International Journal of Advanced Science and Technology

Vol. 29, No. 05, (2020), pp. 322-329

Performance Analysis of Artificial Neural Network Algorithms for

Automatic Handwritten Devanagari Text Generation in Marathi Styles
Mr. Vajid Khan1, Dr. Yogesh Kumar Sharma2
1
Research Scholar, 2Professor, Research Coordinator JJTU Rajasthan
1-2
Deparment of Computer science and engineering,
1
Research Scholar, Shri JJTU, Rajasthan, Professors, 2Head of Department, Research
Coordinator Shri JJTU, Rajasthan.

Abstract
Devanagari content extensively utilized the critical content in India. Generally that content can be
utilized to compose Sanskrit, Nepali, Marathi and Hindi dialects. With the expanded interest,
investigation and globalization of advanced Devanagari records, diverse printed and manually
written archive acknowledgment systems have included since most recent 20 years. The means that
are engaged with character acknowledgment and content age are pre-handling, division, include
extraction and order. There are various methods for devanagiri content like K Nearest Neighbour
(KNN), Artificial Neural Network (ANN) and Support Vector Machine (SVM). The ANN strategy is
actualized inside the python stage and it contrasted and the Artificial Neural Network (ANN). The
presentation of the proposed strategy is broke down with factual estimations of exactness, accuracy
and review.
Keyword: Devanagari script, Marathi, AMA, ANN, KNN, feature extraction, pre-processing,
segmentation.

I. INTRODUCTION
Text generation is one of the subfields of natural language processing. It was knowledge leverages on
the artificial intelligence and computational linguistics for the natural language texts generated by
naturally, certain communicative requirements can have satisfied by these texts. The combination of
practical and different commercial applications is called text generation, in which manuscripts,
reading forms and their archival etc. Like as a less uses only interact the keyboard facilitates system.

As well as the text is moreover hand-written or printed which terms are directly transferred to the
machine. Mainly the visually handicapped getting more benefits by this text generation when
interfaced with a voice synthesizer. The building changes of text recognition system can combine the
performance of human competition as well as providing a strong motivation to the field researches.
The creation of Devanagari script is nature and symbols and character composition are being in the
two-dimensional word that can make the difference from the ideographic and Roman scripts.

The basic leadership part of an acknowledgment framework is the characterization stage and it
utilizes the highlights extricated from the past stage. There are different techniques for order. K
Nearest Neighbor (KNN), Artificial Neural Network (ANN) and Support Vector Machine (SVM).
The qualities of the some characterization strategies that have been effectively applied to Offline
Devnagari Marathi and Sanskrit word acknowledgment and aftereffects of Neural Network order is
superior to anything other arrangement techniques, applied on Handwritten Devnagari word.

Artificial neural networks are one of the mainstream strategies utilized for order because of their
learning and speculation capacities. A multilayer perceptron (MLP) is a feedforward fake neural
framework model that maps sets of data onto a great deal of fitting yields. A MLP comprises of
numerous layers of hubs in a coordinated diagram, with each layer completely associated with the
following one. The advancement strategy utilized for preparing this engineering was the Scaled
Conjugate Gradient (SCG) technique. SCG technique was utilized on the grounds that it gave better

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
322
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

outcomes and has been found to take care of the improvement issues experienced when preparing a
MLP organize more productively than the inclination plunge and conjugate angle strategies.
II. RELATED WORK

Namita Dwivedi have depicted acknowledgment of Sanskrit word utilizing Prewitt's administrator for
removing the highlights from a picture diminishing procedure is applied in pre-preparing system
Thinning is a significant pre-handling step in OCR. The motivation behind diminishing is to erase
excess data and simultaneously hold the trademark highlights of the picture. Freeman Chain code is
one of the portrayal methods that is valuable for picture handling, shape investigation and example
acknowledgment fields is utilized with heuristic methodology for include extraction.

Chhabra, A., (2019) have inspected the inconsistencies in transcribed Devanagari content. This takes
while composing and it will lead the acknowledgment procedure increasingly mind boggling. A
portion of the inconsistencies in written by hand Devanagari is referenced as pursues; nearness of
undesirable or additional strokes, unusual size of a vowel image, composing vowel images in
detachment, inappropriate connection of lower vowel images, inadequate and erroneous portrayal of a
vowel image, over composition, wrong position of header line, converging of vowel images with
feature, tight composition, interruption of upper vowel images with center district, deficient
composition and interruption of lower vowel images with mid locale.

