ISSN 2278-3091

Yellapragada SS Bharadwaj et al., International JournalVolume

of Advanced Trends
9 No.2, in Computer
March -AprilScience
2020 and Engineering, 9(2), March - April 2020, 1335 – 1339
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at http://www.warse.org/IJATCSE/static/pdf/file/ijatcse66922020.pdf
https://doi.org/10.30534/ijatcse/2020/66922020

Effective Handwritten Digit Recognition using Deep Convolution

Neural Network

Yellapragada SS Bharadwaj1, Rajaram P2, Sriram V.P3, Sudhakar S4, Kolla Bhanu Prakash5
1
Dept. of CSE, Koneru Lakshmaiah Education Foundation, India,
saisrinivasabharadwaj@gmail.com
2
Dept. of CSE, AKT Memorial College of Engineering and Technology, Kallakurichi, Tamil Nadu, India,
rajaramnov82@gmail.com
3
Dept. of Management Studies, Acharya Bangalore B School (ABBS), Bengaluru,India
dr.vpsriram@gmail.com
4
Dept. of CSE, Sree Sakthi Engineering College, Coimbatore, Tamil Nadu India,
sudhasengan@gmail.com
5
Dept. of CSE, Koneru Lakshmaiah Education Foundation, India,
drkbp@kluniversity.in

 OCR where instead of taking the whole character's data, HDR

ABSTRACT detects digits. Comparing to OCR, HDR is light and faster. In
fields like medical, banking, student management, and
This paper proposed a simple neural network approach taxation process, HDR possesses great flexibility.[3]
towards handwritten digit recognition using convolution.
With machine learning algorithms like KNN,SVM/SOM, A human brain can easily interpret a sophisticated image and
recognizing digits is considered as one of the unsolvable tasks can extract data from it, but for a computer, image is a
due to its distinctiveness in the style of writing. In this paper, collection of pixels which are nothing but a list of
Convolution Neural Networks are implemented with an combinations of numbers ranging from 0 to 255, blends being
MNIST dataset of 70000 digits with 250 distinct forms of
RGB, B.W., greyscale, etc. Information in an image can be
writings. The proposed method achieved 98.51% accuracy for
extracted from the pixel values. A human eye contains
real-world handwritten digit prediction with less than 0.1 %
light-sensitive cells that can segment the image into partitions
loss on training with 60000 digits while 10000 under
validation. which help to interpret the information like shape, size, and
color which are called features, these features are later sent to
Key words: Convolution neural networks, MNIST dataset, the visual cortex that analyses the characteristics and maps it
TensorFlow, OCR, Segmentation, Cross-Validation to our memory for identification. Thus our brain is capable of
identifying an unseen image.
1. INTRODUCTION
Neural Networks are a branch of machine learning inspired
Advancements in the field of computer vision using deep by the human brain. It adapts layered architecture where the
neural networks attract attention; thus, many A.I. layers represent mathematical functions, and each neutron is
practitioners are moving towards it.[1] One of the influencing a carrier of weights. Layers that lie between input and output
projects that opted for deep learning is OCR (Object layers are called hidden layers, which can be changed based
Character Recognition). OCR is a mechanism that converts on the task. For image classification CNN (Convolution
printed or documented letters into encoded text. Intuitively Neural Networks) performs better, and for speech
OCR tool scans a document to extract the information and recognition, Long short-term memory network delivers
store it in a digital document. There are two major ways of better.[4] Intuitively convolution is a process of applying
implementing OCR, one by recognizing the patterns in the various filters on images to highlight the details in it. Each
characters, and the other one is through segmentation.[1][10] binary classifier called perceptron, many perceptron mapping
Either way, this problem comes under the radar of machine from each neuron to every other neuron of forwarding layers
learning. Handwritten digit recognition (HDR) is a snippet of together forms a neural network.[2] Consider an input
greyscale image of size 28x28, and the input layer consists of
784 neurons, each with a value of grey level corresponding to

