Merge
Ncbi
Racine
Merge
Checkpoint
Ncbi
Racine
Ncbi
Exploration
Accueil
Racine
Racine

Serveur d'exploration sur l'OCR

Attention, ce site est en cours de développement !
Attention, site généré par des moyens informatiques à partir de corpus bruts.
Les informations ne sont donc pas validées.

A hybrid recognition system for off-line handwritten characters

Identifieur interne : 000258 ( Ncbi/Merge ); précédent : 000257; suivant : 000259

A hybrid recognition system for off-line handwritten characters

Auteurs : Gauri Katiyar [Inde] ; Shabana Mehfuz [Inde]

Source :

RBID : PMC:4801835

Abstract

Computer based pattern recognition is a process that involves several sub-processes, including pre-processing, feature extraction, feature selection, and classification. Feature extraction is the estimation of certain attributes of the target patterns. Selection of the right set of features is the most crucial and complex part of building a pattern recognition system. In this work we have combined multiple features extracted using seven different approaches. The novelty of this approach is to achieve better accuracy and reduced computational time for recognition of handwritten characters using Genetic Algorithm which optimizes the number of features along with a simple and adaptive Multi Layer Perceptron classifier. Experiments have been performed using standard database of CEDAR (Centre of Excellence for Document Analysis and Recognition) for English alphabet. The experimental results obtained on this database demonstrate the effectiveness of this system.


Url:
DOI: 10.1186/s40064-016-1775-7
PubMed: 27066370
PubMed Central: 4801835

Links toward previous steps (curation, corpus...)