Srivastava, S., et al., (2019) have talked about a few given in regards to OCR structuring for Indian
contents. The order of determined content and textual style is finished utilizing the element
determination device. Certain devices like division examination device and information assortment
was additionally introduced by the specialists. The computerized cameras were connected by a few
gadgets which will be handheld.

These gadgets help to catch some typical scene pictures in the extraction of printed data will be an
issue. The data that was extracted from the content was further send to the content to discourse motor
or OCR for preparing. Sahare, P. what's more, Dhok, S.B., (2019) have introduced the digitalization
of the world as of late. In future the information on each human will be in the advanced configuration
inside the web.

A few issues emerge in the event of uncontrolled traffic. This prompts long traffic clog, mishaps
lastly the long periods of voyaging will be deferred. Because of increment in populace the utilization
of fuel get improved in top hours. The stopping the executives can't direct delicate procedure of
stopping in these pinnacle hours.

To defeat these issues many research were being made. At last the programmed number plate
acknowledgment was created to control the ideal opportunity for traffic clog issues, for example,
stopping assortment, cost and so forth. Vaishnav, A. what's more, Mandot, M., 2020 have inspected
the acknowledgment of character in number plates which pays the route for the structuring of the
ANPR framework. In vehicles the number plate is the noteworthy segment. This will encourages us to
discover various vehicles. The planning of the character help to discover the character in the number
plate utilizing programmed number plate acknowledgment (ANPR). In India various kinds of text
style faces are utilized to for writing in the number plates. Numerous specialists were keen on
creating character acknowledgment for the number plates. For perceiving the characters in number
diverse Hindi and English textual style faces database were created. For perceiving character in
number plates the regularly utilized textual style faces incorporates Kruti dev, Calibri and times new
roman.

The acknowledgment framework utilized for perceiving the character in number plate was the optical
character acknowledgment. For performing investigation in the introduced framework diverse number

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
323
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

plates with English and Hindi characters were utilized. In information pictures which the example the
acknowledgment was and division pace of around 98% and 96% individually.
By methods for devoted libraries numerous creators have created picture pre-handling techniques for
object acknowledgment and furthermore for optical character acknowledgment. Rizvi, M., et al.,
(2019) have planned article acknowledgment and optical character acknowledgment which are PC
vision advances by methods for MNIST database. At present the PC form innovation was developing
to gain ground in programming and equipment innovation.

Through these procedures the vital data can be separated from the pictures and this will help to settle
on better choices. For adjusting the pictures the pre-preparing strategy was utilized. In the main page
of the test duplicate the data's were separated. By methods for highlight extraction calculation the
mark between the analyst and invigilator were grouped and these marks were considered as articles.

III. ANNS FOR TEXT GENERATION SYSTEM

ANNs (fake neural system) can be considered as structure that contains versatile basic preparing
components called counterfeit neurons, which are firmly between associated. These counterfeit
neurons have the ability to perform broadly equal calculations for information handling and
information portrayal. Neural systems are normally layered structures.

There are interconnected hubs in each layer. Hubs contain an enactment work. Info layer gets input
information, the yield layer gives characterization results and others are shrouded layers present
between the info and yield layer. Associations between hubs have loads that are changed by the info
and yield gave. Info is passed into the main layer.

Singular neurons get the contributions, with every one of them getting a particular worth. After this, a
yield is delivered dependent on these qualities. The yields of the primary layer are then passed into
the subsequent layer to be prepared. This proceeds until the last yield is created. The supposition that
will be that the right yield is predefined. The system, as a result, trains itself.
IV. ANALYSIS OF ANN ALGORITHMS

Another existing method which is stated in this chapter is ANN. similarly to evaluate the performance
of ANN the statistical parameter is evaluated for ANN also. The values of statistical parameter for
ANN are given in the table 4.1, 4.2, 4.3, 4.4.

