Transfer learning based handwritten character recognition of tamil script using inception-V3 Model

Abstract

The paper describes the excellent method to get first-rate accuracy and performance in the discipline of Tamil character recognition in a handwritten mode. However, the subject is still at a nascent stage and grossly lacks adequate accuracy in the Tamil language, even though several studies have been conducted within the discipline of handwritten character recognition. This paper draws the attention to the offline handwritten recognition for the Tamil language using the Inception-v3 based transfer learning method. The proposed work is conducted on the readily available HP Tamil handwritten character offline dataset (Hewlett-Packard Lab: hpl-tamil-iso-char-offline-1.0.). It reveals that with the suitable arrangement of transfer learning approach with Inception-v3, the pre-trained model can achieve the recognition accuracy of 93.1%, overtaking the former deep learning designs. The achieved accuracy is due to the use of a pre-trained version with transfer learning that regularly hastens the method of the training process on a new task. Overall, this results in higher accuracy and a more capable version.

Full text

Uncorrected Author Proof
Journal of Intelligent & Fuzzy Systems xx (20xx) x–xx
DOI:10.3233/JIFS-212378
IOS Press
1
Transfer learning based handwritten
character recognition of tamil script using
inception-V3 Model
1
2
3
R. Gayathri∗and R. Babitha Lincy4
Department of ECE, Sri Venkateswara College of Engineering, Sriperumbudur, India
5
Abstract. The paper describes the excellent method to get first-rate accuracy and performance in the discipline of Tamil
character recognition in a handwritten mode. However, the subject is still at a nascent stage and grossly lacks adequate
accuracy in the Tamil language, even though several studies have been conducted within the discipline of handwritten
character recognition. This paper draws the attention to the offline handwritten recognition for the Tamil language using the
Inception-v3 based transfer learning method. The proposed work is conducted on the readily available HP Tamil handwritten
character offline dataset (Hewlett-Packard Lab: hpl-tamil-iso-char-offline-1.0.). It reveals that with the suitable arrangement
of transfer learning approach with Inception-v3, the pre-trained model can achieve the recognition accuracy of 93.1%,
overtaking the former deep learning designs. The achieved accuracy is due to the use of a pre-trained version with transfer
learning that regularly hastens the method of the training process on a new task. Overall, this results in higher accuracy and
a more capable version.
6
7
8
9
10
11
12
13
14
15
Keywords: Handwritten character recognition, inception-v3, tamil language, transfer learning16
1. Introduction
17
Understanding the handwritten characters or typed18
files is straightforward for human beings because we19
have the potential to learn. An equal potential may be20
precipitated to the machines additionally, by using the21
procedure of Artificial Intelligence, Neural Network,22
Machine Learning, and deep learning algorithms. The
23
discipline which offers this technology is referred to
24
as OCR, which stands for “Optical Character Recog-
25
nition (OCR).” The OCR is the method of changing
26
the image into an editable digital character [1]. There27
are various applications for categorizing handwritten28
characters. It can be utilized to digitize the ancient29
∗Corresponding author. R. Gayathri, Associate Professor
Department of ECE, Sri Venkateswara College of Engineering,
Sriperumbudur, India. E-mail: rgayathri@svce.ac.in.
records in healing centers or workplaces. Moreover, 30
it can be considered in the post office for sorting let- 31
ters for different regions. The application of OCR can 32
reduce the time utilized in entering the information 33
and the storing space capacity required by the reports. 34
In other words, it can be recovered quickly. 35
By using the OCR in the field of banking, law, 36
and so on, numerous critical and important archives 37
can be prepared promptly without human media- 38
tion. The OCR is categorized into two types, (a) 39
handwritten character recognition and (b) printed 40
character recognition. Further, based on acquiring 41
the input documents, handwritten OCR is categorized 42
into Offline and Online recognition systems [2]. The 43
offline mode deals with recognizing the pre-written 44
report obtained through diverse input methods. But 45
in the online recognizing order, the writing is diag- 46
nosed the moment it is written. The device used for the 47
ISSN 1064-1246/$35.00 © 2021 – IOS Press. All rights reserved.