1335
Yellapragada SS Bharadwaj et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 1335 – 1339

the pixels, and the weight carried by that neuron is called Mahmoud M. Abu Ghosh [8] compared CNN, DNN, DBN
activation. The last layer, aka the output layer, contains approaches to determine which neural network is useful in the
neurons equal to the number of target classes; in case of the field of computer vision, specifically in image classification
handwritten digit, recognition number of categories are ten like OCR. Their analysis claims that DNN's(Deep Belief
ranging from 0 to 9 with probability values.[3][4] The number Network) accuracy beats other neural networks with 98.08%
of input neurons and output neurons depends on the task, but accuracy but falls behind CNN in terms of execution time.
the hidden layers are arbitrary and are not dependent on the they concluded that any algorithm would only have 1 or 2
task. That's why they are hidden layers, but the propagations percent error rate towards the similarity of digits [13] [14]
between the hidden layers depend on the activation of the
previous layers.[11] [12] Intuitively, the pattern of activations A.K. Jain [9] have presented an approximation based KNN
in the presentation layer causes some specific patterns in the classification algorithm for handprinted digit recognition. He
next layer; thus, the highest activation neuron is the network's performed matching on two character's deforming edges and
choice of class. In this paper, we are implementing dissimilarities. Their proposed work is on patterns in
convolution neural network architecture with relu and low-dimensional space where the scaling is 2000 times lesser
sigmoid activation functions to predict a real-world results 99.25% accuracy.
handwritten digit by training the network with the MNIST
dataset.[10] R.Alhajj [10] has applied a completely new approach called
the agent-oriented approach. In a sentence, they appoint
2. RELATED WORK agents to each character where their only purpose is to
identify hills and valleys, which are nothing but blacks and
Many techniques have been developed to recognize whites. The ability of agents to socialize with each other is
handwritten digits; most of the A.I. practitioners use this to highlighting features compared with any other image
test their model's performance. In the past decades, a classification techniques. Overlapping digits are identified
segmentation-based approach was used to solve this problem based on the cut-points, which is nothing but the intersection
later with the advancements in machine learning a of agent paths. The results of their methods are surprisingly
segmentation less approach was introduced. Even though the higher when compared to many other ANN-based approaches
implementation changes, the issue still remains the same and at about 97% accuracy.
open for anyone to solve.
Bailing Zhang [5] utilizes the ASSOM technique to provide 3. METHODOLOGY
numerical stability that can predict precise digits. Their main
idea was to use the SOM clustering algorithm with In this paper, we used MNIST as a primary dataset to train the
autoencoder neural networks in a nonlinear approach. The model, and it consists of 70,000 handwritten raster images
modularity of SOM helped in extracting several features in a from 250 different sources out of which 60,000 are used for
digit. For each digit, individual ASSOM was constructed and training, and the rest are used for training validation. MNIST
compared with ten several construction-related errors to data is represented in the IDX file format and are look like in
minimize the misclassification. Their model shows promising figure 1. Our proposed method mainly separated into stages,
results, even with small training samples. as shown in Figure 2, pre-processing, Data Encoding, Model
Construction, Training & Validation, Model Evaluation &
Saleh Aly [3] proposed a technique for handwritten numerical Prediction. Since the loading dataset is necessary for any
strings of arbitrary length recognition using SVM and PCA, process, all the steps come after it.
addressing the major challenge in word detection, which is
overlapping characters. Their method uses hybrid PCA called
PCANet for segmentation and SVM for segmentation
classification together called PCA-SVMNet. Their
experiment shows high efficiency in recognizing unknown
handwritten number classification without any segmentation
method applied [15]

Yue Yin; Wei Zhang [1] have concluded that out of all neural
network implementations CNN method is valid for OCR
based image classification systems. They claimed that OCR
had become a preliminary technique in the field of computer
vision; they state that if an image classification model
performs well in OCR, then it can be used for any image
classification systems. Figure 1: Sample MNIST data
1336
Yellapragada SS Bharadwaj et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 1335 – 1339

3.1 Pre-Processing

After loading the data, we separated the data into X and y

where X is the image, and y is the label corresponding to X.
As shown in figure 3, the first layer/input layer for our model
is convolution. Convolution takes each pixel as a neuron, so
we need to reshape the images such that each pixel value is in
its own space, thus converting a 28x28 matrix of greyscale
values into 28x28x1 tensor. With the right dimensions for all
the images, we can split the images into train and test for
further steps [19][20][21].

Figure 2: Flowchart

3.2 Data Encoding Figure 4: Model Summary

This is an optional step since we are using the 3.4 Training & Validation
cross-categorical entropy as loss function; we have to specify
the network that the given labels are categorical in nature. After building the model [24], we compiled a model with
adam optimizer and particular cross-entropy loss function,
3.3 Model Construction which are standard in making a convolution neural network.
After data encoding, the images and labels are ready to be Once the model is successfully assembled, then we can train
fitted into our model [22] [23]. Summary of the model can be the model with training data for 100 iterations, but as the
seen in Figure 4 number of iteration increases, there is a chance for overfitting.
Therefore we limit the training up to 98% accuracy, as we are
using real-world data for prediction, test data was used to
validate the model [16].