Table 4.1: Statistical Measurements for ANN for 10 classes

S.No Classes ANN

1 0 Precision Recall Accuracy
2 1 0.85 0.87 0.84
3 2 0.77 0.86 0.84
4 3 0.79 0.88 0.91
5 4 0.82 0.94 0.77
6 5 0.89 0.75 0.75
7 6 0.81 0.68 0.98
8 7 0.75 0.45 0.64
9 8 0.61 0.91 0.451
10 9 0.75 0.91 0.542

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
324
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

Table 4.2: Statistical Measurements for ANN for 10 classes

ANN
Sl. No Classes
Precision Recall Accuracy
11 10-ka 0.71 0.75 0.664
12 11-kha 0.79 0.79 0.5441
13 12-ga 0.75 0.85 0.758
14 13-gha 0.79 0.74 0.658
15 14-kna 0.86 0.79 0.755
16 15-cha 0.91 0.71 0.454
17 16-chha 0.84 0.78 0.874
18 17-ja 0.75 0.79 0.91
19 18-jha 0.79 0.69 0.54
20 19-yna 0.81 0.68 0.874

Table 4.3: Statistical Measurements for ANN for 10 classes

ANN
Sl. No Classes
Precision Recall Accuracy
21 20-taamatar 0.92 0.89 0.95
22 21-thaa 0.87 0.87 0.84
23 22-da 0.87 0.82 0.84
24 23-dhaa 0.92 0.84 0.91
25 24-adna 0.86 0.86 1
26 25-tabala 0.85 0.89 1
27 26-tha 0.87 0.81 0.95
28 27-da 0.91 0.86 0.99
29 28-dha 0.95 0.81 0.75
30 29-na 0.99 0.91 0.99

Table 4.4: Statistical Measurements for ANN for 16 classes

Sl. ANN
Classes
No Precision Recall Accuracy
31 30-pa 0.81 0.84 0.84
32 31-pha 0.89 1 0.82
33 32-ba 0.87 1 0.83
34 33-bha 0.85 0.91 0.845
35 34-ma 0.94 0.95 0.814
36 35-yaw 0.87 0.87 0.65
37 36-ra 0.84 0.85 0.89
38 37-la 0.86 0.84 0.89
39 38-waw 0.83 0.82 0.91
40 39-motosaw 1 0.87 0.92
40-
41 0.81 0.91 0.87
petchiryakha
42 41-patalosaw 0.82 0.89 0.81
43 42-ha 0.91 0.85 0.83
44 43-chhya 0.81 0.81 0.94
45 44-tra 0.85 0.80 0.92
46 45-gya 0.81 0.79 0.91
Average Value 0.88 0.86 0.8825

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
325
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

Figure 4.1: Analysis of Precision for ANN

Figure 4.2: Analysis of Recall for ANN

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
326
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

Figure 4.3: Analysis of Accuracy for ANN

In the above table the values for statistical parameter like accuracy, recall and precision for method
ANN is given. This table is given in graphical format in the following section for clear analysis
purpose. In figure 4.1, 4.2, 4.3 the graphical representation of the method ANN is given

In the above figure the graphical representation of the existing method ANN is provided. This
statistical parameter is used to demonstrate the performance of method ANN and it is also used for
comparison purpose.
V. CONCLUSIONS
This paper proposed an ANN calculation for acknowledgment and content age of Devanagari content
in Marathi manually written archives. The acknowledgment and content age of the Devanagari
content can be accomplished with the diverse procedure, for example, pre-handling, division,
highlight extraction and acknowledgment with content age. The pre-handling comprises of clamor
evacuation, binarization and slant remedy; division comprises of word division, character division
and line division; Feature extraction Recognition and content age is accomplished with the assistance
of the ANN calculation. The proposed technique is actualized in python stage and it is contrasted and
the current strategies, for example, ANN and KNN. This examination work can be additionally
stretched out for acknowledgment of various Devanagari contents for instance Marathi with various
component extraction strategy.

VI. ACKNOWLEDGMENT

I might want to communicate my exceptionally incredible thankfulness to Dr. Yogesh Kumar Sharma
for his important and valuable recommendations during the arranging and advancement of this
examination work. His ability to give his time so liberally has been particularly valued.

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
327
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