Uncorrected Author Proof
2R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition
online machine is an Electric pen where it is used for48
writing the letters or words on the tool, known as the
49
digitizer, based on the pen movement. The handwrit-50
ten recognition tool needs high identification ability51
than the online and printed recognition due to the var-52
ious writing styles of the people. Many a time, even
53
the handwriting of the same men or women does not54
match at different points in time.55
In the deep learning process, the “Convolutional56
Neural Network (CNN)” has topped in giving the best
57
results in some of the unique complications such as58
detection and prediction among numerous fields such59
as pattern recognition, character recognition, addi-
60
tionally object detection also [3]. By using some of61
the recent deep learning models like VGG, ResNet,
62
and Inception, the classification and prediction task63
was accomplished with high accuracy. Due to the64
deep and complex structure of these models, it is
65
tough to train, and need images in bulk to train without66
over-fitting. Some researchers have introduced sub-67
stantial data augmentation [4] also to save the model68
from over-fitting for small dataset problems. By using69
a novel approach, called transfer learning with high70
complex model [5], it became possible to enhance71
the performance of the classifier on the small dataset.
72
Presently, it is the best known standard approach in
73
deep learning. Using this approach, one can utilize74
the pre-trained models of the first task as the opening
75
point for the same model on the second task. Transfer76
learning permits us to use learned knowledge of the77
first task and puts them on to newer and any related
78
second task. It means low-level features and some79
of the high-level features are being shared across the
80
tasks, which will permit knowledge transfer between
81
tasks. The proposed method in this research is the82
retraining process of the Inception-v3, using transfer83
learning method for “Tamil Handwritten Character84
Recognition (THCR),” as shown in Fig. 1.85
THCR is defined as the capability of recogniz-86
ing the exact character from digitized print and
87
handwritten Tamil documents with a high degree of88
recognition accuracy for a variety of Tamil digital
89
inputs. Tamil is the longest- surviving and the oldest90
Fig. 1. Transfer learning approach for Tamil Handwritten Charac-
ter Recognition.
language, one of the Dravidian languages primarily 91
vocalized by the Tamil people of India, Sri Lanka, 92
Singapore, and Malaysia. The Tamil alphabets con- 93
sist of 12 vowels, one Aayudham, 18 consonants, 94
and 216 compound characters. Hence Tamil has a 95
total of 247 characters. About 6 Grantham characters 96
are also present in the Tamil language [6]. THCR is 97
more difficult than the printed Tamil character recog- 98
nition due to curves in character, sliding characters, 99
and its various strokes and holes. However, many of 100
the researchers take these as a challenging task, and 101
consequently, reasonable accuracy, speed, and per- 102
formance have not been obtained. So the idea behind 103
this work is to identify and analyze a Tamil handwrit- 104
ten document image, using the Inception-v3 model 105
with the transfer learning approach. This work is car- 106
ried on in the HP lab Offline THCR Dataset, which 107
includes 156 classes. A sample of print and handwrit- 108
ten images of Tamil language characters are shown 109
in Fig. 2. 110
2. Related works 111
Recognition and classification are significant prob- 112
lems in deep learning. Many of the deep learning 113
researchers gave some new signature models for 114
recognition and classification. Many scientists have 115
studied many machine learning methods such as 116
“Support Vector Machine (SVM)”, ANN, HMM 117
[7], HLP, and deep learning model algorithms like 118
CNN [8]. The researchers used these methods to 119
solve the OCR problems for many languages like 120
Japanese, Chinese, English, Tamil, Devanagari, Tel- 121
ugu, Gujarati, and so on [9]. Similarly, hybrid models 122
by combining deep learning with machine learning 123
were also introduced, like CNN-SVM [10]. 124
Ning Bi and others [11] identified the Chinese 125
handwritten character with the help of GoogLeNet, 126
which is one of the successful deep models in CNN. 127
This work is carried on the database CASIA-HWDB 128
and HCL2000. This experiment exhibited that the 129
GoogLeNet can provide superior results for the 130
Handwritten Chinese symbol identification than the 131
previous deep models. In this research, GoogLeNet 132
model uses numerous inception stages to construct 133
an efficient deep network, which will find the opti- 134
mal local construction. Saeeda Naz [12] and others 135
introduced a new hybrid model, which was the com- 136
bination of CNN and “Recursive Neural Network 137
(RNN)” for Urdu Nastaliq recognition. In this hybrid 138
model, low-end features like edges and shapes are 139

