See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/331617158

Recognition and Solution for Handwritten Equation Using Convolutional

Neural Network

Article · June 2018

DOI: 10.1109/ICIEV.2018.8640991

CITATIONS READS

4 2,758

5 authors, including:

Md. Bipul Hossain Y. A. Joarder

Noakhali Science & Technology University World University of Bangladesh
8 PUBLICATIONS   10 CITATIONS    11 PUBLICATIONS   5 CITATIONS   

SEE PROFILE SEE PROFILE

Md Zahidul Islam Md Jashim Uddin

Islamic University (Bangladesh) Islamic University (Bangladesh)
23 PUBLICATIONS   100 CITATIONS    13 PUBLICATIONS   5 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Cryptography View project

Point Detection Based A Smart Pen Featured Listening and Writing Capable Robotic Arm View project

All content following this page was uploaded by Y. A. Joarder on 22 September 2019.

The user has requested enhancement of the downloaded file.

[Format: Times New Roman, 10 font, bold] - 2018 Joint 7th International Conference on Informatics, Electronics
& Vision (ICIEV) and 2018 2nd International Conference on Imaging, Vision & Pattern Recognition (icIVPR)

Recognition and Solution for Handwritten Equation

Using Convolutional Neural Network
Md Bipul Hossain 1, Feroza Naznin 2, Y.A. Joarder 3, Md Zahidul Islam 4, Md Jashim Uddin 5
Information & Communication Engineering
Islamic University,Kushtia,Bangladesh
1
bipulhossainice@gmail.com, 2 liaiceiu@gmail.com, 3 yajoarder@gmail.com,
4
zahidimage@gmail.com, 5 jashim.iu@gmail.com

Abstract— Robust handwritten character recognition is a tricky
job in the area of image processing. Among all the problem
handwritten mathematical expression recognition is one of the
complicated issue in the area of computer vision research.
Segmentation and classification of specific character makes the
task more difficult. In this paper a group of handwritten
quadratic equation as well as a single quadratic equation are
considered to recognize and make a solution for those equation.
Horizontal compact projection analysis and combined connected
component analysis methods are used for segmentation. For
classification of specific character we apply Convolutional
Neural Network. Each of the correct detection, character string
operation is used for the solution of the equation. Finally the
experimental results shows the great effectiveness of our
proposed system.
Keywords-Character segmentation; Convolutional neural Network;
Projection analysis; Quadratics;Connected component .
I. INTRODUCTION
Due to the swift improvement in computer technology and
internet technology, most of the documents, books and
literatures in the area of computer science as well as others are
increasingly becoming digitalized. Mathematics is broadly
used in almost all areas of science, such as physics,
engineering, medicine, economics, etc[6]. Digital document
analysis and understanding is the major research concern
today. For the recognition of English characters and numbers
in electronic books OCR (optical character recognition) can
attain higher recognition exactness[7]. Handwritten
mathematical expression recognition is still a most challenging
job to do in the area of computer vision[3]. Due to the two
dimensional nesting assembly and different sizes, the
