Available online at www.sciencedirect.

com
ScienceDirect
ScienceDirect
Procedia Computer Science 00 (2018) 000–000
Available online at www.sciencedirect.com
Procedia Computer Science 00 (2018) 000–000 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Computer Science 143 (2018) 528–535

8th International Conference on Advances in Computing and Communication (ICACC-2018)

8th International Conference on Advances in Computing and Communication (ICACC-2018)
8th International Conference on Advances in Computing and Communication (ICACC-2018)
BornoNet: Bangla Handwritten Characters Recognition Using
BornoNet: Bangla HandwrittenNeural
Convolutional Characters Recognition Using
Network
Convolutional Neural Network
AKM Shahariar Azad Rabbya*, Sadeka Haquea†, Md. Sanzidul Islama, Sheikh Abujara
AKM Shahariar Azad Rabbya*,
andSadeka Haquea†,
Syed Akhter Md. Sanzidul
Hossain a Islama, Sheikh Abujara
a
and Syed Akhter Hossain
a
Dept. of Computer Science and Engineering, Daffodil International University, Dhanmondi, Dhaka-1205, Bangladesh
a
Dept. of Computer Science and Engineering, Daffodil International University, Dhanmondi, Dhaka-1205, Bangladesh

Abstract
Abstract
Bangla handwriting recognition is becoming an important issue in several years but it becomes a challenge to get good performance
due to the
Bangla alignment recognition
handwriting and many ofisthem are similar.
becoming A simple,
an important lightweight
issue in severalCNN
yearsmodel
but it has been proposed
becomes in this
a challenge paper
to get goodfor classifying
performance
Bangla
due Handwriting
to the Character,
alignment and many ofwhich
them contains 50 A
are similar. basic Bangla
simple, characters
lightweight CNN(11model
vowelshas and 39 proposed
been consonants). Experiments
in this have been
paper for classifying
made
BanglaonHandwriting
three datasets along with
Character, the BanglaLekha-Isolated
which contains 50 basic Bangla [1] characters
CMATERdb (11[2] and the
vowels andISI
39[3] dataset. ForExperiments
consonants). character recognition,
have been
the proposed
made on threeBornoNet model
datasets along gets
with the98%, 96.81%, 95.71%, [1]
BanglaLekha-Isolated andCMATERdb
96.40% validation accuracy
[2] and the ISI [3]respectively
dataset. For for CMATERdb,
character ISI,
recognition,
BanglaLekha-Isolated
the proposed BornoNetdataset
modeland mixed
gets 98%,dataset.
96.81%,Also proposed
95.71%, andmodel
96.40%wasvalidation
trained with one dataset
accuracy and cross-validated
respectively for CMATERdb,with other
ISI,
two datasets. Proposeddataset
BanglaLekha-Isolated model and
achieved
mixedthe best accuracy
dataset. rate so model
Also proposed far forwas
BanglaLekha-Isolated, CMATERdb
trained with one dataset and ISI datasets.
and cross-validated with other
The proposedProposed
two datasets. BornoNetmodel
modelachieved
can be found on accuracy
the best https://github.com/shahariarrabby/BornoNet
rate so far for BanglaLekha-Isolated, CMATERdb and ISI datasets.
The proposed BornoNet model can be found on https://github.com/shahariarrabby/BornoNet
© 2018 The Authors. Published by Elsevier B.V.
© 2018
This The
is an Authors.
open accessPublished by Elsevier
article under B.V.
the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
© 2018 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
Selection
This is an and
open peer-review
access under
article underresponsibility
the CC of the scientific
BY-NC-ND committee of the 8th International Conference on Advances in
license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
Selection and peer-review under responsibility of the scientific committee of the 8th International Conference on Advances in
Computing
Selection and
and Communication
peer-review under (ICACC-2018).
responsibility
Computing and Communication (ICACC-2018). of the scientific committee of the 8th International Conference on Advances in
Computing and Communication (ICACC-2018).
Keywords: Handwritten Recognition; Pattern Recognition; Document Image Analysis; Machine Learning; Computer Vision.
Keywords: Handwritten Recognition; Pattern Recognition; Document Image Analysis; Machine Learning; Computer Vision.

* AKM Shahariar Azad Rabby.