Uncorrected Author Proof
R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition 3
Fig. 2. Sample image of Tamil language: (a) Printed Tamil Character and (b) Handwritten Tamil Character.
extracted by the CNN then forwarded to the RNN140
architecture to recognize the character. This work is141
verified on the openly existing “Urdu Printed Text-
142
line (UPTI)” dataset by using the proposed hybrid143
grouping of the CNN and the RNN for 44-classes,
144
which achieved the superior results on the UPTI145
dataset.146
Adnan Taufique and others [13] recognized the147
Bangla character by using CNN with the inception148
module. The dataset includes 85,000 images used149
for training and 3000 images used for testing. The
150
planned method showed competitive performance
151
with the existing methods based on the test set accu-152
racy for the dataset. The accuracy of this work is
153
better than other models. Many investigators clearly
154
explain the inception models for their studies. In the
155
next stage of an advanced concept, transfer learning
156
is catching the attention to record the progress of the157
performance of traditional architectures by different158
researchers.159
Le Zhang and others [14] uses the transfer learn-160
ing technique to identify the numeral digits with161
the help of multi-layer perceptron and CNN sys-
162
tems. The authors select the five scripts such as
163
Tibetan, Telugu, Arabic, Devanagari, and Bangala.
164
The researchers presented that the transfer learning
165
model is the best model based on less training time,
166
but this model somewhat decreases the accuracy rate.167
Mohammed Aarif and others [15] select the trans-168
fer learning approach with AlexNet and GoogleNet169
deep learning models to identify the Urdu characters.170
This research work also especially concentrated on
171
the different fonts and size characters also. AlexNet
172
and GoogleNet generate the recognition rate as 96.3%
173
and 94.7%, respectively. Satyasangram Sahoo and
174
others [16] suggested transfer learning technique with 175
CNN architecture to get the outstanding performance 176
for Telugu and Kannada letters. Usually, Telugu and 177
Kannada letters are almost in similar shape. 178
Chunmian Lin and others [5] have introduced a 179
new model for traffic sign identification and classi- 180
fication based on transfer learning, which is useful 181
for road infrastructure and driver assistant systems. 182
Using the Inception-v3 model significantly reduced 183
the training data size and computation expense. 184
In this project, Belgium Traffic Sign Dataset was 185
chosen and was augmented through the data pre- 186
processing technique. In this model, the features from 187
different layers using convolution and pooling pro- 188
cesses were compared and analysed. As a result, the 189
transfer learning-based inception model cyclically 190
retrained numerous times with fine-tuning parame- 191
ters at different learning rates. Excellent reliability 192
and repeatability were also observed based on statis- 193
tical analysis. The result of this work showed that the 194
transfer learning model could achieve the best recog- 195
nition performance in traffic sign recognition. Jyotsna 196
Bankar and others [17] proposed the system based on 197
the Inception-v3 design of TensorFlow platform, in 198
which they used the transfer learning technology to 199
train the animal classification model on a mammal’s 200
dataset. The classification accuracy rate of the model 201
is approximately 95% on a given dataset, which is 202
higher than the other methods available for classifi- 203
cation. Nagender Aneja and others [18] used the same 204
transfer learning with the Inception-v3 technique to 205
recognize Devanagari handwritten character. Results 206
of this work depicted that the proposed model can per- 207
form better in terms of accuracy per average epoch 208
time. 209

Uncorrected Author Proof
4R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition
Fig. 3. General architecture of THCR using Inception-V3.
Much research was undertaken for the Tamil lan-210
guage also. Kavitha and Srimathi C. [19] used the
211
CNN model to recognize handwritten Tamil charac-212
ters in the offline mode. They used HP Labs India213
dataset to understand the character. They trained the
214
model from scratch, which produced the state-of-art
215
result in Tamil character recognition. S. Kowsalya
216
and P. S. Periasamy [20] introduced a new model
217
called the Neural Network with Elephant Herding
218
Optimization to recognize the handwritten Tamil219
character. Shanthi and Duraiswamy [21] described220
a model for identifying handwritten Tamil characters221
by SVM, offline. They used their dataset to recognize222
the character. Various pre-processing operations were
223
performed on the scanned image. The features were224
extracted for 64 different zones, and those extracted
225
features trained the SVM. This model achieved good
226
recognition accuracy on the Tamil symbol database.
227
3. THCR recognition by inception V3 with
228
transfer learning229
From the above literature, it can be concluded that230
still, THCR is in its very early stage. So THCR has
231
suggested a novel Inception-v3 model with trans-232
fer learning technique to enhance the recognition233
rate. The general architecture of the proposed THCR234
is shown in Fig. 3. Before entry into the architec- 235
ture model, dataset loading, preparing dataset, and 236
encoding dataset class labels into numeric values 237
are essential. The HP lab THCR dataset is loaded 238
with some underlying dependencies such as resizing, 239
binarisation and noise removal into the model. Data 240
augmentation step is added by the Image Data Gen- 241
erator framework of Keras, to reduce the over-fitting 242
problem. The dataset images get altered by this data 243
augmentation step with some of the image renova- 244
tion processes such as shearing, rotation, zooming, 245
and translation. Due to these random transformations, 246
the model does not get the same images each time. 247
Then the HP lab THCR dataset is passed through 248
the dataset split module, where the dataset images 249
are split into training, validation, and testing of the 250
set images. The planned technique in this study con- 251
sisted of three phases, namely, pre-trained model, 252
retraining process with transfer learning technique, 253
and modified recognition portion. 254
3.1. THCR by pre-trained Inception-v3 255
A pre-trained model is a saved architecture, which 256
was previously trained on a massive dataset. Pre- 257
trained models are a brilliant source of researchers 258
to learn an algorithm or try out an existing frame- 259
work for future problems. Due to time boundaries or 260

