Access to this full-text is provided by IGI Global.
Content available from International Journal of Knowledge Management
This content is subject to copyright. Terms and conditions apply.
DOI: 10.4018/IJKM.2021100103
Volume 17 • Issue 4
This article, originally published under IGI Global’s copyright on October 1, 2021 will proceed with publication as an Open Access article
starting on February 15, 2024 in the gold Open Access journal, International Jour nal of Knowledge Management (converted to gold Open
Access January 1, 2022), and will be distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/
licenses/by/4.0/) which permits unrestricted use, distribution, and production in any medium, provided the author of the original work and
original publication source are properly credited.
*Corresponding Author
52
Rohit Rastogi, ABES Engineering College, Ghaziabad, India
https://orcid.org/0000-0002-6402-7638
Himanshu Upadhyay, ABES Engineering College, Ghaziabad, India
Akshit Rajan Rastogi, ABES Engineering College, Ghaziabad, India
Divya Sharma, ABES Engineering College, Ghaziabad, India
Prankur Bishnoi, ABES Engineering College, Ghaziabad, India
Ankit Kumar, ABES Engineering College, Ghaziabad, India
Abhinav Tyagi, ABES Engineering College, Ghaziabad, India
In handwriting recognition, traditional systems have relied heavily on handcrafted features and a
massive amount of prior data and knowledge. Deep learning techniques have been the focus of
research in the field of handwriting digit recognition and have achieved breakthrough performance in
the last few years for knowledge extraction and management. KM and knowledge pyramid helps the
project with its relationship with big data and IoT. The layers were selected randomly by which the
performance of all the cases was found different. Data layers of the knowledge pyramid are formed
by the sensors and input devices, whereas knowledge layers are the result of knowledge extraction
applied on data layers. The knowledge pyramid and KM helps in making the use of IoT and big data
easily. In this manuscript, the knowledge management principles capture the handwritten gestures
numerically and get it recognized correctly by the software. The application of AI and DNN has
increased the acceptability significantly. The accuracy is better than other available software on the
market.
Convolutional Neural Network (CNN), Handwritten Digits Recognition, Hidden Layers, Machine Learning,
MNIST Dataset
Handwriting recognition is a process, which has been a centre of attraction for analysts since 1998,
with nearly all the calculations structured at a point. In this process, the error rate obtained by direct
classifier was 12% in 1998, which gradually decreased to 0.23% in 2012 when applied by convolution
nets. Nowadays, as there is a drastic increase in the amount of information, Researchers and AI
specialists are trying their level best to create and approve solo learning strategies, for example,
auto-encoders, profound learning models, etc.
Volume 17 • Issue 4
53
In this research work, a handwriting recognition framework is actualized with the celebrated
MNIST informational index. This index is a well known dataset. It consists of a large database, which
is required for training and assessment of various image processing systems. Machine learning is
one of the fields where this data is regularly used in order to test and train various systems. By using
KM and converting all the data achieved to knowledge it can help in solving many other problems.
Digit recognition system (DRS) is a subfield of image processing that is used to scan the image
or documents with the help of the MNIST database or dataset. The main goal of this research work
is to implement a project and discuss its accuracy which scans the image digits using basic image
correlation, also known as matrix matching. The six cases performed different outputs as shown in
subsequent sections, all the cases are different from each other which show different accuracy and
different performances. Still, the rapid and high growth of the handwritten numerical data generated
recently and rising availability of massive processing power urge for improvements in recognition
accuracy and deserves much needed further investigation (Hamad, K. et al., 2016).
Convolutional neural networks (CNNs) are found to be very effective in recognizing the structure
of handwritten digits/characters in many ways, which also helps in automatically finding distinct
features. This scenario makes CNN the best available solution for solving handwritten digit recognition
problems. We have introduced a suitable combination of learning parameters in designing a CNN
that helps us achieve a new absolute record in classifying MNIST handwritten digits. Knowledge
obtained can help in further training machines and making predictions and reducing computation
required in understanding the whole image.
Digit recognition systems are being used on a very large scale in today’s world because the handwritten
digits are not always in the correct size, shape or style. That’s why, a lot of problems may occur to
find the correct digit in the real world with the help of a machine (Dorosh, N. et al., 2020).
When we include ensemble architecture, it helps in increasing the network’s recognition accuracy.
So here, the objective is to accomplish comparable accuracy by using a pure CNN architecture without
using ensemble architecture. As, ensemble architecture introduces computational cost and high testing
complexity. Therefore, in order to acquire accuracy even better than ensemble architecture, we are
trying CNN architecture, with additional bonus profit of reduced operational complexity and cost.