correction rate of symbol segmentation and recognition still
cannot achieved its actual requirements. The primary task for
the recognition of mathematical expression is to segment the
character and then classify those character. An operation that
seeks to decompose an image of a sequence of characters into
sub images of individual symbols is Character segmentation
[19][1]. Histogram grounded projection analysis is a well-
known method and is commonly used in the segmentation of
line and character from the image and also we can practice it
at some detection stage[20][2]. Convolutional neural network
(CNN) is one of the mostly used classification model in
computer vision area[24]. In the last few years, deep
Convolutional Neural Network(CNN) leaning has proved
the outstanding performance in the field of image
classification, machine learning and pattern recognition.
Above all existing model, CNN is one of the most popular
models and has been providing the state-of-the-art
recognition accuracy on object recognition [10], segmentation
[11], human activity analysis [12],image super resolution [13],
object detection [14], scene understanding [15], tracking [16],
and image captioning [17]. For the task of image classification
CNN is outperforms above all the previous classification
method [10]. CNN extract feature from the image by a series
of operations. In this research we mainly focus on the
recognition of quadratic equation and after successful
detection of quadratics we apply character string operation
for the solution of those equation. The main contribution of
recognition of the mathematical equation in this paper are
summarized here. In this research we consider a single or a
series of handwritten quadratic equation in the form
ay2+by+c=0 with any combination of it’s part. Our job is to
recognize the handwritten quadratics form the image and for
each successful detection finding the solution of that equation
. In the case of segmenting each quadratics from the original
image the square term is sometimes slightly placed out of the
other parts of the equations horizontal projection curve. We
solve this problem by compact horizontal projection. Another
difficulty exists for applying connected component in the case
of mathematical symbol such as for ‘=’, which is a
combination of two connected component. We solve this
problem by using combined connected component analysis
method. Feature extraction of each individual character is
most difficult part for any handwritten due to it’s different
shape and structure. To solve this problem we apply
convolutional neural network which doesn’t required any
predefined feature for classification of specific character. We
use smallest number of hidden layer in order to reduce the
training time with tolerable error rate. In section II literature
review of the related work is described and in section three
represents the proposed method of this work. The following
section describe the experimental results and finally describe
the summary of this work.
II. LITERATURE REVIEW
Several work has been done so far for handwritten character
recognition[25][27]. A scheme for mathematical expression
recognition (MER) based on SVM and projection histogram
for simple ME’s have proposed in [18]. which is a part of