Uncorrected Author Proof
R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition 5
computational limits, it is not always possible to build261
a model from scratch. Pre-trained representations are
262
introduced to resolve these issues. Pre-trained model263
is a standard model to either develop the performance264
of the existing model or test the new model against it.265
The perception behind the pre-trained model for clas-
266
sification problem is that if a model was trained on267
a massive and universal dataset, this model would be268
successfully considered as a standard model of the269
optical world. In general, for classification applica-
270
tions, specific standard models like VGG, ResNet-50,271
XCeption and Inception-v3 models are presented,272
which were trained on standard ImageNet dataset.
273
Hence, it would be beneficial for researchers to use274
these models. The ImageNet dataset covered 14 mil-
275
lion images of 1000 groups. The THCR is proposed276
based on the Inception-v3 model as a pre-trained277
model, which is trained on ImageNet weight. The
278
Inception-v3 is an extensively used image classifi-279
cation model that has been developed by concluding280
many ideas of multiple researchers over the centuries.281
This is established from the original research paper282
[22] by Szegedy and others. This model is the third283
version of the series made by Google Deep Learning284
Convolutional Architectures. The Inception-v3 got
285
the first runner up on ImageNet Large Visual Recog-
286
nition Challenge, which attained 21.2% top-1 and287
5.6% top-5 error rate. Visualization of the Inception-
288
v3model architecture is presented in Fig. 4. The289
model is made up of many building blocks, including290
convolutions, average pooling, max pooling, concate-
291
nation, dropouts, and fully-connected layers. Batch292
normalization is used comprehensively, all over the
293
model, and applied to activation inputs. The image
294
label is computed by the probability value, which is295
calculated by the Softmax classifier.296
3.2. Transfer learning concept for THCR 297
The corresponding teams have publically shared 298
a lot of their great deep learning designs. Millions 299
of parameters, feature maps, and weights of these 300
designs were saved as customers to help new users. 301
That publically shared model is called a pre-trained 302
model, which is processed on a particular problem in 303
a stable mode. Due to deep learning believes in shar- 304
ing, by using these learned feature maps, millions 305
of parameters and weights can train large models on 306
the big dataset without having to start from scratch, 307
which is defined as transfer learning. Keras is the 308
famous deep learning Python library, which offers 309
an interface to use and download these pre-trained 310
models. But one essential requirement of transfer 311
learning is the presented pre-trained design, which 312
has been proven to be a well-performing model on the 313
source tasks. Transfer learning model with Inception- 314
v3 architecture for THCR is displayed in Fig. 5. In 315
this work, the Imagenet classification with Inception- 316
v3 model is considered as the source task, and THCR 317
is the target task. The trained features of the source 318
task are transferred to the new THCR task. The target 319
HP Tamil dataset is small and similar to the source 320
task, Imagenet dataset. If the entire feature map- 321
files of Imagenet are transferred to the new THCR 322
model, over-fitting will occur. To avoid this problem, 323
train only the classification part. Due to the require- 324
ments of only the high-level range features, freeze 325
all the Inception-v3 layers and remove the classifi- 326
cation layers of the source task. After removing the 327
old classification layers, add the new classification 328
layers on top of the model depending on the target 329
task. Now the model trains only the newly added 330
classifier layers. By using this model, processing 331
Fig. 4. Inception-v3 model for classification.

