Writer Identification via CNN Features and SVM

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 04 | Apr 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 3444
Writer Identification via CNN Features and SVM
Mahmoud rahhal
Aleepo, Syria, Computer Engineering Department, College of Electrical and Electronic Engineering, University of
Aleppo
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract –Writer recognition is the process of identifying
the author of a document based on his or her handwriting.
Recent advances in computational engineering, artificial
intelligence, data mining, image processing, pattern
recognition, and machine learning have demonstrated that it
is possible to automate writer identification. This paper
proposes a writer recognition model based on Arabic
handwriting.
This research presents the study and implementation of the
stages of writer identification, starting from data acquisition
by scanned images of handwritten text, and then augmented
the data through study of Maaz et al. [1] that generatealarge
number of texts from the set of texts available within the
database, and then building a convolutional Neural Network
(CNN) Which is usually useful for extracting features
information and then classification the data, but in this
research it is used for feature extraction, finallysupportvector
machine is used for classification.
The experiments in this study were conducted on images of
Arabic handwritten documents from ICFHR2012 dataset of
202 writer, and each writer have 3 text. The proposed method
achieved a classification accuracy of 97.20%.
Key Words: Arabichandwriting,data augmentation,writer
identification, deep learning, convolutional Neural Network,
support vector machine.
1. INTRODUCTION
Identification of persons is mainly through the physiological
characteristics like fingerprints, face, iris, retina, and hand
geometry and the behavioral characteristics like a voice,
signature, and handwriting. Writer identification is a
behavioral approach of handwriting recognition, which
offers with matching unknown handwriting with a database
of samples with recognized authorship.
Writer identification has been a lively discipline of lookup
over the previous few years and it used in manyapplications
as in biometrics, forensics and historic report analysis.
Identifying people involve a fields of Artificial intelligence
(AI), image processing and pattern recognition, and it
contributes in many applications: biometrics, forensics,
security, and legal matters, and financial field etc. [2]. All
methods of writer identification and verification can be
classified into two categories [3]: text-dependent method
and text-independent method. Text-dependentmethods are
methods where the text in test samples must be the same as
in training samples. Scale to extractfeaturesfromtextimage.
Text-independent methods are methods in which the text in
test samples does not have to be the same in training
samples and in which structural methods for extracting
features from a character shape are involved.
Writer identification is classified into on-line writer and off-
line writer identification [3], in the case of on-line writer
identification, features are collected directly from signals
sent from digital devices while writing, and in the off-line
case, and attributes are collected from written text
Handwriting obtained from the scanned image. . It is
considered as more complex than on-line case due to many
dynamic features of handwriting are missing, for instance
pen-pressure, order of strokes, and writing quickness [4].
The task of classification in pattern recognition is to identify
the pattern to a class of the known group of classes, in which
features will be extracted from images of scanned texts
handwritten and trained it by using a classification
algorithm. This study proposed a system for identifying
people through Arabic handwritten text by extracting
features by deep learning, which required data
augmentation. Because of limited number of handwritten
text samples, it is impossible to gather a wide number of
samples from different writers. Thetextisdividedintoparts
and then Gathering all text structures randomly to obtain
several lines, which enables us to generate many texts from
them, and then using a convolutional neural network [1] for
feature extraction , finally using support vector machine
(svm) for classification which proves its ability to achieve
high accuracy in writer identification.This researchincludes
a number of previous studies on handwriting features
extraction, person identification techniques, and some data
augmentation techniques.
2. RELATED WORKS
In this paragraph, the studies will be discussed that have
been done on writeridentificationandthetechniquesusedso
far, with an overview of writer identification systems in
several languages, such as Chinese, English, Arabic ... etc.
Researcher Bangy Li et al in 2009 proposed a method based
on combining the static feature and dynamic feature, and