This research work gives us the motivation to create a model using CNN (Convolution Neural
Network) that would be able to receive and identify digits in the form of images. Also, the goal of
the program is to learn about CNN, and apply it to handwriting recognition systems (Dorosh, N.
et al., 2020). The aim of this research paper is to find the accuracy of CNN to classify handwritten
digits using multiple hidden layers.
Our motive in this research work is to explore the various design possibilities like kernel size,
receptive field, and stride size, number of layers, padding and dilution for CNN-based handwritten
digit recognition.
Handwritten digit recognition is one of the practically important issues in pattern recognition
applications. Digits are used in different types of data like vehicle number plate, written data, meter
reading etc. In the modern world digit recognition plays an important role in many user authentication
applications. It can be used in vehicle number plate recognition which uses a number plate in order
to identify the vehicle. This vehicle number plate recognition system is used in the area around top
government offices like Parliament, Supreme Court etc.
Digit recognition can also be used in postal mail sorting, handwriting recognition, form data entry,
historical document preservation in archaeology departments, old document automation in libraries
Volume 17 • Issue 4
54
and banks. All these areas deal with large databases therefore demand high recognition accuracy,
lesser computational complexity and consistent performance of the recognition system. In future,
this study has major implications in hand written answers checking in examinations and human hand
writing notifications. It can also be used to perform the predictive analysis of any terrorist messages
after screening their content partially (Ahlawat, S. et al, 2020).
Gesture recognition is a means of human-machine interaction via body actions. The interaction of
human and computer takes place through traditional input devices like mouse, keyboard etc. Hand
gestures can act as a useful medium for the human-computer interaction and can make the interaction
easier. Gestures vary from person to person in orientation as well as shape. Hence non-linearity
exists here.
It has been proved in recent researches that Convolutional Neural Network (CNN) can be used for
image representation and classification as CNN can learn complex and non-linear relationships among
images. And also recently gesture recognition based on machine learning has been developed rapidly
in human-computer interaction as a graphics processor unit (GPU) and artificial intelligence(AI)
image processing has been introduced. Machine learning algorithms such as support vector machine
(SVM), neural networks are widely used in gesture recognition(Islam, M. et al, 2019).
Digit recognition systems can also prove helpful for the visually impaired persons. The input to
such a system is the pressure sensor data. This data is obtained from a sensor pad which has 256
capacitive pressure sensors. This system recognizes correctly the 2-D shape of a digit which is placed
on the sensor pad. In the first step of the recognition process the Euler number is used to achieve a
preliminary classification while in the second step fourth order central moment is applied to correctly
recognize the input. Also, it can be seen that deep learning using Convolutional Neural Networks is
becoming a popular technique for challenging computer vision applications. But deep learning’s power
consumption and bandwidth requirements limit its application in embedded and mobile systems with
tight energy budgets. Bio-inspired Angle Sensitive Pixels(ASP), custom CMOS diffractive image
sensors (Paul, A. et al, 2012).
The latest Knowledge Pyramid tries to set the knowledge hierarchy within the ambience of the true
world. This Knowledge pyramid is inverted (if compared with the old knowledge pyramid). Here,
there is more information than the data. Similarly more knowledge than information and more wisdom
than knowledge.
The data layers of this pyramid are the results of sensors and input devices, whereas the knowledge
layers are the result of knowledge extraction applied on data layers.
As per revised knowledge pyramid, the reality i.e. real world object, person or situation is in centre,
accordingly the suitable data is collected through sensors or social media on which analysis has to be
performed. After refining and preprocessing operations, information is obtained and through mining
exercises, knowledge and through visualizations and graphical analysis, AI and ML applications,
predictive models, wisdom is achieved. This wisdom is cream and can be used for individual to
organizational learning for different purposes.
Knowledge management is the technique or procedure that is used to capture, distribute and use
knowledge in an efficient manner. It is an approach used in various disciplines making best use of KM to
achieve organizational goals. Knowledge Management circumscribes psychology, epistemology, along
Volume 17 • Issue 4
55
with cognitive science. Better understanding of Knowledge Management can increase throughput,
innovation and can amplify what is known for both, an individual and organizational perspective.
In recent times, the ability to refine a significantly large volume of data, information and
knowledge to gain computational and competitive advantage and the significance of raw data and
text analytics to this effort is gaining strength.
The basic components of a knowledge management strategy can be made useful for the
management of pyramid activities with seven steps. The Knowledge Pyramid and KM helps in
making the use of IOT and Big data easily, opening new opportunities for future work too (Jennex,
M.E., 2017).
CNN is playing a major role in image processing and it is being used in natural small localized areas.