Uncorrected Author Proof
6R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition
Fig. 5. Transfer learning approach with Inception-v3 for THCR.
time also gets reduced, which is one of the most top332
advantages.
333
3.3. Modified version of classification layer to334
recognize tamil character
335
In this proposed work the Inception-v3 is used as
336
a feature extractor by freezing all inception blocks
337
for THCR. Freezing of the inception blocks is proper
338
because, in transfer learning model, there is no need339
for weight updating in base layers during model train-
340
ing [18]. The Inception-v3 pre-trained model learned341
a definite hierarchy of features from Imagenet dataset.342
Therefore, the learned model with a good represen-
343
tation of features from a million images in 1,000
344
different categories can perform as a suitable feature345
extractor to input image of new target classification346
problems. Even though the target images might not347
even exist in the ImageNet dataset or might be of348
entirely different categories, the model can extract349
relevant features. During transfer learning, there is350
no necessity for fully-connected layers, since the pro-351
posed model uses their fully-connected dense layers352
to classify Tamil characters. Thus the Inception-v3353
model is improved by adding fully-connected and354
modified layers. The trained feature extracted layers
355
of the Inception-v3 from Imagenet dataset, get flat-
356
tened, and serve the dense layer of the fully-connected
357
modified deep classifier. The dense layer is one of the 358
actual network layers, where all outcomes of the pre- 359
vious layer are feeds to the following layer in that 360
model. The dropout of 0.3 is added, to enable reg- 361
ularization. Fundamentally, dropout is a dominant 362
technique of regularising in deep neural nets [23]. 363
The modified version of the fully-connected layer of 364
THCR is shown in Fig. 6. 365
4. Experimental setup 366
The key indication of this project recognizes the 367
handwritten Tamil characters. The weights and biases 368
of the Imagenet dataset-based Inception-v3 model 369
are used as the re-used model for Tamil character 370
recognition training model. In this study, those pre- 371
trained model, act as a feature extraction part of the 372
new model, as mentioned earlier. The inception layers 373
are frozen to update weight, or else the key point of 374
transfer learning cannot be confirmed. The top layers 375
of the proposed model are modified depending on the 376
THCR application. Due to the inception blocks being 377
in the frozen stage, the training process is carried only 378
on modified fully-connected layers. 379
The experimental setup and system specification 380
for THCR are given in Table 1. The model using 381
the Adam optimizer for modified layer parameter 382

Uncorrected Author Proof
R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition 7
Fig. 6. Modified version of the fully-connected layer of THCR.
Table 1
Experimental setup and system specification for THCR
Specifications Parameter Value
Model LENOVO IDEAPAD 330
System Specifications Operating System Windows 10
Processor Intel core i5
RAM 8GB
Graphics Card NVIDIA GEFORCE
Parameter Specifications Dataset split ratio 80 :20
Batch size 32
Optimizer ADAM
Epochs 500
Learning rate 0.001
Loss function Categorical cross-entropy
Source, Target datasets ImageNET, HP Tamil dataset
Pre-trained model Inception-v3
Classifier Softmax
(a) (b)
Fig. 7. Performance analysis of the THCR using Inception with transfer learning : (a) Accuracy and (b) Loss.

Uncorrected Author Proof
8R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition
(a)
(b)
(c)
(d) (e)
Fig. 8. Performance analysis of the THCR (a) simple CNN (b) VGG-16 (c) VGG-19 (d) CNN with modified Lion optimizer (e) CNN with
modified sea lion optimizer.