then the Nearest Neighbor Classifier was used as a
classification methodology [5]. In 2009, researcher Zhu et al
presented a study in this field using the Shape Codebook and
Multi Class SVM Classifier, and their proposed method was
evaluated by collecting images of texts written in eight
languages(Arabic, Chinese, English, Hindi, Japanese, Korean,
Russian and Thai). The Chinese identification rate was much
lower (55.1%) compared to the eight other languages [6]. In
2013, researcherDjeddiandhiscolleaguesproposedawayto
identify the identity of the text writer based on a set of
features extracted throughtheGrayLevelRunLength(GLRL)
matrices method and comparing them with the latestknown
features at the time, and good results were shown in
identifying the writer’sidentity,andK-Nearestwasalsoused.
Neighbors and Support Vector Machines and their
experiments were conducted on handwritten Greek and
English texts of 126 writers and each writer 4 samples,
achieving an interesting performanceinauthoridentification
and verification in a multitext environment [7]. A study
presented by the researcher Alaei et al in 2014 proposing a
method for identifying the writer based on Histogram, using
two different sets of handwritten data written in Canadian
and English languages. Their experiments were conducted
on 228 Canadian-language texts for 57 writers with 92.79%
accuracy for F-measure and 330 texts in English for 55
writers with 26.67%accuracyforF-measure[8]. Thendralet
al conducted a 2015 study of person-identification using
images of Tamil handwriting of 300 writers and each writer
100 samples using a discriminant model to categorize with
features aggregated from handwriting and use it in the
Support Vector Machine, using three kernel function: Linear
The RBF and Polynomial and together with Polynomial and
RBF gave the highest prediction accuracy [9]. In 2015,
researcher Fiel and Sablatnig [10] presented a study on the
databases ICDAR2013, ICDAR2011 and CVL, in which they
propose to use the Convolutional Neural Network (CNN) to
generate a ray of traits for each writer that is compared with
the rays of traits stored in the database. The method gave
98.6% on ICDAR-2011, 97.6% on CVL and 40.5% on ICDAR-
2013 might be because of missing Greek training data and
wrong segmentation.
The Arabic writer identification wasnotcoveredaswidelyas
English or Chinese writer identification until the last few
years, and the first study goes back to the proposal of the
researcher AL-Zoubeidy et al in 2005 using multi -channel
Gabor filtering and gray scale co-occurrence matrices to
distinguish writer from his writing style [11]. Fouad and
Volke in 2014 proposed a system for Arabic writer
identifition using only 21 feature, and Gaussian Mixture
Models (GMMs) were also used as the kernel of this system,
where GMMs provide a strong representation of the
distribution of the extracted features using a fixed-length
sliding window It is constructed and trained using images of
words and texts for writer [12]. Djeddi et al in 2014
conducted their study on handwritten texts of 1,000 writers
using three methods to extract features: run-length
distribution, edge-hinge distribution, and edge-direction
features. They used Multiclass SVM (Support Vector
Machine), and then hadan identification rateof 84.10%[13].
In 2016, researcher Elleuch and his colleagues presented a
study based on Convolutional Neural Network (CNN) and
SVM (Support Vector Machine) for handwriting recognition.
The performance was evaluated with HACDB and IFN/ENIT
database for character images. The results showed the
effectiveness of the proposed method compared to CNN
Standard Classifier[14].In2018,theresearcherAmarandhis
colleagues proposed a deep learning methodology forslicing
and recognizing printed and handwritten characters from
words. An algorithm was proposed to slice the word into
letters, and then extract the features using a wavelet
transform, and these extracted features are exploited as
connection weights to build a convolutional neural network
for each letter shape, The proposed method was tested on
APTI and IESKarDB database, and their results showed the
effectiveness and speed of the proposed method for each of
the databases [15]. Dengel et al presented the Arabic text
recognition system using the (KHATT) database. This study
mainly contributes to three aspects: preprocessing, deep
learning, and data augmentation by creating copies of the
data. The first version is the blurred version of the data, the
second version defines the contours, the third version is
enhanced edges, and the fourth version is making effectsand
where Dataaugmentation with a deep learningapproachhas
been shown to produce a better and more promising
improvement in outcomes. This study achievedarecognition
accuracy of 80.02%, which is better compared to other
studies conducted on the same database [16]. In 2019, Aref
and colleagues presented a study based on integrating OBI
features with the character codebook using IAM databases
for English texts and ICFHR-2012 for Arabic texts. The
classification accuracy with the KNN classifier was96%[17].
In 2019, Rehman and his colleagues applied their study to
identify the writer’s identity on the QUWI database that
contains Arabic texts and English texts for 1017 writers, and
each writer has 4 texts, using techniques to increase the data
to improve performance, by obtaining from each original
image an image containing the boundaries surrounding the
words ( Contours),as wellasobtainingnewsamplesfromthe
original images by increasing their sharpness and the image
unit obtained in the first stage of data augmentation when
obtaining the bounding boundaries of words (Sharped
Contours), and finally applying Negatives to the original
images and images resulting from the previous steps to
augment the data, And using the AlexNet structure to extract
attributes to then apply the Support Vector Machine as a
classification method, the accuracy was 92.78% in English
texts, 92.20% in Arabic and 88.11% when combining Arabic
and English, respectively [18].