Links to Exploration step

PMC:4801835

Le document en format XML

<record>
<TEI>
<teiHeader>
<fileDesc>
<titleStmt>
<title xml:lang="en">A hybrid recognition system for off-line handwritten characters</title>
<author>
<name sortKey="Katiyar, Gauri" sort="Katiyar, Gauri" uniqKey="Katiyar G" first="Gauri" last="Katiyar">Gauri Katiyar</name>
<affiliation wicri:level="1">
<nlm:aff id="Aff1">Department of Electrical Engineering, Jamia Millia Islamia, New Delhi, India</nlm:aff>
<country xml:lang="fr">Inde</country>
<wicri:regionArea>Department of Electrical Engineering, Jamia Millia Islamia, New Delhi</wicri:regionArea>
<wicri:noRegion>New Delhi</wicri:noRegion>
</affiliation>
<affiliation>
<nlm:aff id="Aff2">ITS Engineering College, 46 Knowledge Park, Greater Noida, Uttar Pradesh 201308 India</nlm:aff>
<wicri:noCountry code="subfield">Uttar Pradesh 201308 India</wicri:noCountry>
</affiliation>
</author>
<author>
<name sortKey="Mehfuz, Shabana" sort="Mehfuz, Shabana" uniqKey="Mehfuz S" first="Shabana" last="Mehfuz">Shabana Mehfuz</name>
<affiliation wicri:level="1">
<nlm:aff id="Aff1">Department of Electrical Engineering, Jamia Millia Islamia, New Delhi, India</nlm:aff>
<country xml:lang="fr">Inde</country>
<wicri:regionArea>Department of Electrical Engineering, Jamia Millia Islamia, New Delhi</wicri:regionArea>
<wicri:noRegion>New Delhi</wicri:noRegion>
</affiliation>
</author>
</titleStmt>
<publicationStmt>
<idno type="wicri:source">PMC</idno>
<idno type="pmid">27066370</idno>
<idno type="pmc">4801835</idno>
<idno type="url">http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4801835</idno>
<idno type="RBID">PMC:4801835</idno>
<idno type="doi">10.1186/s40064-016-1775-7</idno>
<date when="2016">2016</date>
<idno type="wicri:Area/Pmc/Corpus">000039</idno>
<idno type="wicri:Area/Pmc/Curation">000039</idno>
<idno type="wicri:Area/Pmc/Checkpoint">000008</idno>
<idno type="wicri:Area/Ncbi/Merge">000258</idno>
</publicationStmt>
<sourceDesc>
<biblStruct>
<analytic>
<title xml:lang="en" level="a" type="main">A hybrid recognition system for off-line handwritten characters</title>
<author>
<name sortKey="Katiyar, Gauri" sort="Katiyar, Gauri" uniqKey="Katiyar G" first="Gauri" last="Katiyar">Gauri Katiyar</name>
<affiliation wicri:level="1">
<nlm:aff id="Aff1">Department of Electrical Engineering, Jamia Millia Islamia, New Delhi, India</nlm:aff>
<country xml:lang="fr">Inde</country>
<wicri:regionArea>Department of Electrical Engineering, Jamia Millia Islamia, New Delhi</wicri:regionArea>
<wicri:noRegion>New Delhi</wicri:noRegion>
</affiliation>
<affiliation>
<nlm:aff id="Aff2">ITS Engineering College, 46 Knowledge Park, Greater Noida, Uttar Pradesh 201308 India</nlm:aff>
<wicri:noCountry code="subfield">Uttar Pradesh 201308 India</wicri:noCountry>
</affiliation>
</author>
<author>
<name sortKey="Mehfuz, Shabana" sort="Mehfuz, Shabana" uniqKey="Mehfuz S" first="Shabana" last="Mehfuz">Shabana Mehfuz</name>
<affiliation wicri:level="1">
<nlm:aff id="Aff1">Department of Electrical Engineering, Jamia Millia Islamia, New Delhi, India</nlm:aff>
<country xml:lang="fr">Inde</country>
<wicri:regionArea>Department of Electrical Engineering, Jamia Millia Islamia, New Delhi</wicri:regionArea>
<wicri:noRegion>New Delhi</wicri:noRegion>
</affiliation>
</author>
</analytic>
<series>
<title level="j">SpringerPlus</title>
<idno type="eISSN">2193-1801</idno>
<imprint>
<date when="2016">2016</date>
</imprint>
</series>
</biblStruct>
</sourceDesc>
</fileDesc>
<profileDesc>
<textClass></textClass>
</profileDesc>
</teiHeader>
<front>
<div type="abstract" xml:lang="en">
<p>Computer based pattern recognition is a process that involves several sub-processes, including pre-processing, feature extraction, feature selection, and classification. Feature extraction is the estimation of certain attributes of the target patterns. Selection of the right set of features is the most crucial and complex part of building a pattern recognition system. In this work we have combined multiple features extracted using seven different approaches. The novelty of this approach is to achieve better accuracy and reduced computational time for recognition of handwritten characters using Genetic Algorithm which optimizes the number of features along with a simple and adaptive Multi Layer Perceptron classifier. Experiments have been performed using standard database of CEDAR (Centre of Excellence for Document Analysis and Recognition) for English alphabet. The experimental results obtained on this database demonstrate the effectiveness of this system.</p>