Uncorrected Author Proof
R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition 9
Table 2
Test accuracy for first 20 classes from the dataset
Class Class Number of Accuracy Class Class Number of Accuracy
Name samples Name samples
0 568 92.24 10 563 90.45
1560 91.89 11 517 90.99
2561 94.01 12 497 94.51
3544 94.21 13 548 94.56
4556 89.84 14 542 93.91
5565 91.78 15 565 93.02
6551 91.11 16 564 92.42
7561 92.39 17 553 90.56
8551 93.57 18 528 91.83
9535 90.09 19 556 92.61
updates. Also, the proposed system planned to use383
the Categorical cross-entropy loss as the error func-
384
tion. Based on this error function value, the parameter385
values are modified. To train the model, the THCR386
system is using the 0.001 value as the learning rate387
and selects 500 as the Epoch value and 32 as the Batch
388
Size.389
5. Results and discussions390
From Fig. 7, it can be concluded that the proposed
391
Inception-v3 model is the best model for THCR.392
The gap between the training and validation accu-393
racy Fig. 7(a) shows the model is the best without394
over-fitting. The loss graph Fig. 7(b) indicates the
395
system learning with proper parameters. In this inves-396
tigational arrangement, the Tamil character dataset
397
includes the 155 classes, where all classes have more398
algorithm controls over the power of adaptive learn-399
ing rates methods to find individual learning rates for400
each parameter.401
The Loss function performs as monitors to the opti-402
mizer if it is moving in the right way to reach the403
global minimum. In the proposed work, categorical404
cross-entropy is used as a loss function to optimize
405
the parameter values of the projected model. The loss
406
value suggests how a model performs at every end
407
of the iteration of the training process. For compar-408
ison purpose, the THCR system using simple CNN,
409
VGG-16, VGG-19, CNN with modified lion opti-410
mizer model and CNN with modified sea lion model411
is shown in Fig. 8.412
Accuracy is used to find the performance met-413
rics of the proposed algorithm of the THCR model.414
The training process results for THCR are shown415
in Figs. 7, 8 and Table 1. The baseline architec-416
ture of the proposed model gave 93.1% test accuracy417
Fig. 9. Accuracy comparisons between 20 different classes of
Tamil language.
for Tamil handwritten recognition, which produced 418
Training accuracy of 95.45% and 91.82% as valida- 419
tion accuracy. The efficiency is further improved by 420
introducing the fine-tuning methods, where all the 421
inception blocks were not frozen. 422
Table 2 displays the testing accuracy for 20 indi- 423
vidual classes from the 155 classes of HP dataset, and 424
Fig. 9 shows the accuracy comparison between these 425
20 different classes of Tamil language. From Table 1 426
and Fig. 9, the accuracy for the dataset increases 427
when increasing the depth of the model architecture, 428
which also increases when introducing transfer learn- 429
ing model with less processing time. Some of the test 430
images recognition is shown in Fig. 10. 431
The Inception-v3 model trains the planned THCR 432
with modified fully-connected layers by selected 433
hybrid parameters. The trained model file is saved. 434
The real-time input image is considered as a query 435
image, which is passed through the pre-processing 436
process such as resizing, noise removal, slant correc- 437
tion and slope removal. Then the pre-processed query 438
image is given into the saved proposed model file. 439
Based on that saved model file, labels are assigned to 440
the output in the form of the class of the given query 441

Uncorrected Author Proof
10 R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition
Fig. 10. Test image recognition.
(a) (b)
(c) (d)
Fig. 11. Query image output (a) Input Image (b) Preprocessed
Image (c) Segmented Image (d) Recognized output.
image. Some of the real-time output of input queries442
is shown in Fig. 11.443
The THCR system is trained with different deep444
learning architectures based on heuristic-based and445
meta-heuristic based optimizer. Based on the exper-446
iments, the comparison table for Tamil handwritten 447
character recognition is shown in Table 3. 448
The best model is based on accuracy and also learn- 449
ing speed. When considering the Inception-v3 model 450
without transfer learning approach, the model takes 451
a long time to train, since the Inception-v3 is a very 452
deep model. Because the simple CNN model without 453
transfer learning techniques takes nearly 2217 s per 454
epoch. When considering 30 epochs, it is a long time 455
process. At the same time the Inception-V3 model 456
with transfer learning technique takes only 774us per 457
epoch, even though it is a very deep model. Based on 458
the speed and accuracy rate, the proposed model is 459
the best model. The comparison work based on the 460
THCR system with different existing work is shown 461
in Table 4. From the Figs. 7, 8, Tables 3 and 4, it 462
can be concluded that the proposed transfer learning- 463
based THCR system with the Inception-v3 model is 464