[Format: Times New Roman, 10 font, bold] - 978-1-5386-5163-6/18/$31.00 ©2018 IEEE Best Regards, ICIEV -
IVPR 2018
offline handwritten expression recognition. This paper
basically focus on numerous technique used for feature
extraction and recognition. An effective and robust system
for recognition of printed and handwritten mathematical
characters have been proposed by Zanibbi et. al. (2002).
Which evaluates an expression by using three successive
passes i.e., the layout pass, lexical pass and operator tree.
This tree manipulation which used in each pass can be
represented by tree transformation. Bage et.al.(2013) has
proposed recognition system for offline handwritten
mathematical symbol. For feature extraction, shape of the
character is considered. The proposed system is based on
relative study of feature extraction methods. Recognition
has been carried out via SVM. Ramteke et.al. (2012) have
transformed scientific and engineering documents into
electronic form by bearing in mind of printed mathematical
symbols and expressions. Features are computed using Figure 1. Workflow diagram of proposed method
centroid of the image here and for segmentation bounding For each correct detection we finally find the solution of the
box technique is used. Classification has been carried out quadratics. In Figure 1 workflow diagram of our proposed
using neural network approach which recognition rate of method is shown.
90%. Ahmad-Montaser Awal et al(2010) discussed some
issues related to the problem of online mathematical A. Dataset preparation
expression recognition. Ha. et. al. (1995, August) have
Preparation of the dataset is the fundamental concern for this
proposed a system that understand mathematical expression
work. Characters such as English digit, alphabet and
from the printed document images. In the development of
mathematical symbol all can be well defined by their edges.
this system, they have adapted object oriented methodology
For this reason we first prepare dataset as the most precedence
to describe the data abstraction for the hierarchical structure
given to its edges that is enlighten the edges. We prepare some
of the mathematical expression is given in the form of the
dataset by ourself and we also used some modified version of
expression tree. Pradeep et. al. (2010) have proposed
NIST dataset which is similar as the popular MINIST dataset
diagonal feature extraction technique for the handwritten
for digit. For each category we use 2000 data item for the
character by using feed-forward neural network algorithm.
training of the network. And most of the case our network
This technique uses diagonal, horizontal and vertical features
training we gained more than 98.5% training accuracy. Image
for the classification purpose. A few paper are available
size we used in dataset is 32×32 gray level image.
online for recognition of mathematical expression using
convolutional neural network(CNN)[28][3].In recent years, B. Pre-Processing
Convolutional Neural Network (CNN) has made a series of Pre-processing of the input image is the procedure which
revolution research results in the fields of image classification encompasses changes and modifications to the image to make
and detection[10]. The past few years a branch of it fit for recognition. The following techniques may be used
artificial neural networks called deep learning has shown for image enhancement.
great potential in solving classification problems. The field of
deep learning started gaining popularity in 2012 when Alex et 1) Conversion of RGB to Gray-Scale
al. demonstrated their deep network architecture named First of all this coloured image is transformed into a typical
AlexNet for image classification which outperformed[5]. gray-scale image and is represented through a single matrix
III. PROPOSED METHOD because the detection of characters on a coloured image is
more challenging than on a gray-scale image. IF the gray
In our proposed method at first noise from the original input bitmap Y and color bitmap is R,G and B then the formula we
image is removed by applying binarization to it. After that we used is:
use compact horizontal projection for segmenting each line of Y = 0.299R + 0.587G + 0.114B (1)
equation from the input image. Then we consider each part of
the segmented image as a full image for further process. For
each line of equation image we then find specific characters in 2) Binarization
the form of connected component. Each segmented character Binarization is the procedure of choosing a threshold value is
is then providing as input to the convolutional neural network for adaptation of pixel values into 0’s and 1’s. In This research
model for classification of the character. The resulting for horizontal projection calculation 1’s represent the black
character that is the output of CNN is then used for making a pixels while 0’s characterize the white pixels. The threshold
character string which is similar to the original equation. choice of binarization can approve two ways: overall threshold
and partial threshold. Otsu’s method is an overall threshold
method, basing on image statistical characteristics. This
method can make the computer select automatically
threshold[7].
3) Noise Reduction
Excessive pixels those are present in an image is called noise.
Noise may be in the in the form of Salt and Pepper noise or
Gaussian noise. Low pass filtering is used to remove the
Gaussian noise from the image [22] and there is no need to
filter Salt and Pepper noise as it is very low as compared to the
Gaussian noise. In our proposed method we removed all
component which is less than 5 pixel for simplicity of small
unwanted pixel noise.
C. Segmentation
Identifying the objects or other significant information in