</div>
</front>
<back>
<div1 type="bibliography">
<listBibl>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Arica, N" uniqKey="Arica N">N Arica</name>
</author>
<author>
<name sortKey="Yarman Vural, Ft" uniqKey="Yarman Vural F">FT Yarman-Vural</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Avi Itzhak, Hi" uniqKey="Avi Itzhak H">HI Avi-Itzhak</name>
</author>
<author>
<name sortKey="Diep, Ta" uniqKey="Diep T">TA Diep</name>
</author>
<author>
<name sortKey="Garland, H" uniqKey="Garland H">H Garland</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Chen, My" uniqKey="Chen M">MY Chen</name>
</author>
<author>
<name sortKey="Kundu, A" uniqKey="Kundu A">A Kundu</name>
</author>
<author>
<name sortKey="Zhou, J" uniqKey="Zhou J">J Zhou</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Cho, Sb" uniqKey="Cho S">SB Cho</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Das, N" uniqKey="Das N">N Das</name>
</author>
<author>
<name sortKey="Sarkar, R" uniqKey="Sarkar R">R Sarkar</name>
</author>
<author>
<name sortKey="Basu, S" uniqKey="Basu S">S Basu</name>
</author>
<author>
<name sortKey="Kundu, M" uniqKey="Kundu M">M Kundu</name>
</author>
<author>
<name sortKey="Nasipuri, M" uniqKey="Nasipuri M">M Nasipuri</name>
</author>
<author>
<name sortKey="Basu, Kd" uniqKey="Basu K">KD Basu</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="De Stefano, C" uniqKey="De Stefano C">C De Stefano</name>
</author>
<author>
<name sortKey="Della Cioppa, A" uniqKey="Della Cioppa A">A Della Cioppa</name>
</author>
<author>
<name sortKey="Marcelli, A" uniqKey="Marcelli A">A Marcelli</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="De Stefano, C" uniqKey="De Stefano C">C De Stefano</name>
</author>
<author>
<name sortKey="Fontanella, F" uniqKey="Fontanella F">F Fontanella</name>
</author>
<author>
<name sortKey="Marrocco, C" uniqKey="Marrocco C">C Marrocco</name>
</author>
<author>
<name sortKey="Scotto Di Freca, A" uniqKey="Scotto Di Freca A">A Scotto di Freca</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Govindan, K" uniqKey="Govindan K">K Govindan</name>
</author>
<author>
<name sortKey="Shivaprasad, Ap" uniqKey="Shivaprasad A">AP Shivaprasad</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Hanmandlu, M" uniqKey="Hanmandlu M">M Hanmandlu</name>
</author>
<author>
<name sortKey="Murthy, Ovr" uniqKey="Murthy O">OVR Murthy</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Hanmandlu, M" uniqKey="Hanmandlu M">M Hanmandlu</name>
</author>
<author>
<name sortKey="Mohan, Krm" uniqKey="Mohan K">KRM Mohan</name>
</author>
<author>
<name sortKey="Chakraborty, S" uniqKey="Chakraborty S">S Chakraborty</name>
</author>
<author>
<name sortKey="Goyal, S" uniqKey="Goyal S">S Goyal</name>
</author>
<author>
<name sortKey="Choudhury, Dr" uniqKey="Choudhury D">DR Choudhury</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Hull, Jj" uniqKey="Hull J">JJ Hull</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Katiyar, G" uniqKey="Katiyar G">G Katiyar</name>
</author>
<author>
<name sortKey="Mehfuz, S" uniqKey="Mehfuz S">S Mehfuz</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Lee, Sw" uniqKey="Lee S">SW Lee</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Lee, S W" uniqKey="Lee S">S-W Lee</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Liang, Z" uniqKey="Liang Z">Z Liang</name>
</author>
<author>
<name sortKey="Xia, S" uniqKey="Xia S">S Xia</name>
</author>
<author>
<name sortKey="Zhou, Y" uniqKey="Zhou Y">Y Zhou</name>
</author>
<author>
<name sortKey="Zhang, L" uniqKey="Zhang L">L Zhang</name>
</author>
<author>
<name sortKey="Li, Y" uniqKey="Li Y">Y Li</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Mehfuz, S" uniqKey="Mehfuz S">S Mehfuz</name>
</author>
<author>
<name sortKey="Katiyar, G" uniqKey="Katiyar G">G Katiyar</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Mori, S" uniqKey="Mori S">S Mori</name>
</author>
<author>
<name sortKey="Suen, Cy" uniqKey="Suen C">CY Suen</name>
</author>
<author>
<name sortKey="Yamamoto, K" uniqKey="Yamamoto K">K Yamamoto</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Parkins, Ad" uniqKey="Parkins A">AD Parkins</name>
</author>
<author>
<name sortKey="Nandi, Ak" uniqKey="Nandi A">AK Nandi</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Plamondon, R" uniqKey="Plamondon R">R Plamondon</name>