CNN is also used in Nano-technologies like classification (Dorosh, N. et al., 2020). There are many
different other methods of image processing but CNN is best because of using Keras with the one or
multiple layers, which provide multiple layers at a same time to detect fault and classification, that’s
why high accuracy takes place (Gulli, Antonio et a., 2020).
As shown in Figure-1 of CNN layers, there are seven layers that take place in CNN method in
which five layers are followed by one layer as input and other one remaining layer works as an output
layer which is used to recognize handwritten digits (as per Fig. 1).
Being a problem for well-differentiated, multi-machine learning techniques such as regression,
KNN, SVMs and Neural Networks have been investigated to obtain digits with high accuracy (as per
Fig. 2) (Umapada Pal et al., 2012).
As with each existing research, KNN and SVM in the thinking class may cause scalability
difficulties, and may be comparing the default of the second classifier.
Figure 1. A Seven Layered Convolution Neural Network for Digit Recognition
Volume 17 • Issue 4
56
To build a strong Neural Network for this project, we should make use of Knowledge Management(KM).
Knowledge Management (KM) involves several disciplines such as psychology, cognitive science etc.
The objective of KM is to get insights of the data that we are using and also enable people and firms to
cooperate, share, create, use and reuse knowledge. We have a very large amount of data present around
us and all over the internet but to acquire knowledge from that data we require understanding of KM.
Neural networks can be trained in a better way with the help of this knowledge (Jennex, M.E., 2017).
OmniPage Ultimate
This software is combined with artificial intelligence and neural networks. It provides the best
results.It includes more than 120 language recognition which makes sure that the converted text is
as accurate as possible. It has the latest OCR technology which picks up handwritten text very well.
It easily allows to edit, share and search the PDF document. Free trial is also available from their
official website. It reduces the amount of time spent in conversion of files, lower operational costs,
and managing and sharing documents (Hamad, K. et al, 2016).
TopOCR
It is one of the best handwriting recognition software. It is an optical character recognition application.
It uses a sourced image captured by a scanner or digital camera and offers a dual pane format which
displays the original image on the left and the conversion on the right. When the text is loaded on the
right-side panel, one can read it and also correct any mistakes. After the completion the text can be
saved to an array of different formats. It also has a text-to-speech engine. It is available in windows
only. It supports 11 languages and also has a PDF export feature. It’s free trial version is also available
(Sahu, N. et al, 2017).
Simple OCR
It is a freeware tool which recognizes approximately 120,000 words. It allows to add more words to
it’s dictionary. It has more than 97% accuracy. It even identifies formatted text and it is also possible
to set it to ignore formatting. It is a speedy tool. It is available on desktop only. It is the popular
freeware OCR software with hundreds of thousands of users worldwide. It is 100% free and not
limited in any way. It uses the despeckle or noisy document feature if the handwriting which is being
converted is messy. It can be set to decipher whole documents, portions, or multiple documents in
batches (Hamad, K. et al, 2016).
Software For Handwriting Recognition & Transformation For Dark Data
Since, everywhere in the globe, the govt. of many countries are facing pressure to be more active
and transparent in their functioning so all the processes and happenings has to be brought before
Figure 2. Error Rate of Different Classification Algorithms
Volume 17 • Issue 4
57
all to make the system transparent. Specially the secret operations, foreign policies, heavy defense
and healthcare purchases etc. are areas where the challenges are hidden transactions are in cursive
handwriting, signatures and constrained and unconstrained handprints has to be understood.
These software are able to detect and curb the high instruction on image and noise levels to
improve the accuracy and optimize the performances. Hand written data and fields available in free
formats are challenges and contain structure in big data which plays a crucial role in big organizations.
(Ref. https://www.parascript.com/blog/best-icr-software-for-handwriting/)
Besides this, the professionals are also trying to design the software which can read and understand
the complex and crypto coded human hand writing and so the terrorist messages can be understood
and can be dealt with in advance.
During the project we have used various tools and techniques to get best results. We used python as
our base tool for implementation of the text recognition through CNN algorithm with 5 layers along
with the MNIST dataset of numerals.
CNN algorithm is one of those algorithms which are being used prominently in the field of
computer vision. It is widely used in the field of object detection and recognition in an image with
a great efficiency as compared to other neural network algorithms.
This can be easily observed through “Figure 2” of this document.
MNIST stands for “Modified national institute of standards and technology”. It is a large sized dataset
widely used in the field of computer vision and convolutional neural networks(CNN). MNIST dataset
is used for training and testing the systems which includes numerals as data. it holds binary images
of handwritten digits.
This Dataset contains 60,000 images which is used for training and cross validation.