3.5 Model Evaluation & Prediction

For real-world image classification prediction, we need to do

a little image pre-processing on the real-world images as
model training was done with greyscale raster images. The
steps of image pre-processing are,
Figure 3: Proposed model 1. Loading image
2. Convert the image to greyscale
Our model is composed of feature extraction with convolution 3. Resize the image to 28x28
and binary classification. Convolution and max-pooling are 4. Converting the image into a matrix form
carried out to extract the features in the image, and a 32 3x3 5. Reshape the matrix into 28x28x1
convolution filters are applied to a 28x28 image followed by a After pre-processing, we predict the label of the image by
max-pooling layer of 2x2 pooling size followed by another passing the pre-processed image through the neural network.
convolution layer with 64 3x3 filters. In the end, we obtain The output we get is a list of 10 activation values 0 to 9,
7x7 images to flatten. Flatten layer will flatten the 7x7 images respectively. The position having the highest value is the
into a series of 128 values that will be mapped to a dense layer predicted label for the image [18].
of 128 neurons that are connected to the categorical output
layer of 10 neurons. 4. RESULTS AND DISCUSSION
Our model is built to work on real-world data, and real-world
images are not even close to MNIST raster images, a lot of
pre-processing was done to make a real image to look like a
raster image.

1337
Yellapragada SS Bharadwaj et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 1335 – 1339

4.1 Accuracy score 5. CONCLUSION

Our model stopped training at the 2nd epoch as it reached The performance of CNN for handwritten recognition
98.21% training accuracy and 98.51% validation accuracy performed significantly. The proposed method obtained 98%
with 5% training loss and 4% validation loss. The progression accuracy and is able to identify real-world images as well; the
of accuracy and loss are represented in Figure 5. loss percentage in both training and evaluation is less than
0.1, which is negligible. The only challenging part is the
noise present in the real-world image, which needs to look
after. The learning rate of the model is much dependent on the
number of dense neurons and the cross-validation measure

REFERENCES
1. Y. Yin, W. Zhang, S. Hong, J. Yang, J. Xiong, and G.
Gui, "Deep Learning-Aided OCR Techniques for
Chinese Uppercase Characters in the Application of
Internet of Things," in IEEE Access, vol. 7, pp.
47043-47049, 2019.
https://doi.org/10.1109/ACCESS.2019.2909401
2. N. B. Venkateswarlu, "New raster, adaptive document
binarization technique," in Electronics Letters, vol. 30,
no. 25, pp. 2114-2115, 8 Dec. 1994.
https://doi.org/10.1049/el:19941439
3. S. Aly and A. Mohamed, "Unknown-Length
Figure 5: Loss and Accuracy Learning Curves Handwritten Numeral String Recognition Using
Cascade of PCA-SVMNet Classifiers," in IEEE
From the above curve, we can observe that the loss and Access, vol. 7, pp. 52024-52034, 2019.
accuracy are cooperatively changed at every fold during 4. Seong-Whan Lee and Sang-Yup Kim, "Integrated
k-fold cross-validation. Before two folds, efficiency almost segmentation and recognition of handwritten
reached 98%, and that's why the number of iterations stopped numerals with cascade neural network," in IEEE
at the 2nd epoch. Stability inaccuracy score can be observed Transactions on Systems, Man, and Cybernetics, Part C
from 2nd iteration. (Applications and Reviews), vol. 29, no. 2, pp. 285-290,
May 1999.
4.2 Prediction https://doi.org/10.1109/5326.760572
5. Bailing Zhang, Minyue Fu, Hong Yan, and M. A. Jabri,
Our model is able to recognize computer-generated digits as "Handwritten digit recognition by adaptive-subspace
well as handwritten digits. Computer-generated digit self-organizing map (ASSOM)," in IEEE Transactions
prediction is more accurate compared to real-world digit on Neural Networks, vol. 10, no. 4, pp. 939-945, July
prediction, which can be observed in Figure 6. 1999.
6. L. Deng, "The MNIST Database of Handwritten Digit
Images for Machine Learning Research [Best of the
Web]," in IEEE Signal Processing Magazine, vol. 29,
no. 6, pp. 141-142, Nov. 2012.
https://doi.org/10.1109/MSP.2012.2211477
7. S.Sudhakar, V.Vijayakumar, C.SathiyaKumar, V.Priya,
LogeshRavi, V.Subramaniyaswamy, Unmanned Aerial
Vehicle (UAV) based Forest Fire Detection and
monitoring for reducing false alarms in forest-fires,
Elsevier- Computer Communications 149 (2020) 1–16,
https://doi.org/10.1016/j.comcom.2019.10.007
8. M. M. A. Ghosh and A. Y. Maghari, "A Comparative
Study on Handwriting Digit Recognition Using
Figure 6: Raster vs. Real image prediction Neural Networks," 2017 International Conference on
Promising Electronic Technologies (ICPET), Deir
El-Balah, 2017, pp. 77-81.
9. A. K. Jain and D. Zongker, "Representation and
recognition of handwritten digits using deformable