Uncorrected Author Proof
R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition 11
Table 3
Comparison between different models for THCR system
Model Optimizer Optimizer Transfer Accuracy (%) Training time
type learning per Epoch
CNN Adam Heuristic No 83.1 2217s
CNN Modified Lion [24] Meta- heuristic No 84 2626s
CNN Modified Sea Lion Meta- heuristic No 86 3219s
VGG-16 Adam Heuristic No 85.3 5425s
VGG-19 Adam Heuristic No 87.2 6658s
Inception-v3
(Proposed) Adam Heuristic Yes 93.1 774us
Table 4
Comparison between the various existing works
Existing work Dataset Method Accuracy %
Kavitha et al. [19] HP Labs India CNN 95.1%
Sornam and Vishnu IWFHR-10 PCA and CNN 85.05%
priya [25]
Kowsalya and Own dataset ANN and EHO 93%
Periasamy [26]
Raj and Abirami [6] HP-India 2013 hierarchical SVM 90.3%
Bhattacharya HPLabs dataset Clustering and group 92.77%
et al. [27] wise classification %
Shanthi and Own dataset SVM 82%
Duraiswamy [28]
Vijayaraghavan HPLabs dataset CNN 99 %(only 35 labels)
and Sra [29]
Proposed work HPLabs dataset Transfer learning 93.1% (155 labels)
with Inception-V3
the best model in terms of accuracy and less learning465
period.466
6. Conclusion467
Transfer learning allows retraining only the top468
layer of a proposed model, causing a significant469
reduction in both training time and also the size of470
the dataset. A prominent model that can be used471
for transfer learning is the Inception-v3, to recog-
472
nize the handwritten Tamil characters. As expressed,
473
this model was initially prepared with the assis-
474
tance of over a million pictures from 1,000 labels
475
on some extremely incredible models. Being able
476
to retrain the final layer signified that the model477
could maintain the knowledge that it had learned478
during its original training, and could apply it to479
a smaller HP Tamil handwritten character dataset.
480
The result is with highly accurate classifications,481
without the need for extensive training and computa-
482
tional power. The proposed THCR system achieved483
93.1% testing accuracy, which is higher than THCR484
using the CNN model. The main identification errors485
were due to abnormal writing and ambiguity among486
similar shaped characters. Future work can include487
more robust extracting features for the classifier to 488
achieve better discrimination power by performing 489
a fine-tuning process in the Inception layers. The 490
recognition accuracy of the individual characters can 491
be additionally enhanced by combining the hybrid 492
models. And also, the future work will consider the 493
special segmentation technique for the identification 494
of abnormal writing and among similar shaped char- 495
acters. 496
References 497
[1] D. Bouchain, Character Recognition Using Convolutional 498
Neural Networks, 6(2006), 1903–1907. 499
[2] M. Agarwal, V.T. Shalika and P. Gupta, Handwritten char- 500
acter recognition using neural network and tensor flow, 501
Int. J. Innov. Technol. Explor. Eng., 8(6) Special Issue 4, 502
1445–1448. (2019), doi: 10.35940/ijitee.F1294.0486S419 503
[3] M. Soomro, M.A. Farooq and R.H. Raza, Performance eval- 504
uation of advanced deep learning architectures for offline 505
handwritten character recognition, Proc. - 2017 Int. Conf. 506
Front. Inf. Technol. FIT (2017), 2017 (2017), 362–367, doi: 507
10.1109/FIT.2017.00071 508
[4] D.X. Xue, R. Zhang, H. Feng and Y.L. Wang, CNN-SVM 509
for Microvascular Morphological Type Recognition with 510
Data Augmentation, J Med Biol Eng 36(6) (2016), 755–764, 511
doi:10.1007/s40846-016-0182-4 512