</author>
<author>
<name sortKey="Srihari, Sn" uniqKey="Srihari S">SN Srihari</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Pradeep, J" uniqKey="Pradeep J">J Pradeep</name>
</author>
<author>
<name sortKey="Srinivasn, E" uniqKey="Srinivasn E">E Srinivasn</name>
</author>
<author>
<name sortKey="Himavathi, S" uniqKey="Himavathi S">S Himavathi</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Murthy, Ovr" uniqKey="Murthy O">OVR Murthy</name>
</author>
<author>
<name sortKey="Hanmandlu, M" uniqKey="Hanmandlu M">M Hanmandlu</name>
</author>
</analytic>
</biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Shrivastava, S" uniqKey="Shrivastava S">S Shrivastava</name>
</author>
<author>
<name sortKey="Singh, Mp" uniqKey="Singh M">MP Singh</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct>
<analytic>
<author>
<name sortKey="Trier, Od" uniqKey="Trier O">OD Trier</name>
</author>
<author>
<name sortKey="Jain, Ak" uniqKey="Jain A">AK Jain</name>
</author>
<author>
<name sortKey="Taxt, J" uniqKey="Taxt J">J Taxt</name>
</author>
</analytic>
</biblStruct>
<biblStruct></biblStruct>
<biblStruct></biblStruct>
<biblStruct></biblStruct>
</listBibl>
</div1>
</back>
</TEI>
<pmc article-type="research-article">
<pmc-dir>properties open_access</pmc-dir>
<front>
<journal-meta>
<journal-id journal-id-type="nlm-ta">Springerplus</journal-id>
<journal-id journal-id-type="iso-abbrev">Springerplus</journal-id>
<journal-title-group>
<journal-title>SpringerPlus</journal-title>
</journal-title-group>
<issn pub-type="epub">2193-1801</issn>
<publisher>
<publisher-name>Springer International Publishing</publisher-name>
<publisher-loc>Cham</publisher-loc>
</publisher>
</journal-meta>
<article-meta>
<article-id pub-id-type="pmid">27066370</article-id>
<article-id pub-id-type="pmc">4801835</article-id>
<article-id pub-id-type="publisher-id">1775</article-id>
<article-id pub-id-type="doi">10.1186/s40064-016-1775-7</article-id>
<article-categories>
<subj-group subj-group-type="heading">
<subject>Research</subject>
</subj-group>
</article-categories>
<title-group>
<article-title>A hybrid recognition system for off-line handwritten characters</article-title>
</title-group>
<contrib-group>
<contrib contrib-type="author" corresp="yes">
<name>
<surname>Katiyar</surname>
<given-names>Gauri</given-names>
</name>
<address>
<email>gaurikatiyar0208@yahoo.com</email>
<email>gaurikatiyarrl@its.edu.in</email>
</address>
<xref ref-type="aff" rid="Aff1"></xref>
<xref ref-type="aff" rid="Aff2"></xref>
</contrib>
<contrib contrib-type="author">
<name>
<surname>Mehfuz</surname>
<given-names>Shabana</given-names>
</name>
<address>
<email>shabana_mehfuz@yahoo.com</email>
</address>
<xref ref-type="aff" rid="Aff1"></xref>
</contrib>
<aff id="Aff1">
<label></label>
Department of Electrical Engineering, Jamia Millia Islamia, New Delhi, India</aff>
<aff id="Aff2">
<label></label>
ITS Engineering College, 46 Knowledge Park, Greater Noida, Uttar Pradesh 201308 India</aff>
</contrib-group>
<pub-date pub-type="epub">
<day>22</day>
<month>3</month>
<year>2016</year>
</pub-date>
<pub-date pub-type="pmc-release">
<day>22</day>
<month>3</month>
<year>2016</year>
</pub-date>
<pub-date pub-type="collection">
<year>2016</year>
</pub-date>
<volume>5</volume>
<elocation-id>357</elocation-id>
<history>
<date date-type="received">
<day>24</day>
<month>8</month>
<year>2015</year>
</date>
<date date-type="accepted">
<day>12</day>
<month>2</month>
<year>2016</year>
</date>
</history>
<permissions>
<copyright-statement>© Katiyar and Mehfuz. 2016</copyright-statement>
<license license-type="OpenAccess">
<license-p>
<bold>Open Access</bold>
This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (
<ext-link ext-link-type="uri" xlink:href="http://creativecommons.org/licenses/by/4.0/">http://creativecommons.org/licenses/by/4.0/</ext-link>
), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.</license-p>
</license>
</permissions>
<abstract id="Abs1">