digital images segmentation is mostly used in image
processing and computer vision application. Which is the
process of dividing an image into multiple parts[19]. The
segmentation involves two main steps in our proposed
method.
a) Equation line Segmentation
Equation segmentation is the split-up of the different lines of
characters existing in the image. Each line is well-defined by a
minimum vertical gap between the characters existing on a
line and on the line overhead and below it. This gap can be
used for the detection and separation of not the same lines of
characters. For the solution of square term in quadratics as
mentioned above we use compact horizontal projection.
Combining of two curve into one that is representing one
equation line with two curve is done by comparing with a
threshold value that is obtained by averaging the maximum
value of each horizontal projection vector. If any curve’s
maximum value is less than the threshold this curve is then
considered as a part of the next projection curve.
Algorithm 1: Compact horizontal projection
Step 1. Calculate N horizontal projection vectors, X.
Step 2. Find the maximum of each vector.
Step 3. Calculate Threshold,
Th=
Step 4. For each curve if max(X)< Th ,
Combine the curve with next one for segmentation
b) Character Segmentation
Character segmentation is a procedure that look for to
decompose an image of a series of characters into sub images
of individual symbols[19]. In our proposed method we use
connected component analysis method for the segmentation of
specific character from the image. Problem arises at the point
of extracting of math symbol ‘=’ which is a combination of
two connected component. If two consecutive component has
the minimum height and are in a same horizontal direction we
can consider this two into a single component.If an outside
rectangle of two symbols region with connected regions is set
to X(x1, y1, x2, y2) and Y(x3, y3, x4, y4). Now if this two
symbol X and Y has minimum height and belong to a whole
horizontal direction then the combination of X and Y produce
a third component Z that is our desired. The position of Z can
be obtained as
Z.x5=min(x1,x3)
Z.x6=max(x2,x4)
Z.y5=max (y2, y4)
Z.y6 =min(y1 ,y3)
Where the even x and y variable represent the right and top
position of the rectangle respectively, odd x and y variable
represent respectively the left and bottom position of the
rectangle.
D. Classification model
In this research convolutional neural network is used as a
classification model , 32×32 gray scale image is used as input
of the CNN input layer for the training section and also for the
testing section. 7×7 filter is used at the convolutional layer
after the convolution of the input image with the filter for each
input image a 28×28 feature vector is produced. It is concord
to apply a nonlinear layer (or activation layer) instantly after
conv layer,. The persistence of this layer is to introduce
nonlinearity to a system[9]. To performs down-sampling by
separating the input into rectangular pooling regions, and
figuring the maximum of each region a max pooling layer is
used[5]. Pooling layers perform down-sampling operations.
Calculation of the output size, O of a pooling layer with input
size ,I Pooling filter size,F Padding ,P and stide ,S has done
by
O = (I – F + 2×P) / S + 1 (2)
In our proposed method we use 2×2 pool size and the output
of the pooling layer is 14×14. Random selection of input
elements to zero with a given probability is done in a dropout
layer. It is a simple way to prevent the overfitting in the
neural network. Overfitting is a stern difficulty in such
networks. Large networks are also sluggish to use, making it
more challenging to contract with overfitting by combining
the predictions of many different large neural nets at the test
time.The technique for addressing this problem is Dropout .
The core idea is to randomly drop units (along with their
connections) from the neural network during training [23]. In
our proposed method we use the probability 0.3 at the dropout
layer during the training. After the convolutional and down-
sampling layers one or more fully connected layers is used.
The layer in which the neurons associate to all the neurons in
the previous layer is fully connected layer. Fully connected
layer combines all the features learned by the previous layers
through the image to recognize the larger patterns. To classify
the images the last fully connected layer combines the
features. Hence, the Output parameter in the last fully
connected layer is equal to the number of classes in the target
data. In our work, the output size is 14, corresponding to the
14 classes. An activation function softmax which normalizes
the output of the fully connected layer is also used in the
convolutional neural network model. The softmax layer output
comprises of positive numbers that sum to one, in the nest that
can be used as the classification probabilities at the
classification layer. At the final layer ,classification layer uses
the probabilities given by the softmax activation function to
find the input image classes and find the loss by comparing it
with the pre assigned ground truth classes.
problem by considering two consecutive component with
smallest width and also considering two component with same
horizontal direction as shown in figure 3 (d). Most complex
part is the classification part generally several character are
distinct in shape. But for the different handwritten strategy
some the character creates ambiguity to each other categories
for example ‘y’ and ‘2’ are two separate character with
different structure. But when ‘2’ is used as a square term it
shape is then something like as the figure 3 which is
sometimes similar to the structure of the ‘y’.