Additional 10,000 images are also present for testing purposes. All the images of dataset are
into 28*28 sizes which makes the total dimension vector as 28*28=784(Sharma, I., et al., 2020).
As shown figure 3, all the digits are in 28*28 size which makes a unique and stable combination
to recognize the handwritten digits.
As per the current structure of knowledge pyramid, since it is inverted now so, the knowledge obtained
from the data set will be raw and it will be refined from semi and fully structured data. To convert
into wisdom, complex mathematical operation will be performed so that the knowledge obtained will
be useful for individual and for organization and can be applied in designing much runtime software
to check fraud and crime. The knowledge management will play a crucial role here.
Step 1:
In this step, the handwritten document is scanned and digits are extracted from that document by the
system. Then, the digits are arranged in a sequence of characters as a string. This input can serve as
a prominent base for gathering various vital information derived from text extraction. The Devices
used in this process can be a digital camera or a scanner.
Volume 17 • Issue 4
58
Step 2:
In this step, extracted image’s digits gets resized to 28*28 dimension. This step is important because
the dimensions of a handwritten document’s data is specific to the writing sense and stroke style
of the writer. So this step brings the input data to a fixed standard size. Also, the size of images in
MNIST data set is with the dimensions of 28*28 which makes the comparison more convenient. This
modified data further can be used as a base data for other comparisons. This data can pile up and
contribute to a very large dataset i.e. Big Data. This big data can be used to give us vital information
which can give useful outcomes on processing.
Step 3:
In this step, multiple layers of CNN are employed on the dataset which makes the digits more clear.
Along with this, image processing techniques are employed to classify the input. This step brings the
handwritten document closer to the digital document. After the processing the handwritten document
becomes a more convenient source through which one can get some insight of the real data present
in the handwritten document file and can provide knowledge which can be used for decision making
and further uses.
Step 4:
In this step, handwritten digits are tested against the MNIST dataset. In this process, each numeral is
checked against every numeral present in the dataset giving out accuracy percentage.This phase can
be used to get and create wisdom for a particular individual.
Figure 3. Samples of MNIST Dataset
Volume 17 • Issue 4
59
All the Necessary Diagrams to explain the functionality of our Research work are mentioned below
( Fig. 5 to Fig. 8)
Here, the activity diagram depicts how the procedure is undertaken by the digit recognition system.
Firstly, after starting the process, we will upload the digit image. If the format of the image is not
supported, then the system will show an error message and will take us back to the initial stage. But
if the format of the image is supported, then the system will pass the subject image through the image
acquisition stage, and then through the image preprocessing stage where the image is converted to the
grayscale image, which is then converted to the binary format and after that, it is then normalized.
But, if the preprocessed image is not clear, then the system will again show an error message and
will again take us back to the initial stage. But, if the preprocessed image is clear, then it will pass
through various stages which include image segmentation, feature extraction and then the classification
or recognition. Then after completing all the steps, the system will generate a process at the end.
The network contains 784 neurons as input data which means the input layers consists of 28*28
pixels. on the grayscale the white pixel with the value of 0 and the black pixel with the value of 1.
Hence, CNN models have 5 hidden layers ( as per Equations 1,2 and 3) (as per Fig. 9 and 10).
For the performance of cost function the equation is expressed as -
The 1st hidden layer is the decision layer responsible for extraction in input data. Addition to
this, it contains many feature maps with readable headers and repair units.
The kernel size determines the filter area. The repair unit is used as a trigger function at the last
of each case layer and layer that is fully connected to enlarge model performance. Next hidden layer
pool layer 1. Reduces flow information to the display layer then decreases the no. of parameters
again the complexity of the model. Various large pool lakes, central swimming, and L2 swimming.
Figure 4. Simple Steps for Digits Recognition with MNIST
Volume 17 • Issue 4
60
Here, maximum swimming used to decrease in size maps for each feature. Conversion layer 1
and 2 have the same function as output layer 1 and also no ponding layer 2.
To decrease the excess weight, stop the standard method is applied to a fully connected layer
1.It occasionally quenches specific neurons during training in order to improve network performance
by providing more efficient mightily. The output of the layer network has 10 neurons and finds the
digits from 0 - 9 output layers.
This is done by an algorithm known as gradient descent (as per Fig. 10).
As shown in figure the graphical representation of cost vs. weight .Where the gradient descent
algorithm is unstable when training the very large data (as per Fig. 12).
Convolutional Neural Network (CNN) is one of the main algorithms which is used to perform image
recognition, object detection and object recognition from a given image. This algorithm is widely used
in object detection and recognition as it has very good accuracy as compared to other neural networks.
