.97 98.97 98.47 97.22 96.97 97.72 99.47 98.41
m 99.83 99.33 99.33 98.83 97.58 97.33 97.83 99.83 98.74
n 99.76 99.26 99.26 98.76 97.51 97.26 97.76 99.76 98.67
o 99.80 99.30 99.30 98.80 97.55 97.30 97.80 99.80 98.71
p 91.40 90.90 90.90 90.40 89.15 88.90 89.40 91.40 90.31
q 94.42 93.92 93.92 93.42 92.17 91.92 92.42 94.42 93.33
r 96.44 95.94 95.94 95.44 94.19 93.94 94.44 96.44 95.35
s 99.48 98.98 98.98 98.48 97.23 96.98 97.23 99.48 98.36
t 97.84 97.34 97.34 96.84 95.59 95.34 95.84 97.84 96.75
u 95.46 94.96 94.96 94.46 93.21 92.96 93.46 95.46 94.37
v 99.96 99.46 99.46 98.96 97.71 97.46 97.96 99.96 98.87
w 96.47 95.97 95.97 95.47 94.22 93.97 94.47 96.47 95.38
x 97.47 96.97 96.97 96.47 95.22 94.97 95.22 97.47 96.35
y 98.83 98.33 98.33 97.83 96.58 96.33 96.83 98.83 97.74
z 99.02 98.52 98.52 98.02 96.77 96.52 96.77 99.02 97.90
0 99.40 98.90 98.90 98.40 97.15 96.90 97.40 99.40 98.31
1 99.72 99.22 99.22 98.72 97.47 97.22 97.97 99.72 98.66
2 97.26 96.76 96.76 96.26 95.01 94.76 95.26 97.26 96.17
3 98.30 97.80 97.80 97.30 96.05 95.80 96.05 98.30 97.18
4 94.40 93.90 93.90 93.40 92.15 91.90 92.15 94.40 93.28
5 94.92 94.42 94.42 93.92 92.67 92.42 92.67 94.92 93.80
6 99.96 99.46 99.46 98.96 97.71 97.46 97.71 99.96 98.84
7 99.95 99.45 99.45 98.95 97.70 97.45 98.20 99.95 98.89
8 99.98 99.48 99.48 98.98 97.73 97.48 97.73 99.98 98.86
9 97.96 97.46 97.46 96.96 95.71 95.46 95.71 97.96 96.84

6. CONCLUSION
This research work deals with the recognition of handwritten numerals and alphabets by applying
the support vector machine technique. However, many variations of the same alphabet or number,
with different text styles and sizes for recognition, become complex and difficult. Therefore, the
SVM algorithm can be used for optimal pattern recognition from several attributes. In this task,
offline handwritten numeral and alphabet recognition is used for training and managing the dataset,
which is obtained manually or from a scanned document of handwritten numerals and alphabets.
The main work focuses on classifying digits and alphabets in a digitally desirable format so that
they can be easily modified and processed by the machine intelligence system. The proposed
research work provides more efficient and accurate results that obtained an overall 97-99%
recognition rate. Therefore, SVM works well with a clear & best margin of separation and high
dimensional space. The recognition rate needs to be tested by increasing size of datasets for future
work.

7
ICCSSS 2020 IOP Publishing
IOP Conf. Series: Materials Science and Engineering 1084 (2021) 012021 doi:10.1088/1757-899X/1084/1/012021

References

[1] Gupta, A., Srivastava, M., Mahanta, C.: Offline handwritten character recognition using neural
network. In: International Conference on Computer Applications and Industrial Electronics (ICCAIE),
pp. 102-107. 2011 IEEE International Conference on IEEE (2011)
[2] Bellili, A. et al.: An hybrid MLP-SVM handwritten digit recognizer. In: Conference Proceedings of
Sixth International Conference on Document Analysis and Recognition, vol. 1, pp. 0028. (2001)
[3] Nasien, D., Haron, H., Yuhaniz., S.S.: Support Vector Machine (SVM) for English handwritten
character recognition. In: 2010 Second International Conference on Computer Engineering and
Applications, (2010)
[4] Priya et al.: Review on handwritten digit recognition. In: International Journal of Novel Research and
Development, vol. 2, Issue 4, pp. 91-94. April (2017)
[5] Vamvakas, G., Gatos, B., Perantonis, S.J.: A Novel Feature Extraction and Classification Methodology
for the Recognition of Historical Documents. In: proceeding of IEEE 10th International Conference on
Document Analysis and Recognition, pp. 491-495. (2009)
[6] Gattal, A. et al.: Isolated Handwritten Digit Recognition Using oBIFs and Background Features. In:
Conference Proceedings 12th IAPR Workshop on Document Analysis Systems (DAS), vol. 1, pp. 305-
310. (2016)
[7] Neves, R.F.P. et al.: A SVM Based Off-Line Handwritten Digit Recognizer. In: International
conference on Systems, Man and Cybernetics, pp. 510-515. IEEE Xplore, (2011)
[8] Kadam, D., Chavan, P., Pandhara, P.: Literature Survey on Recognition and Evaluation of Optical
Character Recognition (OCR). In: International Journal of Scientific & Engineering Research, vol. 9,
Issue 2, pp. 72-75. February (2018)
[9] Khedidja, D., Hayet, M.: Multiple Classifiers and Invariant Features Extraction for Digit Recognition.
In: International Journal of Computer Electrical Engineering, vol. 11, Number 1, pp. 41-52. March
(2019)
[10] Mishra, A., Singh, D.: Handwritten digit recognition using combined feature extraction technique and
neural network. In: Computer Modelling & New Technologies, vol. 21, issue 2, pp. 80-88.
[11] Sivaraman, P “A New Method of Maximum Power Point Tracking for Maximizing the Power
Generation from a SPV Plant” Journal of scientific and Industrial Research Vol.74, No.3 pp.411 - 415
AUG 2015.
[12] Senthil kumar J Charles Raja S Dipti S Venkatesh P 2018 Hybrid Renewable Energy based Distribution
System for Seasonal Load Variations International Journal of Energy Research 42 3 1066 - 1087 2018
[13] Kadam, D. et al.: Literature Survey on Recognition and Evaluation of Optical Character Recognition
(OCR). In: International Journal of Scientific & Engineering Research, vol. 9, Issue 2, pp. 72-75.
February (2018)
[14] Veerakumar Nnirmalkumar Sathishkumar Rajesh Novel harmonic elimination technique for cascaded
h-bridge inverter using sampled reference frame, Journal of Theoretical and Applied Information
Technology Vol. 58 No.2 2013