E. Solution of Quadratics
After successful detection of the quadratics we organized the
character string in the form of ‘+ay^2+by+c=0’. Now for
solving this quadratics our job is to find the value of a, b, c.
where a,b,c can be a single digit number or a multiple digit a)
number. To find the value of a, b, c we apply a character string b)
operation algorithm as Algorithm 2. After finding the
parameter of the quadratics we apply the formula for the
solution of the quadratic equation.

Algorithm 2 Quadratics parameter calculation

Step 1. Find the position i of y^2, y and ‘=’ in the c)

character string s. Initialize sum=0.
Step 2. [For each of a, b, c ]From j= i-1 to till s(j)!= ‘+’ or
s(j)!= ‘-’decrease j by 1. sum =10*[s(j)-48]+sum.
Step 3. If s(j)== ‘-’ then sum= - sum
d)
Figure 2. Horizontal projection curve (b) of the original image (a) due to
image pixel orientation, Compact horizontal projection curve(d) of the
original image(c), starting point of image is upper left in left figure and
lower left in the right represent starting same point.

IV. EXPERIMENTAL RESULTS

After training of the network we use more than 1000 test
image for the recognition of the quadratics. we use image with
one quadratics ,and more than one quadratics. we consider
each of the quadratics character are separating with each other
in the test image. For applying the traditional algorithm we
face some problem and we modify those part in our system
for getting correct solution. Firstly we train the network after
completing the training we apply each test image for a) b)
recognition. For each of the quadratics recognition process in
the case of correct recognition we will provide a command for
finding a solution otherwise providing a command for next
quadratic recognition. For the segmentation of each line of
quadratics we use horizontal projection at each row of the
experimental image. Problem arises at the time of
segmentation when the square term makes a different
projection curve that is just above the equation’s other parts c) d)
projection as shown in figure 2 (a) and 2(b).In our proposed Figure 3. Applying simple connected component analysis to the original
compact horizontal projection based segmentation of image(a) resulting region selected output (c) problem is indicated with red
quadratics solve this problem by combining this two circle, combined connected component analysis to the original noise free
image(b) resulting region selected output(d) solution is indicated with red
projection curve into single one. Then the selection region for circle.
segmentation is as shown in the figure 2(c) and 2(d).
Another problem occurs when we apply connected component Here we use convolutional neural network for the solution of
analysis algorithm to the segmented quadratics. The symbol ‘= critical classification problem. In figure 4 the input image
‘ is a single character but composed with two distinct part as consist of three quadratics that is shown in the first part of the
shown in the figure 3 (c). In this research we solve this image. After that the noise free image , expected outcome and
actual outcome are also shown. Here a single character For each case we calculate the accuracy of the test character
recognition is misclassified that is shown in the figure 4 with and equation separately. we examine the total number of
red mark. If the detection of quadratics is absolutely correct correct recognition with respect to the input recognition . Im1
we then further process this equation for create the solution of ,Im2 and so on represents here the test image of equation.
the quadratics. Wrong detection does not need to further
process for the solution. We define this part as to select
manually whether the detection is correct or not.

a) b)
y2+2y=-1 y2+2y=-1
2 2
2y -10y+12=0 2y -10y712=0
3y2-6y+3=0 3y2-6y+3=0
c) d)
Figure 4. Experimental output a) original image b) Noise free image c)
Expected outcome from the input image d) system output where the symbol ‘+’
is misclassified as ‘7’ shown in red mark.
In figure 4 the actual outcome are shown where the symbol ‘+’
is misclassified as ‘7’ for this recognition the solution part is
not necessary that is why we command for next processing.
For a set of 1000 test equation image our system can
recognize 39.11% equation correctly, that is our system works
well with 39.11% equation detection accuracy. For the case of
character segmentation it segment character with 91.08%
accuracy. In case of correct detection our system successfully
find the solution for each cases. In Table I accuracy
calculation procedure for our test image the data used in table
are obtained from our testing from the system.
TABLE I. Accuracy calculation procedure in our test image in both
character(termed here symbol also character) and equation . Here “…” is used
to represents Im6 to Im 999.
Test
Figure 5. Training time elapsed comparision for 100 training data, traditional
model requires lot more training time than our proposed approach.
Traditional method with higher hidden layer though
sometimes provide high accuracy rate for the classification
task but requires lots of training time as shown in the Figure 5.
For the case of 100 training data elapsed time for training is
several times larger than our proposed network layer.
Handwritten character are sometime even becomes difficult
for human to recognize due to several reason ,such as different
people writes character in different shapes and style, different
writing tools are used by people that makes the character
sometimes more complex, another most important reason is
writing style of single character by a man is most often
different from a group of words or specially in mathematical
equation. For classification part the actual performance, P can
be measured by
P= ×100% (3)
Where X is the number of image character given as input for
classification and Y is actual number for correct classification
and m is the number of character those are conflict shape with
other character that even can’t classified by human. In figure 6
shown the comparison of the accuracy of equation in our
proposed one and with and without considering the