Figure 5. Activity Diagram for Digit Recognition Systems
Volume 17 • Issue 4
61
Figure 6. Flow Chart for Digit/Character Recognition System.
Volume 17 • Issue 4
62
Figure 7. Sequence Diagram for Digit Recognition System.
Volume 17 • Issue 4
63
Types of Layers
Let’s understand the working of layers by taking an example of an image of dimension 32 x 32 x 3.
1. Input Layer:This layer stores the raw input of an image with dimension 32 x 32 x1.
2. Convolution Layer: This layer performs the computations to obtain the output volume. Output
volume can be obtained by performing dot product between all filters and image patches.
Figure 8. Use Case Diagram for Digit Recognition System.
Volume 17 • Issue 4
64
3. Activation Function Layer: This layer applies activation function to each and every element
of the output of the second layer. Some most frequently used activation functions are Sigmoid,
Tanh, RELU, Leaky RELU, etc. The volume does not change in this layer.
4.Pool Layer: Main function of this layer is to decrease the size of volume. Decreasing the volume
makes the computation fast and reduces the memory. Pool layer also prevents over fitting.
Two common types of pooling layers are max pooling and average pooling.
5 .Fully-Connected Layer: This layer is a regular neural network layer. It takes input from the
previous layer and outputs the class using an activation function and classifies the input image.
In this, CNN is applied on MNIST dataset to handwritten digits recognition to observe the accuracies.
The accuracy is validation using python training and accuracy (As per Table 1).
Figure 9. The Mathematical Operations Occurring at the First and Second Hidden Layer
Figure 10. The Mathematical Operations Occurring at the First and Second Hidden Layer
Volume 17 • Issue 4
65
This Table shows the maximum and minimum training after using six different cases and this
table also provides the validation accuracies of CNN.
Addition to this, in this table all six cases take place which provide the actual results to the
handwritten digits with the help of MNIST dataset (as per Fig. 12 to Fig. 17).
After applying the MNIST recognition the output comes which contains 2 layers and 10 neurons
and recognition the digits from 0 to 9.
• In this case where pooling 1 and convolution 2 are used .which is connected to each other during
the recognition of digits or the testing .
Figure 11. The Gradient Descent Method for Specific Neurons
Figure 12. The Mathematical Operations Occurring at the Intermediate Hidden Layer
Volume 17 • Issue 4
66
• In this case two convolutions are used by pooling layers, flatten layers and connected layers.
In this case a dropout will be used and the validation of the accuracy of digits will take place.
• the minimum training accuracy is found approx. 98.11% at the epoch 12 and the maximum
training accuracy is found approx. 97.18% at the epoch 14.
Figure 13. This figure show the observed accuracy for case one in which the first hidden layer is convolution layer one which
use 32 filters at the time of recognition which are in the size of 3*3 pixels.
Figure 14. This shows the accuracy of training and validation of the layers as shown.
Volume 17 • Issue 4
67
Figure 18. In this case, the total number of hidden layers which make the combination using
pooling and convolutions layers also show the different output which affect the accuracy
of the digit recognition.
• These six cases show the different output with different layers by which during the recognition
of digits every layer affects the output of the recognition.
• From the above cases the maximum accuracy was found nearly 99.13% in the case 5 and the
minimum accuracy was found approx. 96.55% in the case 8 during the digits recognition.
Figure 15. Observed Accuracy for case 3
Figure 16. Observed Accuracy for case 4
Volume 17 • Issue 4
68
Currently, the authors planned to build a multi-layer neural feed forward network using both
Keras and - an interface for displaying machine learning algorithms in Python.
This network contains perceptions inside layers that absorb the input and transfer the information
to the next part. The team planned on building a multifaceted CNN deposition and layers of maxpool
and final exit layer by activmax activation 3,4.
With the help of CNN, multi layers, MNIST dataset and many other required methods, handwritten
digits recognized which give the output as below(as per Fig. 18, 19 and 20)
Our model will pretend that the input image is unseen and new for the system and then our model
will predict the integer that the image represents.
The load_image () function will return the loaded image which will be ready for classification.
After loading the input image our system will force the image to be in grayscale format and then
force the size to be of 28×28 pixels.
Next, we call the predict classes () function to predict the digit that the image represents.
Output-1
As shown in the above image, 2 is written on the input image. So our model has predicted the integer
that the image represents with an accuracy of 99%.
Figure 17. Observed Accuracy for case 5
Volume 17 • Issue 4
69
Figure 18. Figure shows the output that our model has accurately predicted the correct digit i.e 2 with accuracy of 99%
Figure 19. Figure shows the output that our model has accurately predicted the correct digit i.e 6 with accuracy of 77%.