VII. REFERENCES
[1] Naveen malik and Aashdeep singh, “review of character recognition of offline handwritten
Devanagari script”, an international journal of computer science and mobile computing, vol.5,
pp-178-183, 2016.
[2] Vijay Pal, Dhaka and Manoj Kumar Sharma, "An efficient segmentation technique for
Devanagari offline handwritten scripts using the Feed forward Neural Network", an international
journal of Neural Computing and Applications, vol. 26, no. 8, 2015.
[3] Pratibha,Singh, , Ajay Verma, and Narendra S. Chaudhari. "Feature selection based classifier
combination approach for handwritten Devanagari numeral recognition", vol. 40, no. 6 2015.
[4] Pooja,Keshri, , Prabhat Kumar, and Rajib Ghosh, "RNN Based Online Handwritten Word
Recognition in Devanagari Script.", In proceeding of International Conference on Frontiers in
Handwriting Recognition (ICFHR), pp. 517-522, 2018.
[5] Deepu,Kumar and Divya Gupta, "Review On Optical Character Recognition for Off-line
Devanagari Handwritten Characters & Challenges", an international journal on computer science
engineering and technology,vol.3 (2018).
[6] Wan, X., Zhang, J., Yao, J.G. and Wang, T., (2016). Overview of the NLPCC-ICCPOL 2016
Shared Task: Sports News Generation from Live Webcast Scripts. In Natural Language
Understanding and Intelligent Applications, Springer, Cham, 870-875.
[7] Wang, L., Zhao, W., Jia, R., Li, S. and Liu, J., (2019). Denoising based Sequence-to-Sequence
Pre-training for Text Generation. arXiv preprint arXiv:1908.08206.
[8] Wiseman, S., Shieber, S.M. and Rush, A.M., (2018). Learning neural templates for text
generation. arXiv preprint arXiv:1808.10122.
[9] Yadav, D., Sánchez-Cuadrado, S. and Morato, J., (2013). Optical character recognition for Hindi
language using a neural-network approach. JIPS, 9(1): 117-140.
[10] Zhang, H., Liu, J. and Chen, T., (2019), February. Scene Text Detection with Inception Text
Proposal Generation Module. In Proceedings of the 2019 11th International Conference on
Machine Learning and Computing, 456-460.
[11] Zhang, J., Guo, M. and Fan, J., (2019). A novel CNN structure for fine-grained classification of
Chinese calligraphy styles. International Journal on Document Analysis and Recognition
(IJDAR), 22(2): 177-188.
[12] Zinjore, R.S. and Ramteke, R.J., (2016). Identification of Devanagari Script from Bilingual
Printed Text Documents. In Proceedings of the International Conference on Recent Cognizance
in Wireless Communication & Image Processing, Springer, New Delhi, 451-462.
[13] Zouari, R., Boubaker, H. and Kherallah, M., (2019). Multi-language online handwriting
recognition based on beta-elliptic model and hybrid TDNN-SVM classifier. Multimedia Tools
and Applications, 78(9): 12103-12123.
[14] Framework for Privacy Preserving Classification in Data Mining ,DYK Sharma, GM Sharif,
Journal of Emerging Technologies and Innovative Research 5 (9), 178-183
[15] Performance Evaluation of Delay Tolerant Networks Routing Protocols Under Varying Time of
Live, VK Samyal, DYK Sharma International Journal of Advance Research in Computer
Science (IJARCS) 8
[16] Li-Fi the Most Recent Innovation in Wireless Communication S Saini, DYK Sharma
International Journal of Advanced Research in Computer Science and Software.
[17] Web Page Classification on News Feeds Using Hybrid Technique for Extraction AD Patel, DYK
Sharma Information & Communication Technology for intelligent system 107 (6), 399-405.
[18] Impact of buffer size on different drop policies (DLR, MOFO and E-Drop) for MaxProp Routing
Protocol in DTN VK Samyal, DYK Sharma International Journal of Research in Applied
Science & Engineering
[19] A Pragmatic Evaluation of Stress and Performance Testing Technologies for Web Based
Applications P sonkari, DYK Sharma IEEE, Amity International Conference on Artificial
Intelligence (AICAI), 399-403
[20] Impact of Network Load and node Mobility on the performance of Proactive, Reactive and
Hybrid routing protocols of MANET V Singla, YK Sharma International Journal of Advanced
Research in Computer Science 8

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
328
 International Journal of Advanced Science and Technology
Vol. 29, No. 05, (2020), pp. 322-329

[21] Analysis of Selfish Node Behavior in Delay Tolerant Networks Routing Protocols VK Samyal,
DYK Sharma International Journal of Innovative Research in Science and Engineering
[22] Significance Study Of User Web Access Records Mining For Business Intelligence, Ad Vyas,
Yk Sharma Indian Journal Of Applied Research
[23] A Comprehensive Study On Classification Of Automated Categorization Of Web Sites: A
Proposed Methodology Da Vyas, Dyk Sharma Indian Journal Of Applied Research (Ijar)

ISSN: 2005-4238 IJAST

Copyright ⓒ 2020 SERSC
329