3. PROPOSED SYSTEM
In this paragraph, the Arabic writer identification is
presented. First, Image Acquisition of handwritten texts in
Arabic, Secondly, the data-augmentation technique for
increasing the input data images, third, finally, training
process and the experimental are presented.
3.1. Image Acquisition
The study was conducted on an ICFHR2012 database
downloaded from Kaggle for 202 writers, each writer had
three different handwritten texts in Arabic, and the images
were at a resolution of 600 DPP binary [19].
3.2. Pre-processing
The database did not need to be pre-processed so it was
binary images with white background and text in black, this
data just need to be increased.
3.3. Data Augmentation
Handwritten documents may be limited, which prompted a
number of researcherstoincreasedatatechniques,anditwas
necessary to train the neural network from the presence of a
large number of samples. This prompted us to use an
algorithm that seeks to increase the number of images for
handwritten texts in Arabic [1], especially that the database
The data on which the study was conducted (ICFHR2012
dataset) contains only three texts for each writer, and this
number is not enough to train the proposed neural network
in the research, in addition to the images of large and
different sizes, which does not fit with the nature of the
proposed network and to avoid the problem of out of
memory. To solve this problem, huge local image patches
centered on the contour of the handwriting images [20] or
inside the slipping windows on text lines [21] are extracted
by past approaches, which only consider the local structural
information of handwriting.
After reviewing some research in this field, the method of
researcher Maaz et al. was applied to increase the data[1],
and through the following stages as shown in" Figure1":
1) Consequently,theproposedmethodfordividingthetext
into words or word structures wasadopted toavoidthe
problems of using the techniques for dividing the text
into lines and then into words, in proportion to the
database through the following stages:
a) Read original pictures and find their Negative and
then determining the connected components (CCs:
connected-components) from the binary image.
b) Applying the (LoG: Laplacian of Gaussian) filter to
the matrix of connected components, which
resulting from phase (a).
c) Filling the filtered image based on morphological
reconstruction.
d) Determining the connected components (CCs:
connected-components) from the output image of
phase (c). Then, determining the connected
components, which resulting on the original.
e) Deleting the connected components that are within
larger connected components and make the
interconnectedcomponentsasinglecomponentthat
includes them.
2) Collecting all the words or word structures
randomly, which resulting from the previous stage
in one line and repeat this step to get several
different lines.
3) Forming an image of a five-line text generated from
stage (c) in the second stage.
4) Cropping each image generated from the third stage
into a number of images with a size of 400*R,where
(R) represents the height of the image generated
from the third stage, as shown in "Figure 10",
and then the size of all images was standardized
when entered into the deep learning network .
Hundreds of pictures of handwriting texts foreachwriterare
got from this stage; the first 200 handwriting images
resulting from Data Augmentation stage foreachwriterwere
taken to obtain a database of 40400 handwriting scripts for
202 writers and for each writer 200 texts.

Fig -1: Data augmentation process: (a) Negative original image. (b) Image after filtering. (c) Fillingimageregionsandholes
after filtering. (d) Determining connected components on image shown in (a). (e) Delete connected components that are
among other components. (f) Sample of word pictures and word structures for a writer. (g) Sample of generated text line
image. (h) : sample Image of a 5-line generated text. (i) Sample of the text images after cropping.
3.4 Convolutional neural network
After increase the data, a wide number of images are
generated from the training dataset, grounded on which, the
coming step is to extract some distinct features to represent
their characteristic. As done by Christlein et al and Fiel et al.
[20, 21], the Convolutional neural network is employed to
reckon the features in this work.
Convolutional neural network’s architecture has several
famous designs, including: LeNet, AlexNet [22], GoogLeNet
[23], and VGGNet [24].
Three types of A Convolutional Neural Network (CNN):
Convolutional Layer (CONV), Pooling Layer (POOL), and
Fully-Connected Layer (FC).
3.4.1 Convolutional Layer (CONV)
This layer is the most significant in any CNN structure due to
the fact filters are utilized to learn the featuresfromtheenter
image. The 2D function map consists of the fold product
between the filter values and the scanned area with the
corresponding filter dimension of the enter image, i.e. every
neuron of this layer will join to a precise enter location
according to for filter size.
In a CONVlayer, the input volume isrepresentedas[D1xM1x
P1] corresponding to the dimensionsoftheinputimage.Four
hyperparameters are represented as (N, L, Z, W)
corresponding to the number of pollutants, the size of the
filter, the strideand the quantum of zeropadding.Theoutput
volume is represented as (W2 x H2 x D2) corresponding to:
D2= (D1 – L + 2W) / Z + 1 (3.4. 1.1)
M2= (M1 – L + 2W) / Z + 1 (3.4. 1.2)
P2 = N (3.4. 1.3)
3.4.2 Pooling Layer (POOL)
It is an intermediate layer in the network that reduces or
compresses the spatialdimensionsoftheincominginputtoit,
and this reduces the computations within the network.

