International Conference on Communication and Signal Processing, April 4-6, 2019, India

Enhanced Character Recognition using Deep

Neural Network- A Survey
Desai Hinduja, R. Dheebhika and T. Prem Jacob

Abstract—The idea has been popped up suddenly 50 years ago. The rest of this paper Section II is discussed about the
Neural Computing is a separate science and is very extensive. Literature review ,Methodology is briefly explained in
Sample recognition problems are solved by using Neural Section III ,Result are explained in Section IV and finally
Networks. Character Recognition is one among those. The Concludes the paper in Section V.
implementation of the Neural Network is easy for obtaining the
solution for this problem. Artificial Intelligence uses techniques
that gives Computer the capability to learn with data and
II. LITERATURE REVIEW
without being explicitly programmed. The neural network
recognizes characters, numbers and some special symbols. The
In the optical character recognition system, many
solution is obtained by using MATLAB’s Neural Network techniques are determined to realize the centre of character
Toolbox. The accuracy of the output depends on the resolution of recognition. Though, there are many papers and studies
the input to the neural network. which describes the methods for conversion of text from the
document into readable form by the machine. In the process
Index Terms—Neural Network, BPN Network, OCR, CNN,
Artificial Intelligence. of Optical character recognition, the computer automatically
understands the image of script which is handwritten and
I. INTRODUCTION convert into the classify character [1].

M ANY improvements are occurring in the area of Deep

Learning. OCR is one of the active areas in where
Deep Neural Network is used. Recognition of handwriting is
The ability to reduce staff involvement to increase the
productivity and to store text effectively are the main
not much difficult process for humans. It is a complicated advantage in this Optical Character Recognition System. In
process in case of computers. The reason is that handwriting this paper, the data is taken from the faded or poorly printed
varies between person to person and there are various documents which are blurred. Using 53,342 noisy character
characters. images inception V3 network is trained. The result shows
OCR is the fascinating area where the pattern recognition that the error rate is reduced overall to 21.5% and achieved
and image processing is used. The real time application of good recognition accuracy on text images which is poor in
OCR includes Automatic number plate recognition, quality compared to existing OCRs [2].
transforming the handwritten document into the structural text This paper proposes a simple color and size invariable
form, etc. The aim of OCR is to identify the digits, characters character recognition system to recognize alphanumeric
and special symbols. There are various steps which are to be characters based on ANN. Using feed-forward neural
carried out. They are Pre-processing, Scale region Detection, network this system is proposed. It has two layers. Input
Segmentation, Classification. In this process, we can use layer and output layer. The recognition process contains the
MATLAB toolbox which help us to identify the parameters. following steps. They are pre-processing, normalization,
OCR is an electronic or mechanical conversion of typed network establishment and recognition. This system gives
images or, printed text or handwritten into machine encoded better result for when the letters and numeric digits are
text. It is used as data entry from the data records of the trained and tested separately. When they are tested together
printed paper, whether passport documents, printouts of static- it gives satisfactory result [3-8]. Neural Networks are being
data, bank statements, computerized receipts, documentation used in different kinds of pattern recognition. Handwriting
or mail. varies from person to person. Therefore, it is tedious to
recognize handwritten characters. In handwritten character
recognition, Neural network plays an important role. In this
paper, for English character recognition the developmental
outcome shows that the way used gives better recognition
Desai Hinduja ,R. Dheebika and T.Prem Jacob are with Satyabama accuracy and reduced training time [9].
Institute of Science and Technology, Chennai. In the normal optical character recognition systems, an
{email:[email protected], [email protected] }
optical scanner digitizes input characters. For reducing the

978-1-5386-7595-3/19/$31.00 ©2019 IEEE

0438
noise and normalizing the image, every character is placed and ranging to 94% is acquired after training the neural network
divided into segments and the outcome of the image which is a model.
character is processed inside the pre-processor. To classify,
C. Back Propagation Neural Network (BPN)
certain characteristics are extracted from the character. Many
different techniques exist as feature extraction is critical and With mathematical foundation, the back propagation
each has its own potency and weakness. The characters that neural network process is used to train a multi-layer ANN.
are identified during classification are collected to rebuild the The intention of Back Propagation Neural network (BPN)
initial symbol strings and to detect and correct errors context is training the neural network. The patterns which are
may then be applied. taken as input is balanced between them. Using this BPN,
the method is classified into 2 parts. The training section
This paper proposes that using handheld devices and the recognition section. The work flow Shows in Fig.
documents which contains camera captured images containing 2.
texts can be recognized using Optical Character Recognition D. Deep Neural Network (DNN)
(OCR) system. Initially, the text regions of the image are
extracted and skew corrected. After binarizing these regions, DNN is a neural network with more than two layers.
they segmented into lines and characters [10]. In the These layers connected with other nodes called
recognition module the characters are passed. The maximum interconnected nodes. These layers communicate with each
recognition accuracy of 92.74% is obtained when captured other using the hidden layers. The actual processing is
using cell phone camera. This technique consumes low carried out through the weighted connections. These hidden
memory and computationally efficient and hence applicable layers are connected to the output layer.
on handheld devices [11].