Uncorrected Author Proof
12 R. Gayathri and R.B. Lincy / Inception-V3 model based Tamil script recognition
[5] C. Lin, L. Li, W. Luo, K.C.P. Wang and J. Guo, Transfer
513
learning based traffic sign recognition using inception-514
v3 model, Period. Polytech. Transp. Eng. 47(3) (2019),515
242–250. doi:10.3311/PPtr.11480516
[6] M. Antony Robert Raj and S. Abirami, Structural517
representation-based offline Tamil handwritten charac-
518
ter recognition, Soft Comput 24(2) (2020), 1447–1472,519
doi:10.1007/s00500-019-03978-5520
[7] T. Bluche, H. Ney and C. Kermorvant, Tandem Hmm521
With Convolutional Neural Network For Handwritten Word
522
Recognition, Human Language Technology and Pattern523
Recognition (2013), 2390–2394.524
[8] S. Joseph James, C. Lakshmi, P. Uday Kiran and Parthiban,525
An efficient offline hand written character recognition using
526
CNN and xgboost, Int. J. Innov. Technol. Explor. Eng. 8(6)527
(2019), 115–118.528
[9] M. Elleuch, N. Tagougui and M. Kherallah, A novel archi-529
tecture of CNN based on SVM classifier for recognizing
530
Arabic handwritten script, Int. J. Intell. Syst. Technol. Appl.531
15(4) (2016), 323–340, doi:10.1504/IJISTA.2016.080103
532
[10] X.X. Niu and C.Y. Suen, A novel hybrid CNN-533
SVM classifier for recognizing handwritten digits,534
Pattern Recognit. 45(4) (2012), 1318–1325,535
doi:10.1016/j.patcog.2011.09.021536
[11] G. Katiyar, Off-Line Handwritten Character Recog-
537
nition System Using Support Vector Machine,
538
Am. J. Neural Networks Appl. 3(2) (2017), 22.
539
doi:10.11648/j.ajnna.20170302.12
540
[12] S. Naz, et al., Urdu Nastaliq recognition using
541
convolutional–recursive deep learning, Neurocomputing542
243 (2017), 80–87. doi:10.1016/j.neucom.2017.02.081543
[13] M.A. Uddin, Handwritten Bangla Character Recognition544
Using Artificial Neural Network, IOSR J. Comput. Eng.545
16(3) (2014), 33–38. doi:10.9790/0661-16333338
546
[14] L. Zhang, A Transfer Learning Approach for Handwrit-547
ten Numeral Digit Recognition, ICSIM’20: Proceedings of
548
the 3rd International Conference on Software Engineer-
549
ing and Information Management, pages 140–145. doi:550
https://doi.org/10.1145/3378936.3378970551
[15] K.O. Mohammed Aarif and S. Poruran, OCR-Nets:552
Variants of Pre-trained CNN for Urdu Handwrit-
553
ten Character Recognition via Transfer Learning,
554
Procedia Computer Science 171 2294–2301. DOI:555
https://doi.org/10.1016/j.procs.2020.04.248
556
[16] S. Sahoo, B. Prem Kumar and R. Lakshmi, Offline
557
handwritten character classification of the same scrip-558
tural family languages by using transfer learning559
techniques, 3rd International Conference on Emerging
560
Technologies in Computer Engineering: Machine Learn-
561
ing and Internet of Things (ICETCE-2020), (2020). DOI:
562
10.1109/ICETCE48199.2020.9091744563
[17] J. Bankar and N.R. Gavai, Convolutional Neural Network564
Based Inception V3 Model for Animal Classification, Int.565
J. Adv. Res. Comput. Commun. Eng. 7(5) (2018), 142–146.566
doi:10.17148/IJARCCE.2018.7529567
[18] N. Aneja and S. Aneja, Transfer Learning using CNN for 568
Handwritten Devanagari Character Recognition, 1st IEEE 569
Int. Conf. Adv. Inf. Technol. ICAIT 2019 - Proc., (2019), 570
293–296. doi:10.1109/ICAIT47043.2019.8987286 571
[19] B.R. Kavitha and C. Srimathi, Benchmarking on offline 572
Handwritten Tamil Character Recognition using convolu- 573
tional neural networks, J. King Saud Univ. - Comput. Inf. 574
Sci., (2019), doi:10.1016/j.jksuci.2019.06.004 575
[20] S. Kowsalyaand P.S. Periasamy,Recognition of Tamil hand- 576
written character using modified neural network with aid of 577
elephant herding optimization, Multimed. Tools Appl. 78 578
(2019), 25043–25061. doi:10.1007/s11042-019-7624-2 579
[21] N. Shanthi and K. Duraiswamy, A novel SVM-based 580
handwritten Tamil character recognition system, Pattern 581
Anal. Appl. 13(2) (2010), 173–180, doi:10.1007/s10044- 582
009-0147-0 583
[22] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens and Z. Wojna, 584
Rethinking the Inception Architecture for Computer Vision, 585
Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recog- 586
nit., 2016 (2016), 2818–2826, doi:10.1109/CVPR.2016.308 587
[23] M. Elleuch, R. Maalej and M. Kherallah, A New 588
design based-SVM of the CNN classifier architecture 589
with dropout for offline Arabic handwritten recogni- 590
tion, Procedia Comput. Sci. 80 (2016), 1712–1723. 591
doi:10.1016/j.procs.2016.05.512 592
[24] R.B. Lincy and R. Gayathri, Optimally configured convo- 593
lutional neural network for Tamil Handwritten Character 594
Recognition by improved lion optimization model, Mul- 595
timed Tools Appl (2020). https://doi.org/10.1007/s11042- 596
020-09771-z 597
[25] M. Sornam and C. Vishnu Priya, Deep Convolutional Neu- 598
ral Network for Handwritten Tamil Character Recognition 599
Using Principal Component Analysis, Smart and Innovative 600
Trendsin Next Generation Computing Technologies (2018), 601
778–787. 602
[26] S. Kowsalyaand P.S. Periasamy,Recognition of Tamil hand- 603
written character using modified neural network with aid of 604
elephant herding optimization, Multimedia Toolsand Appli- 605
cations 78(17) (2019), 25043–25061. 606
[27] U. Bhattacharya, S.K. Ghosh and S. Parui, A two stage 607
recognition scheme for handwritten Tamil characters, 608
ICDAR 2007. Ninth International Conferenceon Document 609
Analysis and Recognition, IEEE (2007), 511–515. 610
[28] N. Shanthi and K. Duraiswamy, A novel SVM-based hand- 611
written Tamil character recognition system, Pattern Anal. 612
Appl. 13(2) (2010), 173–180. 613
[29] Vijayaraghavan and Sra,” Handwritten tamil recognition 614
using a convolutional neural network”, 2014.m 615