Volume 17 • Issue 4
70
Output-2
As shown in the above image, 6 is written on the input image. So our model has predicted the integer
that the image represents with an accuracy of 77%.
Output-3
As shown in the above image, 5 is written on the input image. So our model has predicted the integer
that the image represents with an accuracy of 70%.
System’s accuracy depends on its training and how good the handwriting of the sample is and how
easy it is to differentiate each word from each other. Some digits are quite similar to each other like
1 and 7, 6 and 8, 3 and 8, 9 and 0, which is quite confusing for computers and proves to be a high
hurdle sometimes and may even reduce the accuracy of the system.
It may also be possible that the exactness among all perceptions in the exhibition was discovered
to be 97.07% when the number was tested to be 6. In addition, the miss ratio in all our tests was
roughly approx 0.049449. Also, the least testing miss ratio was around 0.21313 when the number
was tested to be 1. The lowest miss ratio will provide CNN with better execution to accomplish good
picture goals & commotion handling. Later on, we will try to watch the variety in the general grouping
precision by fluctuating the quantity of cluster size and concealed layers. We can also improve the
accuracy of our model by trying various methods like by Tuning Pixel Scaling, Tuning the Learning
rate and tuning the Model depth.
The six cases performed different outputs as shown above; all the cases are different to each other
which show different accuracy and different performance. The layers were selected randomly by
which performance of all the cases are different.
Figure 20. Figure shows the output that our model has accurately predicted the correct digit i.e 5 with accuracy of 70%.
Volume 17 • Issue 4
71
The maximum and minimum accuracy was observed with different cases with the help of different
algorithms and different methods with MNIST dataset. Data to knowledge conversion requires much
more computational power than just working on data, increasing system requirements of the project.
KM is different from working on data only, hence, much more complex and requires higher skill.
Although the system achieves almost perfect accuracy, it can’t reach the 100% accuracy mark. It’s
computational cost and testing complexity may be low compared to ensemble architecture, it is still
not trivial. Some digits are quite similar to each other like 1 and 7, 6 and 8, 3 and 8, 9 and 0, which
is quite confusing for computers and proves to be a high hurdle sometimes and may even reduce the
accuracy of the system. KM helps in understanding data and making predictions which helps a lot
but it isn’t still 100% accurate or totally safe for public use like traffic fines. For KM large amounts
of data are required to train the machine and improve accuracy by reducing errors, so although the
machine starts with low accuracy, it starts to chase 100% accuracy as data received increases.
Humans have different handwritings and some are not quite understandable or distinguishable even
for humans, these types of cases prove to be a challenge for the system. System’s accuracy depends
on its training and how good the handwriting of the sample is and how easy it is to differentiate each
word from each other. The system doesn’t have capability to identify its errors and notify users or
correct them.
There are multiple methods that can help us in recognizing digits, for example DNN (Deep Neural
Network), DBF (Deep Belief Network), CNN (Convolution Neural Network), etc.. In this research
paper, we combined CNN with MNIST dataset which provided higher accuracy and a clear image
with high performance. In order to enhance the performance of Digit Recognition System, we
tested various versions of this network in order to avoid costly feature extraction and complex digit
image processing. We can also use data and knowledge for future research. IoT and Big data are big
customers for the knowledge attained by our project. This knowledge can also help in training neural
networks and AIs. Also, with the help of this research work, we left behind the traditional recognition
system which was quite a complex ensemble approach. Still the poor hand writing, cursive letters
and intentionally done bad hand writing is a great challenge in this domain. Sufficient knowledge
extraction to convert into wisdom is difficult for current workers of this area.
The capability of this system in thoroughly investigating different parameters and objectives of
CNN architecture that helps us in providing best accuracy in recognition for the combined dataset.
The different cases were tested with the help of different algorithms and maximum & minimum
accuracy were noted. The maximum accuracy was observed as 99.13% and the minimum accuracy
was observed as 92.42%. This digit recognition system provided higher accuracy and it will definitely
help expediting all large-scale exhibition systems.
Research can be performed in various architectures of Convolutional Neural Network (CNN). Also,
investigation of domain specific recognition can be performed. With the help of different algorithms,
we can optimize the training parameters like number of layers, kernel sizes of filters, learning, etc.
The algorithm proposed in this research paper can be modified further to improve the minimum or
maximum accuracy of recognition.
Volume 17 • Issue 4
72
The algorithm we applied in the recognition of handwritten digits, can also be applied further for
the recognition of handwritten characters. Also, the proposed algorithm should be applied on large
databases in order to determine the efficiency of the system developed.