III. METHODOLOGY Hidden layer

A. Feed Forward Backpropagation connections

The feature of backpropagation algorithm is the acquired
error. The error that is obtained by the neural networks get
along its output. The difference between output value and the Input output
real value is equal to the obtained error. In the output, the layer layer
backpropagation algorithm reduces the error. The layers of
Neural network is shown in Fig. 1.

Backpropagation algorithm functions using some small

Fig. 1. Layers of Neural Network
repeated steps. Using the pre-initialized weights, an output
value is generated. Mean square error value will be calculated
using the output values. The required output value is
INPUT IMAGE
compared with the obtained output value. This process will be
followed for all the samples and minimizes the error. This
method doesn’t process precisely but it learns to solve
satisfactorily.
PRE-PROCESSING
B. Convolutional Neural Network (CNN)

Neural Network is a set of algorithms which are

designed to recognize the patterns. CNN has many
applications in the field of natural language processing, SCALE REGION
DETECTION
image recognition, video recognition and recommender
system. It is very good at image recognition and biologically
inspired neural network. In this network, the much essential
layer is convolutional layer. It is used for performing SEGMENTATION
convolutional operation. Next layer is pooling layer. This
layer is useful for larger images. The trainable parameters of
this layer are large. This layer is useful for reducing size of
the image. The original size of the images is 128x128 CLASSIFICATION
pixels. The image size is reduced to 28x28 to decrease the
time taken for training neural network. The accuracy Fig. 2. Flow chart of Work Flow

0439
 IV. RESULT [4] Swati Tomar and Amit Kishore “A Review: Optical Character
Recognition” International Journal of Engineering Sciences &
By plotting into graph, the image is divided as lines and the
Research Technology, ISSN: 2277-9655, CODEN: IJESS7, Tomar
sum of pixel is calculated. From the histograms final text can et al., 7(4): April, 2018.
be predicted. [5] P. Kakkar and U. Dutta, “A Novel approach to Recognition of
V. CONCLUSION English Characters using Artificial Neural Networks”,
International Journal of Advanced Research in Electrical,
In this paper, we studied various algorithms which has been Electronics and Instrumentation Engineering, Vol 3, Issue 6, June
2014.
already implemented. Using these algorithms, recognition of
[6] K Prasad, D. C. Nigam, A. Lakhotiya and D. Umre, “Character
characters has been done but still having some chances for the Recognition using MATLAB’s Neural Network Toolbox”,
improvements in the existing accuracy and efficiency. In International Journal of u- and e- Service, Science and
order to obtain an accurate result, successful method should be Technology, Vol. 6, No. 1, February, 2013.
[7] M. F. Kader and K. Deb, “Neural network based English
used and this can be done with the help of Deep Neural
alphanumeric character recognition”, International Journal of
Network. In future better algorithms can be used which would Computer Science, Engineering and Applications (IJCSEA) Vol.2,
yield better results. No.4, August 2012.
[8] A. F. Mollah, N. Majumder, S. Basu and M. Nasipuri, “Design of
REFERENCES an Optical Character Recognition System for Camera based
Handheld Devices”, IJCSI International Journal of Computer
[1] Jyoti Dalal and Sumiran Daiya “Image Processing Based Optical Science Issues, Vol. 8, Issue 4, No. 1, July 2011.
Character Recognition Using MATLAB” International Journal of [9] A. Pal and D. Singh, “Handwritten English Character Recognition
Engineering Sciences & Research Technology, ISSN: 2277-9655,
using Neural Network”, International Journal of Computer Science
CODEN: IJESS7, Dalal et al., 7(5): May, 2018
& Communication, Vol. 1, No. 2, July-December 2010, pp. 141-
[2] Tan Chiang Wei, U. U. Sheikh, Ab Al-Hadi Ab Rahman. “Improved
144
Optical Character Recognition with Deep Neural Network”, 2018
[10] Prem Jacob, T; Pravin, A; Asha, P. Arduino Object Follower with
IEEE 14th International Colloquium on Signal Processing & its
Augmented Reality. International Journal of Engineering &
Applications (CSPA 2018), 9 -10 March 2018 Technology, [S.l.], v. 7, n. 3.27, p. 108-110, aug.2018. ISSN 2227-
[3] Shyla Afroge, Boshir Ahmed, Firoz Mahmud. “Optical character 524X, doi:http://dx.doi.org/10.14419/ijet.v7i3.27.17665.
recognition using back propagation neural network", 2016 2nd [11] Dr. T. Prem Jacob, Implementation of Randomized Test Pattern
International Conference on Electrical, Computer & Generation Strategy, Journal of Theoretical and Applied
Telecommunication Engineering (ICECTE), 2016. Information Technology 10th March 2015. Vol.73 No.1.

0440