E-mailShahariar
* AKM address: Azad
[email protected]
Rabby.
* Sadeka Haque [email protected]
E-mail address:
* E-mail
Sadeka address:
Haque [email protected]
E-mail address: [email protected]
1877-0509 © 2018 The Authors. Published by Elsevier B.V.
This is an open
1877-0509 access
© 2018 Thearticle under
Authors. the CC BY-NC-ND
Published license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
by Elsevier B.V.
Selection
This is an and
openpeer-review under
access article responsibility
under of the scientific
the CC BY-NC-ND licensecommittee of the 8th International Conference on Advances in Computing and
(https://creativecommons.org/licenses/by-nc-nd/4.0/)
Communication
Selection (ICACC-2018).
and peer-review under responsibility of the scientific committee of the 8th International Conference on Advances in Computing and
Communication (ICACC-2018).

1877-0509 © 2018 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
Selection and peer-review under responsibility of the scientific committee of the 8th International Conference on Advances in Computing
and Communication (ICACC-2018).
10.1016/j.procs.2018.10.426
 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535 529
2 AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000

1. Introduction

In recent years Convolutional neural network (CNN) becomes popular for complex visual recognition in due to its
architecture and a lot of research on using deep CNN to recognize handwritten digits, characters, etc. And we will
explore the power of deep CNN on the classification of handwritten Bangla characters. Handwritten recognition has
importance for its various applications as Optical Character Recognition (OCR). Several scientific types of research
have been carried out for character recognition of English, Chinese, Arabic handwritten digits and alphabet. Though
automatic recognition of handwriting remains a great challenge: the performance of layout analysis, word/character
segmentation, and recognition is still far behind the human recognition capability.
Bangla is the mother language of Bangladesh, apart from it is the official language of Bangladesh, West Bengal of
India, Tripura, Assam and Jharkhand, Sierra Leone a West African country. Though Bangla is the 7th most popular
language and writing scripts like about 250 million speak in Bangla and 2nd most beautiful language in the world.
Considering those all circumstance the technology in different sectors in these regions, Bangla Handwritten
recognition plays an important role and should overcome the challenge. However, to compare other language writing
script only a few studies are attested on handwritten characters of Bangla scripts there have a sturdy model such as
Latin [4], Chines [5], Japanese [6] have achieved a great success on machine learning and deep learning application.
Conducting with the handwritten Character or digit recognition always a big challenge in due to its variation of shape,
size and complexity of Bangla text such as there have many misleading characters. But several years few works have
done for Bangla handwritten digit recognition using CNN and gain some success of classifying those type of
recognition to the researcher for its educational and economic interesting applications.
Bangla writing scripts different from English or other languages writing scripts because of it came from Sanskrit
script which is completely different and it has alinement and some character are similar to another character as some
of them are different from small dot and line. There is 50 basic character (11 vowels and 39 consonants), 10 modifiers,
10 numerals, more than 300 compound characters. Due to those, all fact Bangla scripts make difficult to achieve a
good result and better performance with Bangla Handwritten character recognition. If we overcome all these
challenges and make a model then we improved many kinds of Handwritten recognition-based application such as
Bangla Handwritten character base OCR (Optical Character Recognition), Picture to text to speech, Bangla ID card
reading, Number plate reading, vehicle tracking, Post office automation etc. Fig 1 is example of Bangla characters.

Fig 1: Example of Bangla Characters

2. Literature review

In the area of Bangla character classification, previous works have mainly focused on Bangla digit, which contains
10 digits.
There are a few works are available for Bangla handwritten character recognition. Some literature has been reported
on Bangla characters recognition in the past years as “A complete printed Bangla OCR system” [7], “On the
development of an optical character recognition (OCR) system for printed Bangla script” [8] but few research on
handwritten Bangla numeral recognition that reaches to the desired recognition accuracy. Pal et al. have conducted
some exploring works for recognizing handwritten Bangla characters those are “Automatic recognition of
unconstrained offline Bangla hand-written numerals” [9], “A system towards Indian postal automation” [10] and
530 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535
AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000 3

“Touching numeral segmentation using water reservoir concept” [11]. The proposed schemes are mainly based on
extracted features from a concept called water reservoir. Apart from there also present several Bangla Handwritten
Character Recognition and had achieved pretty good success. Halima Begum et al, “Recognition of Handwritten
Bangla Characters using Gabor Filter and Artificial Neural Network” [12] works with own dataset that was collected
from 95 volunteers and their proposed model achieved without feature extraction and with feature extraction around
68.9% and 79:4% of recognition rate respectively. “Recognition of Handwritten Bangla Basic Character and Digit
Using Convex Hull Basic Feature” [13] accuracy for Bangla character 76.86% and Bangla numeral 99.45%. “Bangla
Handwritten Character Recognition using Convolutional Neural Network” [14] achieved 85.36% test accuracy using
their own dataset. In “Handwritten Bangla Basic and Compound character recognition using MLP and SVM classifier”
[15] handwritten Bangla character recognition with MLP and SVM has been proposed and they achieved around
79.73% and 80.9% of recognition rate, respectively.