1338
Yellapragada SS Bharadwaj et al., International Journal of Advanced Trends in Computer Science and Engineering, 9(2), March - April 2020, 1335 – 1339

templates," in IEEE Transactions on Pattern Analysis (IJATCSE), Vol.8 No.5,pp:2062-2067,

and Machine Intelligence, vol. 19, no. 12, pp. DOI:10.30534/ijatcse/2019/33852019.
1386-1390, Dec. 1997. 19. K.B., Prakash. "Information extraction in current
10. S.Sudhakar, V.Vijayakumar, C.SathiyaKumar, V.Priya, Indian web documents." International Journal of and
LogeshRavi, V.Subramaniyaswamy, Unmanned Aerial Technology(UAE), 2018: 68-71.
Vehicle (UAV) based Forest Fire Detection and https://doi.org/10.14419/ijet.v7i2.8.10332
monitoring for reducing false alarms in forest-fires, 20. Kolla B.P., Dorairangaswamy M.A., Rajaraman A. "A
Elsevier- Computer Communications 149 (2020) 1–16, neuron model for documents containing multilingual
https://doi.org/10.1016/j.comcom.2019.10.007 Indian texts." 2010 International Conference on
11. R. Alhajj and A. Elnagar, "Multiagents to separating Computer and Communication Technology,
handwritten connected digits," in IEEE Transactions ICCCT-2010,2010: 451-454.
on Systems, Man, and Cybernetics-Part A: Systems and https://doi.org/10.1109/ICCCT.2010.5640489
Humans, vol. 35, no. 5, pp. 593-602, Sept. 2005. 21. Kolla B.P., Raman A.R. "Data Engineered Content
https://doi.org/10.1109/TSMCA.2005.843389 Extraction Studies for Indian Web Pages." Advances
12. H. I. Avi-Itzhak, T. A. Diep and H. Garland, "High in Intelligent Systems and Computing, 2019: 505-512.
accuracy optical character recognition using neural https://doi.org/10.1007/978-981-10-8055-5_45
networks with centroid dithering," in IEEE 22. Prakash K.B., Dorai Rangaswamy M.A. "Content
Transactions on Pattern Analysis and Machine extraction studies using neural network and
Intelligence, vol. 17, no. 2, pp. 218-224, Feb. 1995. attributegeneration." Indian Journal of Science and
https://doi.org/10.1109/34.368165 Technology, 2016: 1-10.
13. Naimin Zhang, Wei Wu, and Gaofeng Zheng, 23. Prakash K.B., Dorai Rangaswamy M.A., Raman A.R.
"Convergence of gradient method with momentum "Text studies towards multi-lingual content mining
for two-Layer feedforward neural networks," in IEEE for web communication." Proceedings of the 2nd
Transactions on Neural Networks, vol. 17, no. 2, pp. International Conference on Trendz in Information
522-525, March 2006. Sciences and Computing, TISC-2010, 2010: 28-31.
14. R.Vasanthi, R.Jayavadivel, K.Prasadh, J.Vellingiri, G.A https://doi.org/10.1109/TISC.2010.5714601
kilarasu, S.Sudhakar, P.M.Balasubramaniam, A novel 24. Prakash K.B., Rangaswamy M.A.D. "Content
user interaction middleware component system for extraction of biological datasets using soft computing
ubiquitous soft computing environment by using the techniques.”, Journal of Medical Imaging and Health
fuzzy agent computing system, Journal of Ambient Informatics, 2016: 932-936.
Intelligence and Humanized Computing (2020), https://doi.org/10.1166/jmihi.2016.1931
Springer, doi.org/10.1007/s12652-020-01893-4.
15. Jagadeesh Gopal, Vellingiri J, Gitanjali J, Arivuselvan
K, Sudhakar S, An Improved Trusted On-Demand
Multicast Routing with QoS for Wireless Networks,
International Journal of Advanced Trends in Computer
Science and Engineering, Vol.9,No.1, January –
February 2020, P.P.:261-265,
https://doi.org/10.30534/ijatcse/2020/39912020.
16. Satheesh N, Sudha D, Suganthi D, Sudhakar S, Dhanaraj
S, Sriram V.P., Priya V, "Certain improvements to
Location aided packet marking and DDoS attacks in
internet," Journal of Engineering Science and
Technology Vol. 15, No. 1 (2020), pp: 94 - 107, School of
Engineering, Taylor's University.
17. Sathiya Kumar C, Priya V, Sriram V P, Sankar Ganesh
K, Murugan G, Devi Mani, Sudhakar S, "An Efficient
Algorithm for Quantum Key distribution with Secure
Communication, Journal of Engineering Science and
Technology Vol. 15, No. 1 (2020), pp:77-93, School of
Engineering, Taylor's University.
18. S.Sudhakar, N.Satheesh, S.Balu, Amireddy Srinish
Reddy, G.Murugan,2019,"Optimizing Joins in a
Map-Reduce for Data Storage and Retrieval
Performance Analysis of Query Processing in HDFS
for Big Data," International Journal of Advanced
Trends in Computer Science and Engineering,
1339