Pooling layers decrease the dimensions of databycombining
the outputs of neuron clusters at one layer into a single
neuron in the subsequent layer. Pooling acts on all the
neurons of the feature map, by combining small clusters,
tiling sizes such as 2 x 2 are frequently used. There are two
common types of pooling in famous use: max and average.
Max pooling uses the maximum value of each local cluster of
neurons in the feature map, while average pooling takes the
average value.
In a POOL layer, the input volume is representedas[D2xM2x
P2] corresponding to the dimensions of the input volume,
two hyperparameters are represented as [L1, Z1]
corresponding to the receptive field or size of the filter and
the stride, and the output volume is represented as [D3 x M3
x P3] corresponding to:
D3= (D2 – L) / Z + 1 (3.4.2.1)
M3= (M2 – L) / Z + 1 (3.4.2.2)
P3 = P2 (3.4.2.3)
3.3.3 Fully-Connected Layer (FC)
The FC layer in a CNN like in ANNs has neurons that are
connected every neurons in one layer to every neuron in the
past layer. This layer is often kept as the final layer of a CNN
with “SOFTMAX” as its activation function for multi-class
classification problems. The FClayerisresponsibletopredict
the final class. Thus, it has an output dimension of [1x1xD]
where D represents the number of classes or labels
considered for classification [25].
The network architecture that was used in this research
consists of 4 convolutional layers, 4 pooling layers, and 2
Fully Connected Layer (FC). [1] Each convolutional layer is
followed by a Batch NormalizationLayer(BN)andRelulayer.
This network also includes a Softmax layer, which is used for
Classification problems.
The description of this architecture convolutional network
layers is as follows and illustrated according to "Figure 2":
Fig -2: The structure of the convolutional network [1]
1. The input layer receives training samples, which is
binary images and size [400,400].
2. ConvolutionalLayer(CONV1):Inthislayer,96filters
of size F = [5, 5] were used, and stride (S = 2)
without applying zero padding.
3. Batch Normalization Layer: It normalizes the input
via mini batch, speeds up the training of
convolutional neural networks, and reduces the
sensitivity of network initialization, so batch
normalization layers are used between
convolutional layers and nonlinear layers, such as
ReLU layers.
The layer first normalizes the activations for its
input by subtracting the mean of the mini batch and
dividing by the standard deviation of the mini batch.
[26]:
4. Relu Layer: This layer applies a threshold to all its
income items from the previous layer according to
the function shown below:
5. The pooling layer: (Pool1) In this layer Max Pooling
has been used in the maximum pool and the size of
the area from which the maximum value will be
chosen is size P=[2,2], and step (S=2).
…. (3.4.2)

6. Convolutional Layer (CONV2): In this layer, 128
filters of size F = [3, 3], and stride (S = 2) without
applying zero padding.
7. Batch Normalization Layer.
8. Relu Layer.
9. The pooling layer: (Pool2): The size of the area from
which the maximum value will be chosen is of size F
= [2,2], and a stride (S = 2).
filters of size F = [3,3], and stride (S = 1) without
12. Relu Layer.
= [2, 2], and a stride (S = 2).
filters of size F = [1, 1], and stride (S = 1) without
16. Relu Layer.
= [2,2], and a stride (S = 2).
underlying networkstructurebetweeniteratorsand
helps prevent the network from over-fitting by
setting the input elements randomly to zero
according to the probability that was defined when
building the network and the most appropriate
probability value for our network is (0.5) [27],[28].
21. Fully Connected Layer (FC2): The output size has
been specified forthislayer (202), whichrepresents
the number of classes.
22. Softmax Layer Activation Function: It is used in
classification problems of several classes. In
classification problems, the last connected layer
must be followed by a softmax layer and a
classification layer, where the activation function of
the output unit is the softmax function
23. Classification Layer.
3.5 Feature extraction using CNN
This section shows how to extract learned image features
from a pertainedconvolutionalneuralnetwork,andusethose
features to train an image classifier. Feature extraction is the
easiest and fastest way to use the representational power of
pertained deep networks. Because feature extraction only
requires a single pass through the data, it is a good starting
point if you do not have a GPU to accelerate network training
with. We showed the network architectureoftheCNNbased-
SVM model in "Figure. 3".
Fig -3: Architecture of the CNN based-SVM model
4 EXPERIMENTS AND RESULTS
It is observed that it looks as follows. First, load the data.
Second, load a pre-trained network. Third, the network
creates a hierarchical exemplification of the input images.
Deeper layers contain higher-level features, which are built
using the lower-level features of past layers. To get
representations of training features and test images, use
activations on the fully connected layer 'FC1'. For a lower
level exemplification of the images, use a past layer in the
grid. Fourth, from the training images has been used the
features extracted as predictor variablesand fit a multi-class
18. Fully Connected Layer (FC1): The FC layer contains
neurons completely connected to the neurons inthe
previous layer, the output volume of this layer is
defined (4500).
19. Relu Layer.
20. Dropout Layer: This layer effectively changes the