3. Proposed methodology

In this paper to build the proposed model “BornoNet” using Convolutional Neural Network (CNN) which has many
phases that are following in below.

3.1. Dataset

The proposed method used three datasets BanglaLekha-Isolated, CMATERdb and ISI datasets. These three datasets
consist of Bangla characters, Bangla numeral digits and Bangla compound characters.
CMATERdb database has total 15,000 characters images. The image was 32 x 32px noise-free images and the
image edge looks blocker. ISI Dataset has total 37,858 images. The image has different pixel size and image edge
look smother. BanglaLekha-Isolated has total 98,950 images. The image is inverted and resized and add some padding
to be square with preserving the aspect ratio.
By joining all these three datasets we make a mixed dataset of 151580 images. Fig 2 showing an example of the
different datasets.

(a) (b) (c)

Fig 2: (a) BanglaLekha (b) ISI (c) CMATERdb

3.2. Dataset preparation

Initially we prepared dataset for building the model. First fixed some incorrect labeling images from the
BanglaLekha-Isolated dataset. Delete some incorrect images such as blank images. This dataset already containing
inverted image prepossessed with its inverted foreground and background, removal of noise with the median filter,
edge thickening.
For CMATERdb and ISI dataset first thresholding, grayscaling and invert all the images.
Secondly for all datasets resized all images to 28 x 28 pixel. Then converted all dataset’s 28 x 28-pixel image into
a 784 + (1 label) D matrix and store all the image pixel into a CSV file to reduce hard disk read and fast computation.
Also, perform a Minmax (1) normalization to reduce the effect of illumination differences. Moreover, the CNN
converge faster on (0...1) data than on (0...255)

𝑋𝑋 −minmum⁡(𝑋𝑋)
𝑖𝑖
𝑍𝑍𝑖𝑖 = maximum (1)
⁡(𝑋𝑋)−minmum⁡(𝑋𝑋)

Then convert the 10 labels into one hot encoding

(1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0) = অ
(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 , 0, 0, 0, 0, 0, 0, 0, 1) = ৺
 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535 531
4 AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000

Then convert the 784 D into 28 x 28 image matrix.

3.3. Proposed model

The proposed model is a 13-layer convolutional neural network with 2 sub-layers. Proposed model use ADAM
[16] optimizer. For first two-layer same padding and ReLU (2) activation used with 32 filters with the 5x5 kernel.
Then a max-pooling layer added with a 2x2 followed by 25% dropout layer. All dropout layer used to reduce
overfitting.
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅(𝑋𝑋) ⁡ = ⁡𝑀𝑀𝑀𝑀𝑀𝑀(0, 𝑋𝑋) (2)

The output from this two-layer goes as an input of two sublayers. Where both sublayers have same 2 convolutional
layers with the same padding, ReLU activation, 64 filters with a 5x5 kernel, followed by another convolutional layer
with a 3x3 kernel. Later the output of last 2 sub convolutional layer added together and pass through a Max-Pooling
layer with 25% dropout [17] layer.
Then flatten the layer and used a fully connected layer with 1280 hidden node followed by 25% dropout. Final
output layer has 50 nodes with SoftMax (3) activation. Fig 3 is showing the architecture of the network.
𝑧𝑧
ⅇ 𝑗𝑗
𝜎𝜎(𝑧𝑧)𝑗𝑗 = 𝐾𝐾 ⁡𝑓𝑓𝑓𝑓𝑓𝑓⁡𝑗𝑗 = 1, … 𝑘𝑘 (3)
∑𝑘𝑘=1 ⅇ 𝑧𝑧𝑘𝑘 ⁡

Fig 3: Prosed BornoNet architect.

Table 1. BornoNet model architect summary.