The Knowledge Pyramid and KM helps in making the use of IOT and Big data easily, opening
new opportunities for future work too. Advancements in related technologies opens many ways to
modify and improve the workings of this project. The data and Knowledge extracted by this project
is extremely precious and flexible for future researchers. With the IOT machines can be trained to not
require human interaction at all like automatic traffic fining through traffic cameras. Accuracy of the
project is yet to achieve flat out accuracy which can be a great topic of research (Jennex, M.E., 2017).
The proposed algorithm can be modified further for the recognition of broken characters or
digits. This kind of recognition system is really important as this can help us in various fields like
forensics, reading postal addresses, forms filled by candidates, bank check amounts. Also, with the
help of machine learning, the system will be able to produce precise and accurate outcomes.
In this research paper we have implemented a CNN algorithm with 5 layers in order to make digit
recognition in handwritten documents. To make it happen we have used MNIST dataset which consist
of more than 60,000 images for training and additional 10,000 images for testing purposes.
While researching, we found out that this CNN approach is much better than “Random Forest
classifier”, “K nearest neighbors” and “supervised vector machine” with utmost 1.28 error rate.
Also, it gave better results because of use of multiple layers and keras which helped in the reduction
of error in classification.
This system is very helpful for solving real time problems such as the problem of faulty written
cheques in banking documents, Forensic etc. Also, by employing the same strategy with alphabets
we may use it as a full-fledged similarity index checker of handwritten documents along with their
digital conversion.
This technique is also helpful for visually impaired people, as that digital document can be
further translated to an audio file.
Knowledge attained through this project not only helps in automating other projects it also helps
with Big data and IOT. AI and Neural Network will be specially helped by this. Compatibility of the
project with knowledge based technologies is really high due to its nature of data based project. It
has really good data for big data operations.
KM is different from working on data only, hence, much more complex and requires higher skill
but as we are moving towards more advanced times and researching better and advanced technologies,
this will be really impactful on society.
While working on this we found out that our technique still has some weak areas, for example
sometimes a fluent stroke of digit ‘3’ may appear similar to digit ‘8’ and same with the ‘6’ and ‘0’,
‘7’ and ‘1’. This could be a good replacement of traditional OCR technique. Although we received
exceptionally good results but still we have approx. 1.28 error rate.
Volume 17 • Issue 4
73
Ahlawat, S., Choudhary, A., Nayyar, A., Singh, S., & Yoon, B. (2020). Improved Handwritten Digit Recognition
Using Convolutional Neural Networks(CNN). Sensors (Basel), 20(3344), 1–18. doi:10.3390/s20123344
PMID:32545702
Dorosh, N., & Fenenko, T. (2020, May 19). Recognition of MNIST Handwritten Digits And Character Set
Research. National Metallurgical Academy of Ukraine. Retrieved from https://www.citethisforme.com/cite/
sources/journalautociteeval
Gulli, A., & Pal, S. (2020). Deep Learning With Keras (2nd ed.). Digit Recognition System.
Hamad, K., & Kaya, M. (2016). A Detailed Analysis of Optical Character Recognition Technology. International
Journal of Applied Mathematics, Electronics and Computers, 4(Special Issue), 244-249.
Islam, M., Hossain, M., Islam, R., & Andersson, K. (2019). Static Hand Gesture Recognition Using Convolutional
Neural Network with Data Augmentation. Joint 8th International Conference on Informatics, Electronics
&Vision(ICIEV) and 3rd International Conference on Imaging, Vision & Pattern Recognition (icIVPR).
doi:10.1109/ICIEV.2019.8858563
Jennex, M. E. (2017). Big Data, the Internet of Things and the Revised Knowledge Pyramid. The Data Base for
Advances in Information Systems, 48(4), 69–79. doi:10.1145/3158421.3158427
Pal, U., Jayadevan, R., & Sharma, N. (2012, March). Handwriting Recognition In Indian Regional Scripts: A
Survey Of Offline Techniques. ACM Transactions on Asian Language Information Processing, 11(1), 1–35.
doi:10.1145/2090176.2090177
Paul, A., & Bhattacharya, N. (2012). Digit recognition from pressure sensor data using Euler number and central
moments. International Conference on Communications, Devices and Intelligent Systems(CODIS). , 93-96.
doi:10.1109/CODIS.2012.6422144
Sahu, N., & Sonkusare, M. (2017). A Study On Optical Character Recognition Techniques. The International
Journal of Computational Science, Information Technology and Control Engineering, 4(1), 1–14.