support vector device (SVM).Finally,fromthetestimagesthe
trained SVM model andfeaturesextractedareusedtoclassify
the test images
The experiments were conducted on MatlabversionR2020a.
The database includes 202 writers and each writer has 3
texts to be increased by using the Data Augmentation
techniques proposed in this research, so that each writer has
200 texts and the dataset becomes 40400 handwriting
images. The dataset was divided into 3 groups 90% as a
trainingset, chose 5% of them as a validation set, and10%as
a test set.
To know the effectiveness of the proposed method, the
accuracy was measured using relationship (4.1) to obtain
96.5099% classification accuracy.
= (4.1)
Table -1: Table compare result between Proposed System
and Maaz et al.
Method
Feature
Extraction
Classification
Model
Accuracy
Maaz et al. CNN CNN 96.51%
Proposed
System
CNN SVM 97.20%
5 CONCLUSION AND FUTURE WORK
Writer identification by handwriting is a relatively new
biometric method that has received significant research
interest in recent years. Handwriting biometrics can be used
in forensic applications to identify individuals based on the
characteristics of their writing by comparing unsorted
handwritten texts with written samples Handwritten and
categorized.
Handwritten texts were used to verify the identity of the
author of a document. The font analyzer usually depends on
studying some features of the writing style of each writer.
This research present using a convolutional neural network,
which required the use of a data augmentation method that
fits the characteristics of the text written in Arabic to extract
the features and then Using support vector machine (SVM).
In the future, experiments will be focused interest on
applying this study to a new and more diverse database of
texts that contain structures in Arabic and English and
numbers, in order to determine the identification accuracy
with these cases in the text.
6 REFERENCES
[1] MAAZ, SH., ISSA, H., 2020- Using Deep Learning For
Arabic Writer Identification. International Journal of
Coputer Applications. Vol. 175, No 25.
[2] SREERAJ, M., IDICULA, S., 2011- A Survey on Writer
Identification Schemes. International Journal of
Computer Applications, (26).
[3] SHAHABINEJAD, F., RAHMATI, M.,2009 A New Method
for Writer Identification of Handwritten Farsi
Documents, ICDAR 2009, Spain, pp. 426 – 430
[4] SARANYA, K., VIJAYA, M. S., 2013- Text Dependent
Writer Identification using Support Vector Machine.
International Journal of Computer Applications, (65).
[5] Li,B., Sun, Z., Tan,T.N., 2009- Hierarchical Shape
Primitive Features for Online Text-independent Writer
Identification. 2009 10th International Conference on
Document Analysis and Recognition, pages 201–210.
[6] ZHU, G., YU, X., LI, Y., DOERMANN, D. 2009-Language
identification for handwritten document imagesusinga
shape codebook. Pattern Recogn.(42), 3184-3191.
[7] DJEDDI, C., SIDDIQI, I., MESLATI, L., ENNAJI, A., 2013-
Text-independent writer recognition using multi-script
handwritten texts. journal homepage, 1196–1202.
[8] ALAEI,A., ROY, P.P., 2014- A New Method for Writer
Identification based on Histogram Symbo lic
Representation. 14th International Conference on
Frontiers in Handwriting Recognition, 2167-6445.
[9] THENDRAL, T., VIJAYA, M. S., KARPAGAVALLI, S., 2015-
Prediction of Writer Using Tamil Handwritten
Document Image Based on Pooled Features.
International Journal of Computer, Electrical,
Automation, Control and Information Engineering, (9).
[10] Fiel, S., Sablatnig, R., Writer identi_cation and retrieval
using a convolutional neural network, in Proc. Int. Conf.
Comput. Anal. Images Patterns. Cham, Switzerland:
Springer, 2015, pp. 26_37.
[11] AL-ZOUBEIDY ,L. M. AL-NAJAR ,H. F., 2005- Arabic
writer identification for handwriting images,
International Arab Conference on Information
Technology, pp. 111-117.
[12] SLIMANE ,F., MARGNER, V., 2014- A New Text-
Independent GMM Writer Identification SystemApplied
to Arabic Handwriting.14thInternational Conferenceon
Frontiers in Handwriting Recogniti.