Layer No (type) Output Shape Param Connected to Layer No (type) Output Shape Param Connected to
(Input Layer) 28, 28, 1 0 - 8 (Conv2D) 14, 14, 64 36928 7
1 (Conv2D) 28, 28, 32 832 Input Layer Add 6 8 (Add) 14, 14, 64 0 6, 8
2 (Conv2D) 28, 28, 32 25632 1 9 (MaxPooling2D) 7, 7, 64 0 Add_ 6_8
3 (MaxPooling2D) 14, 14, 32 0 2 10 (Dropout) 7, 7, 64 0 9
4 (Dropout) 14, 14, 32 0 3 Flatten (Flatten) 3136 0 10
5 (Conv2D) 14, 14, 64 51264 4 11 (Dense) 1280 4015360 Flatten
7 (Conv2D) 14, 14, 64 51264 4 12 (Dropout) 1280 0 11
6 (Conv2D) 14, 14, 64 36928 5 Output (Dense) 50 64050 12
Total parameters: 4,282,258
Trainable parameters: 4,282,258
Non-trainable parameters: 0

3.4. Optimizer and Learning rate

In Deep Learning and computer vision work, the optimization algorithm can change the result and make it pretty
532 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535
AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000 5

sufficient. The Adam paper says, "...many objective functions are composed of a sum of subfunctions evaluated at
different subsamples of data; in this case, optimization can be made more efficient by taking gradient steps w.r.t.
individual sub-functions ..." [16]. The Adam optimization algorithm is straightforward to implement, computationally
efficient and little memory requirements that’s why recently adopting most of the computer vision and natural
language processing application. Proposed method used ADAM (4) Optimizer with learning rate 0.001.
𝜂𝜂
𝜃𝜃𝑡𝑡+1 = 𝜃𝜃𝑡𝑡 − 1 𝑚𝑚
̂ 𝑡𝑡 (4)
√𝑣𝑣̂𝑡𝑡 +𝜀𝜀

When using a neural network to perform classification and prediction task we need to calculate the error rate. A
recent study shows that cross entropy function performs better than classification error and mean square error. [18]
Proposed method used categorical cross entropy (4) as loss function.

Li = − ∑j t ⁡i,⁡⁡⁡j ⁡log(pi,⁡⁡⁡j ) (5)

To make the optimizer converge faster and closer to the global minimum of the loss function, using an automatic
Learning Rate reduction method [19]. Learning rate is the step by which walks through the minimum loss. If the
learning rate is too low, it will take more time to reach the global minima, and if the learning rate is too high then the
training may not be converging or even diverge. To keep the advantage of the fast computation time we set a high
learning rate which is automatically decreased by monitoring the validation accuracy.

3.5. Data augmentation

To avoid overfitting, artificially expand the handwritten dataset. This data transformation will create some variance
that can occur when someone else writing the digits. For Data augmentation, several methods are chosen:
• Randomly shifting height and width 10% of the images.
• Randomly rotate our training image 10 degrees.
• Randomly 10 % zoom the training image.

3.6. Training the model

The proposed model was trained with different training and validation set with the batch size of 86. During training,
automatic learning rate reduction formula mentored the validation accuracy and reduce the learning rate if required.

4. Evaluate the model

The proposed model was applied to the three-different dataset and gain good result on train, test, validation sets.

4.1. Train, Test and Validation sets

For CMATERdb 12,000 images used as train image and 3,000 for the test image. We take 10% of train images for
validation purpose. After training model was tested with 3000 images of CMATERdb test dataset also with 37,858
images of ISI dataset and 98,722 images of the BanglaLekha-Isolated dataset.
For ISI, the dataset has 37,858 images, where we use 7000 as validation images and 12,859 as test images also
used 15,000 images of CMATERdb dataset and 98,722 images of BanglaLekha-Isolated for testing the model.
For BanglaLekha-Isolated 98,722 images we took 88,849 images for training and 9,873(100%) image as a
validation set. After training model tested with 15,000 images of CMATERdb dataset and 37,858 images of ISI
dataset.
For Mixed Datasets 151,580 images we make the training set of 121,263 images by taking 80% of each dataset.
10% used as validation and other 10% used as test set.
 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535 533
6 AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000

4.2. Model performance

After 30 epoch proposed model gets 98%, 96.81%, 95.71%, and 96.40% validation accuracy respectively for
CMATERdb, ISI, BanglaLekha- Isolated dataset and mixed dataset. Also, all of this dataset cross-validate with each
other and perform accurately. Fig 4 (a) (b) (c) (d) is showing the accuracy and loss of training and validation set of
BanglaLekha-Isolated, CMATERdb, ISI and mixed dataset respectively. Table 2 showing the details accuracy on
different datasets.

(a) (b)

(c) (d)
Fig 4: (a) BanglaLekha-Isolated (b) CMATERdb (c) ISI (d) Mixed Database

Table 2. Result comparison in different Dataset.