equation
Total

character
Total

Character
Correct
Recognized

equation
Correct
Recognized

rate
Recognition
Character

rate
Recognition
Equation

performance matric (3). In Figure 6 with consider performance

evaluation matric accuracy is slight higher than the traditional
one. Though it is not the solution of the actual problem , a
much more intelligent system is required to predict for the
Im1 3 30 29 2 96.67 66.67 solution of such problem what is still a research topic. Here
Im2 2 19 19 2 100 100 we can segment only the character that are separate from each
Im3 3 28 24 1 85.7 33.3
Im4 1 10 10 1 100 100
other. But in real environment some peoples handwritten
Im5 2 20 17 0 85 0 characters are not separate from each other. In such case our
….. .. .. .. .. .. .. segmentation method connected component fails to segment
Im1000 3 25 23 1 92 33.3 each distinct character individually. Another important note is
91.08% 39.11% in this paper we consider a quadratics of only the variable ‘y.
Other variables can also be used by increasing the number of
class in CNN. So other variable may not be applicable for this
system.

Figure 6. Equation recognition rate with considering performance evaluation
matric (P.M) shown in blue color curve and without considering curve shown
in figure with red color. Blue color curve sometimes slightly provide upper
accuracy.
V. CONCLUSION
In this paper we mainly focused on recognizing handwritten
mathematical quadratics. we consider a single quadratics and
also series of quadratics for the recognition. Projection
analysis specially compact horizontal projection is used for the
segmentation of each line of quadratics. Connected component
which has very high success rate is used for character
segmentation [19]. Improved version of connected component
is used for the symbol like ‘=’ detection which is a single
symbol combined with two distinct connected component
.Feature extraction the most complicated part of classification.
Moreover with some predefined feature about handwritten it is
difficult to recognize handwritten. Convolutional Neural
Network the most powerful classification model is used in the
classification part[24]. Once successful recognition of the
quadratics in any combination we further process the detected
equation for finding the solution of the quadratics. In this part
we apply an string operation algorithm for finding the value of
a, b, c of each quadratics in the form of ay2+by+c=0. Finally
we obtained a state of art performance in the detection phases
and also in the solution phases. In future we will try to
improve the accuracy and try to make the system
workable for multiple mathematical formula simultaneously.
Through the use of convolutional neural network we will try
to make a vision based mathematical formula dictionary in
future.
ACKNOWLEDGMENT
We are very much thankful to Computer Vision and Intelligent
Interfacing Lab(CVIIL) at Islamic University, Kushtia,
Bangladesh for providing all kind of logistic and other
supports .
REFERENCES
[1] Widad Jakjoud , Azzeddine Lazrek ‘Segmentation method of offline
mathematical symbols’, International Conference on Multimedia
Computing and Systems (ICMCS), 2011
[2] Zouaoui Abderaouf ‘license plate character segmentation based on
horizontal projection and connected component analysis’ 2014
world symposium on computer applications & research (wscar).
[3] Catherine Lu Karanveer Mohan.‘ Recognition of Online Handwritten
Mathematical Expressions Using Convolutional Neural Networks’
cs231n project report stanford 2015.
[4] Pooja Kamavisdar ,Sonam Saluja, Sonu Agrawal ‘A Survey on Image
Classification Approaches and Techniques’ international Journal of
View publication stats
Advanced Research in Computer and Communication Engineering
Vol. 2, Issue 1, January 2013.