[13] DJEDDI, CH., MESLATI,L.,SIDDIQI,I.,ENNAJI,A.,ABED,H.,
GATTAL,A., 2014- Evaluation of Texture Features for
Offline Arabic Writer Identification. 11th IAPR
International Workshopon DocumentAnalysis Systems.
[14] ELLEUCH,M., MAALEJ,R., KHERALLAH,M., 2016 - A New
Design Based-SVM of the CNN Classifier Architecture
with Dropout for Offline Arabic Handwritten
Recognition. Procedia Computer Science. Volume 80,
2016, Pages 1712–1723.
[15] ELADEL,A., ZAIED,M., AMAR,CH., 2018 - Trained
convolutional neural network based on selected beta
filters for Arabic letter recognition. WIREs Data Mining
Knowl Discov.2018;e1250.
[16] DENGEL,A., LIWICKI,M., RASHID,S., AFZAL,M., NAZ,S.,
AHMAD,R. 2018 - A Deep Learning based Arabic Script
Recognition System: Benchmark on KHAT. The
International Arab Journal of Information Technology,
Vol. 17, No. 3, May 2081.
[17] Durou,A., Aref,.I., Al-Maadeed,S., Bouridane,A., 2019 –
Writer identification approach based on bag of words
with OBI features.ScienceDirect,InformationProcessing
& Management, vol 56 issue (2) pp. 354-366. ISSN
0306-4573 .
[18] REHMAN,A., NAZ,S.,RAZZAK,M.,2019 -AutomaticVisual
Features for Writer Identification: A Deep Learning
Approach. IEEE Volume 7, 2019.
[19] www.kaggle.com/c/awic2012/leaderboard.
[20] Christlein, V., Bernecker, D., Maier, A., Angelopoulou, E.
2015- Offline Writer Identification Using Convolutional
Neural Network Activation Features, Proc. German
Conference on Pattern Recognition, pp. 540-552.
[21] Fiel, S., Sablatnig, R. 2015- Writer Identification and
Retrieval Using a Convolutional Neural Network, Proc.
Computer Analysis of Images and Patterns, pp. 26-37.
[22] Krizhevsky, A., Sutskever, I., Hinton,G.E.2012-Imagenet
classification with deep convolutional neural networks,
Proc. Neural Information Processing Systems,pp.1097-
1105.
[23] Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S. ,
Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.
2014- Going deeper with convolutions, arXiv preprint
arXiv:1409.4842.
[24] Simonyan, K., Zisserman, A. 2014- Very deep
convolutional networks for large-scale image
recognition, arXiv preprint arXiv:1409.1556.
[25] Gogul.,Kumar,S., 2017- Flower Species Recognition
System using Convolution Neural Networks and
Transfer Learning. 2017 IEEE, 2017 4th International
Conference on Signal Processing, Communications and
Networking.
[26] IOFFE, S., SZEGEDY, CH., 2015- Batch Normalization:
Accelerating Deep Network Training by Reducing
Internal Covariate Shift. reprint,arXiv:1502.03167 .
[27] SRIVASTAVA,N., HINTON , G., KRIZHEVSKY, A.,
SUTSKEVER, I., SALAKHUTDINOV, R., 2014- Dropout: A
Simple Way To Prevent Neural Networks From
Overfitting. Journal of Machine Learning Research. Vol.
15, pp. 1929-1958, 2014.
[28] KRIZHEVSKY, A., SUTSKEVER, I., HINTON, E., G., 2012-
Imagenet ClassificationWith DeepConvolutional Neural
Networks. Advances in Neural Information Processing
Systems. Vol. 25, 2012.

Writer Identification via CNN Features and SVM

More Related Content

Similar to Writer Identification via CNN Features and SVM (20)

More from IRJET Journal (20)

Recently uploaded (20)

Writer Identification via CNN Features and SVM