Dataset Name Tr. Loss Val. Loss Tr. Acc. Val Acc. Test Dataset Test Acc.
BanglaLekha 0.1455 0.1679 95.45% 95.71% CMATERdb 95.22%
ISI 93.59%
CMATERdb 0.0475 0.0817 98.44% 98.00% BanglaLekha 78.29%
ISI 92.19%
ISI 0.0814 0.1120 97.37% 96.81% BanglaLekha 84.63%
CMATERdb 94.94%
Mixed Dataset 0.1298 0.1396 95.81% 96.40% Mixed Test 96.44%
BanglaLekha 97.96%
CMATERdb 99.12%
ISI 99.18%
534 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535
AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000 7

4.3. Result analysis

Analyzing the result and confusion, we found that CMATERdb dataset got best validation accuracy but perform
poorly with another dataset. That because CMATERdb dataset contains the image that is noise free. That’s why when
noise image from other datasets came it fails to recognize. Confusion matrix can be found on
https://github.com/shahariarrabby/BornoNet/graph.
On the other hand, BanglaLekha-Isolated gets less accuracy but perform very well on other datasets. BanglaLekha-
isolated has the noisy image. When a noise-free image from another dataset came it perform very well.

4.4. Result comparison

Table 3 showing a comparison between some the previous work. From this table, we found that our proposed
BornoNet got so far, the best accuracy rate for all of the three datasets.

Table 3. Result Comparison in different Model.

Work Accuracy Work Accuracy
Recognition of Handwritten Bangla Characters 79.4% HMM-Based Online Handwritten Bangla 91.85%
Using Gabor Filter and Artificial Neural Network Character Recognition using Dirichlet
[12] Distributions [21]
Recognition of Bangla handwritten basic characters 76.86% Bangla Hand-Written Character Recognition 93.43%
and digits using convex hull-based feature set [13] Using Support Vector Machine [22]
Handwritten Bangla Character Recognition Using 84.00% Bengali handwritten character recognition 95.00%
Neural Network [20] using Modified syntactic method [23]
Bangla Handwritten Character Recognition using 85.36% BornoNet (Proposed) 95.71% (Isolated)
Convolutional Neural Network [14] 98% (CMATER)
96.81% (ISI)

5. Error observation

Analyzing the error from validation set we found that most of the incorrect classification is caused by error labeling
on the dataset. It is clear that model performing great for classifying characters. Some of this mistake can also make
by humans. Fig 5 showing top 6 validation error from each dataset.

(a) (b) (c)

Fig 5: Top six Error form (a) BanglaLekha-Isolated (b) CMATERdb (c) ISI

6. Conclusion and Future work

This research work presented a new CNN model which performs better classification accuracy in the different
dataset for both train and validation set for lesser epochs and less computation time compared to the other CNN model.
CNN's in general costly to train but extremely effective deep learning models. Hence, fast convergence should be
 Akm Shahariar Azad Rabby et al. / Procedia Computer Science 143 (2018) 528–535 535
8 AKM Shahariar Azad Rabby et al./ Procedia Computer Science 00 (2018) 000–000

viewed as an important front of research. Also, the cross-validation from different distribution’s data proposed model
achieve a great result that makes it a robust model that improve any other previous model.
Sometimes proposed model confused to understand overwritten character and dataset contained some incorrect
labeling images. Also, the model performed poorly if the train on noise-free data. In future work fixing dataset and
overcoming the limitation of overwriting Character should fix. Making a benchmark model for Bangla Handwritten
all characters that include the numeral, basic characters, Bangla modifier and compound letter.