[5] Leo Pauly , Peel, Shan Luo, David Hogg and Raul Fuentes.‘Deeper
Networks for Pavement Crack Detection’34th International
Symposium on Automation and Robotics in Construction (ISARC
2017).
[6] Ahmad-Montaser Awal ,Harold Mouchère, Christian Viard-
Gaudin.‘Towards Handwritten Mathematical Expression Recognition’
2009 10th International Conference on Document Analysis and
Recognition.
[7] Haiyan Wang, Shibao Li,Yang Zhang et al. ‘An Improved Algorithm
for Symbol Segmentation of Mathematical Formula Images’ 2016
16th ISCIT.
[8] Nicholas E. Matsakis ‘Recognition of Handwritten Mathematical
Expressions’ massachusetts institute of technology may 1999.
[9] Vinod Nair, Geoffrey E. Hinton” Rectified Linear Units Improve
Restricted Boltzmann Machines”2010.
[10] Krizhevsky, Alex, Ilya Sutskever, and Geoffrey E. Hinton.
"ImageNet classification with deep convolutional neural networks."
Advances in neural information processing systems. 2012.
[11] Long, J., Shelhamer, E., & Darrell, T. (2015). Fully convolutional
networks for semantic segmentation. In Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition (pp. 3431-
3440).
[12] Toshev, A., & Szegedy, C. (2014). Deeppose: Human pose
estimation via deep neural networks. In Proceedings of theIEEE
Conference on Computer Vision and Pattern Recognition (pp.
1653-1660).
[13] Dong, C., Loy, C. C., He, K., & Tang, X. (2014, September). Learning
a deep convolutional network for image super resolution. In
European Conference on Computer Vision (pp. 184-199). Springer
International Publishing.
[14] Shankar, S., Garg, V. K., & Cipolla, R. (2015). Deep-carving:
Discovering visual attributes by carving deep neural nets. In
Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition (pp. 3403-3412).
[15] Zhou, B., Lapedriza, A., Xiao, J., Torralba, A., & Oliva, A.
(2014). Learning deep features for scene recognition using places
database. In Advances in neural information processing systems (pp.
487-495).
[16] Wang, Naiyan, et al. "Transferring rich feature hierarchies for robust
visual tracking." arXiv preprint arXiv: 1501.04587(2015).
[17] Mao, Junhua, et al. "Deep captioning with multimodal recurrent
neural networks (m-rnn)." arXiv preprint arXiv: 1412.6632 (2014).
[18] Gharde, S. S., Baviskar, P. V., & Adhiya, K. P. (2013) Identification
of Handwritten Simple Mathematical Equation Based on SVM and
Projection Histogram. International Journal of Soft Computing and
Engineering (IJSCE),3(2),425-429
[19] Richard G. Casey, Eric Lecolinet ” A survey of methods and strategies
in character segmentation’1996
[20] Zouaoui Abderaouf, Benblidia Nadjia, Oukid-Khouas Saliha “License
plate character segmentation based on horizontal projection and
connected component analysis” 2014 World Symposium on
Computer Applications & Research (WSCAR).
[21] Shilpa C Vijayan, Jyothy R.L et al.” Histogram Based Connected
Component Analysis for Character Segmentation”International Journal
of Scientific and Research Publications, Volume 6, Issue 6, June 2016
[22] Saeeda Naz et al “Urdu Nasta’liqtext recognition using implicit
segmentation based on MDLSTM neural networks” Springer 2016.
[23] Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever. "Dropout: A
Simple Way to Prevent Neural Networks from Overfitting." Journal of
Machine Learning Research. Vol. 15, pp. 1929-1958, 2014.
[24] Course note of CS231n Stanford
[http://cs231n.github.io/convolutional-networks/]
[25] O. D. Trier, A. K. Jain, and T. Taxt. Feature extraction methods for
character recognition-a survey. Pattern Recognition, 29(4):641–662,
1996.
[26] Literature Review on Handwritten Mathematical Expression
Recognition, chapter 2.
[27] D. Cirean, U. Meier, and J. S. Schmidhuber. Multi-column deep neural
networks for image classification. CoRR, abs/1202.2745, 2012.