Sharma, I. (2020, June 14). Handwritten Digits Recognition Using Google Tensorflow With Python”, Data
aspirant. Retrieved from https://dataaspirant.com/handwritten-digits-recognition-tensorflow-python/
Volume 17 • Issue 4
74
As it is already discussed, each image is of 28X28 dimension which gives 784 entries corresponding
to image and pixel intensity.
In the dataset, the first column, called “label”, is the digit that was drawn by the user. The rest of the
columns contain the pixel-values of the associated image.
Tabular Data set, Sample Data Set is given as per Image 22.
1. Test Dataset:
As we have already discussed, each image is of 28X28 dimension which gives 784 entries
corresponding to image and pixel intensity. In this partial screenshot, some of the entries are visible
in which, Vertical entries depict pixel intensities and Horizontal entries depict different images (As
per Figure 21).
2. Training Dataset:
Sample of this data set is given as per Fig. 21
As we have already discussed, each image is of 28X28 dimension which gives 784 entries
corresponding to image and pixel intensity. In this partial screenshot, some of the entries are visible
in which, Vertical entries depict pixel intensities and Horizontal entries depict different images (As
per Figure 22).
Figure 21. MNIST dataset for testing
Volume 17 • Issue 4
75
Rohit Rastogi received his B.E. degree in Computer Science and Engineering from C.C.S. Univ. Meerut in 2003,
the M.E. degree in Computer Science from NITTTR-Chandigarh (National Institute of Technical Teachers Training
and Research-affiliated to MHRD, Govt. of India), Punjab Univ. Chandigarh in 2010. Currently he is pursuing his
Ph.D. In computer science from Dayalbagh Educational Institute, Agra under renowned professor of Electrical
Engineering Dr. D.K. Chaturvedi in area of spiritual consciousness. Dr. Santosh Satya of IIT-Delhi and dr. Navneet
Arora of IIT-Roorkee have happily consented him to co supervise. He is also working presently with Dr. Piyush
Trivedi of DSVV Hardwar, India in center of Scientific spirituality. He is a Associate Professor of CSE Dept. in
ABES Engineering. College, Ghaziabad (U.P.-India), affiliated to Dr. A.P. J. Abdul Kalam Technical Univ. Lucknow
(earlier Uttar Pradesh Tech. University). Also, he is preparing some interesting algorithms on Swarm Intelligence
approaches like PSO, ACO and BCO etc.Rohit Rastogi is involved actively with Vichaar Krnati Abhiyaan and
strongly believe that transformation starts within self.
Himanshu Upadhyay is a B.Tech. in Computer Science Engineering student of Dr. Abdul Kalam Technical University.
He is currently working on “Smart Attendance Monitoring and Marking System” and “IOT based Agricultural
Monitoring System”. His area of interest includes Back End Development, Java, Big Data Analysis, Python, DBMS
etc. His hobbies include listening to music, Swimming, Watching vlogs, Cricket, etc.
Akshit Rajan Rastogi is an engineering student in AKTU University, presently he is in B.Tech 3rd Year CSE Branch
in ABES Engineering College, Ghaziabad. He is currently working on various research papers as an analyst and
content developer. He loves doing research and finding new things related to the Computer Science & Information
Technology field.
Divya Sharma is an engineering student at AKTU University, Lucknow. Presently she is in 3rd year B.Tech CSE
Branch in ABES Engineering College. She is currently working on various research papers as a content writer.
She loves to explore new things in the technical field. She has a keen interest in Machine learning. She wants to
serve the society in future with all her technical resources.
Prankur Bishnoi has achieved B.Tech. in CSE from AKTU University, Lucknow, India. Presently he is working on
various research papers as an analyst and content developer. He loves doing research and finding new things
related to the Computer Science & Information Technology field.
Ankit Kumar is an engineering student in AKTU University, presently he is in B.Tech graduate from ABES Engineering
College, Ghaziabad. He is currently working on various research papers as an analyst and content developer.
He loves doing research and finding new things related to the Computer Science & Information Technology field.
Abhinav Tyagi is a Computer Science Engineering student of Dr. Abdul Kalam Technical University. Currently,
he is in the final year of his graduation. He is currently working on “Smart Attendance Monitoring and Marking
System” and “IOT based Agricultural Monitoring System”. His area of interest includes Front End Development,
Java, python, Machine Learning, Javascript (ES6), Python, DBMS etc. His hobbies include listening to music,
Swimming, Watching vlogs etc. He likes to explore and work on new ideas.
Figure 22. MNIST dataset for training
Content uploaded by Rohit Rastogi
Author content
All content in this area was uploaded by Rohit Rastogi on Oct 14, 2021
Content may be subject to copyright.