References

[1] Biswas, Mithun & Islam, Rafiqul & Kumar Shom, Gautam & Shopon, Md & Mohammed, Nabeel & Momen, Sifat & Abedin, Anowarul. (2017).
BanglaLekha-Isolated: A multi-purpose comprehensive dataset of Handwritten Bangla Isolated characters. Data in Brief. 12. 103-107.
10.1016/j.dib.2017.03.035.
[2] R. Sarkar, N. Das, S. Basu, M. Kundu, M. Nasipuri, and D. K. Basu, “Cmaterdb1: a database of unconstrained handwritten Bangla and Bangla–
English mixed script document image,” International Journal on Document Analysis and Recognition (IJDAR), vol. 15, no. 1, pp. 71–83, 2012.
[3] Bhattacharya, Ujjwal & Chaudhuri, Bidyut. (2009). Handwritten Numeral Databases of Indian Scripts and Multistage Recognition of Mixed
Numerals. IEEE transactions on pattern analysis and machine intelligence. 31. 444-57. 10.1109/TPAMI.2008.88
[4] E. Kavallieratou, N. Liolios, E. Koutsogeorgos, N. Fakotakis, G. Kokkinakis, "The GRUHD Database of Greek Unconstrained Handwriting",
ICDAR, 2001, pp. 561-565.
[5] F. Yin, Q.-F. Wang, X.-Y. Zhang, and C.-L. Liu. ICDAR 2013 Chinese Handwriting Recognition Competition. 2013 12th International
Conference on Document Analysis and Recognition (ICDAR), pages 1464–1470, 2013
[6] B. Zhu, X.-D. Zhou, C.-L. Liu, and M. Nakagawa. A robust model for on-line handwritten Japanese text recognition. IJDAR, 13(2):121–131,
Jan. 2010.
[7] B. B. Chaudhuri, U. Pal, “A complete printed Bangla OCR system,” Pattern Recognition, vol. 31, pp. 531–549, 1998.
[8] U. Pal, “On the development of an optical character recognition (OCR) system for printed Bangla script,” Ph.D. Thesis, 1997.
[9] U. Pal, B. B. Chaudhuri, “Automatic recognition of unconstrained offline Bangla hand-written numerals,” T. Tan, Y. Shi, W. Gao (Eds.),
Advances in Multimodal Interfaces, Lecture Notes in Computer Science, vol. 1948, Springer, Berlin, pp. 371–378, 2000.
[10] K. Roy, S. Vajda, U. Pal, B.B. Chaudhuri, “A system towards Indian postal automation,” Proceedings of the Ninth International Workshop on
Frontiers in Handwritten Recognition (IWFHR-9), pp. 580–585, October 2004.
[11] U. Pal, A. Belad, Ch. Choisy, “Touching numeral segmentation using water reservoir concept,” Pattern Recognition Lett. vol. 24, pp. 261-272,
2003.
[12] Halima Begum et al, Recognition of Handwritten Bangla Characters using Gabor Filter and Artificial Neural Network, International Journal
of Computer Technology & Applications, Vol 8(5),618-621, ISSN:2229-6093
[13] Nibaran Das, Sandip Pramanik, “Recognition of Handwritten Bangla Basic Character and Digit Using Convex Hall Basic Feature”. 2009
International Conference on Artificial Intelligence and Pattern Recognition(AIPR-09)
[14] Md. M. Rahman, M. A. H. Akhand, S. Islam, P. C. Shill, “Bangla Handwritten Character Recognition using Convolutional Neural Network,”
I. J. Image, Graphics and Signal Processing, vol. 8, pp. 42-49, July 15, 2015.
[15] N. Das 1, B. Das, R. Sarkar, S. Basu, M. Kundu, M. Nasipuri, “Handwritten Bangla Basic and Compound character recognition using MLP
and SVM classifier,” Journal of Computing, vol. 2, Feb. 2010.
[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.CoRR, abs/1412.6980,2132014..
[17] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural
networks from overfitting.Journal of Machine Learning Research,20915:1929–1958, 2014.
[18] Katarzyna Janocha and Wojciech Marian Czarnecki. On loss functions for deep neural networks in classification. Arxiv, abs/1702.05659, 2017
[19] Tom Schaul, Sixin Zhang, and Yann LeCun. No more pesky learning rates. In Sanjoy Dasgupta and David McAllester, editors, Proceedings
of the 30th International Conference on Machine Learning, volume 28 ofProceedings of Machine Learning Research, pages 343–351, Atlanta,
Georgia, USA, 17–19 Jun 2013. PMLR.
[20] Md. Alamgir Badsha et al, Handwritten Bangla Character Recognition Using Neural Network, International Journal of Advanced Research
in Computer Science and Software Engineering 2 (11), November- 2012, pp. 307-312
[21] Chandan Biswas et al, HMM Based Online Handwritten Bangla Character Recognition using Dirichlet Distributions, 978-0-7695-4774-9/12
$26.00 © 2012 IEEE DOI 10.1109/ICFHR.2012.213
[22] Riasat Azim et al, Bangla Hand-Written Character Recognition Using Support Vector Machine, International Journal of Engineering Works
vol. 3, Issue 6, PP. 36-46, June 2016.
[23] Mohammad Badiul Islam et al, Bengali handwritten character recognition using modified syntactic method, NCCPB-2005 © Independent
University, Bangladesh