<p>Computer based pattern recognition is a process that involves several sub-processes, including pre-processing, feature extraction, feature selection, and classification. Feature extraction is the estimation of certain attributes of the target patterns. Selection of the right set of features is the most crucial and complex part of building a pattern recognition system. In this work we have combined multiple features extracted using seven different approaches. The novelty of this approach is to achieve better accuracy and reduced computational time for recognition of handwritten characters using Genetic Algorithm which optimizes the number of features along with a simple and adaptive Multi Layer Perceptron classifier. Experiments have been performed using standard database of CEDAR (Centre of Excellence for Document Analysis and Recognition) for English alphabet. The experimental results obtained on this database demonstrate the effectiveness of this system.</p>
</abstract>
<kwd-group xml:lang="en">
<title>Keywords</title>
<kwd>Character recognition</kwd>
<kwd>Multi layer perceptron</kwd>
<kwd>Feature extraction</kwd>
<kwd>Genetic Algorithm</kwd>
</kwd-group>
<custom-meta-group>
<custom-meta>
<meta-name>issue-copyright-statement</meta-name>
<meta-value>© The Author(s) 2016</meta-value>
</custom-meta>
</custom-meta-group>
</article-meta>
</front>
<body>
<sec id="Sec1">
<title>Background</title>
<p>Machine simulation of functional perspective of human being has been a significant and vital research area in the field of image processing and pattern recognition (Plamondon and Srihari
<xref ref-type="bibr" rid="CR31">2000</xref>
; Arica and Yarman-Vural
<xref ref-type="bibr" rid="CR2">2001</xref>
). Handwriting Recognition is the mechanism for converting the handwritten text into a notational representation. It is a special problem in the domain of pattern recognition and machine intelligence. Character recognition can be split up into two modes: Online and offline—depending on the type of available data. Online character recognition involves the identification of character while they are being written and are captured by special hardware (e.g. smart pen or pressure sensitive tablets), which is capable of measuring pen’s pressure and velocity. Offline character recognition on the other hand, are written on paper with ordinary pens and converted into scanned digital images. Systems for recognizing machine printed text originated in the late 1950s and have been in widespread use on desktop computers since the early 1990s. In the early 1990s, image processing and pattern recognition are efficiently and effectively combined with artificial intelligence and statistical technique that is Hidden Morkov Model (HMM) (Avi-Itzhak et al.
<xref ref-type="bibr" rid="CR3">1995</xref>
; Chen et al.
<xref ref-type="bibr" rid="CR4">1994</xref>
).</p>
<p>Off-line handwritten character recognition continues to be an active research area towards exploring the newer techniques because it has various applications such as postal sorting, bank cheque amount reading, and official document reading. Mori et al. (
<xref ref-type="bibr" rid="CR27">1992</xref>
) presented historical review of off-line character recognition research and development. A comprehensive survey of recognition techniques has been reported in (Govindan and Shivaprasad
<xref ref-type="bibr" rid="CR13">1990</xref>
; Mehfuz and Katiyar
<xref ref-type="bibr" rid="CR25">2012</xref>
; Katiyar and Mehfuz
<xref rid="CR19" ref-type="bibr">2012</xref>
). A survey on on-line and off-line handwritten and hand printed character recognition has already been presented (Plamondon and Srihari
<xref ref-type="bibr" rid="CR31">2000</xref>
). Recognition of cursive handwritten character is a difficult task. Most of the researchers have conducted their work on unconstrained character (Hanmandlu et al.
<xref ref-type="bibr" rid="CR17">2003</xref>
; Saurabh and Singh
<xref ref-type="bibr" rid="CR35">2011</xref>
; Lee
<xref ref-type="bibr" rid="CR22">1996</xref>
).</p>
<p>Feature selection is a process of selecting only significant features from a large database to create subset of the original database. This process retains the original feature without changing the features domain. The main objectives of the feature selection are: to reduce database dimensionally, remove irrelevant features, reduce time needed to learn the classification function, increase accuracy and to keep only important features that give comprehensive understandings for all variables. Demand of feature selection is currently rising because of the expanding sizes of databases in various applications. There are many ways of feature selection using search algorithms such as Sequential Forward Selection (SFS), Sequential Backward Selection (SBS), Exhaustive search and Genetic Algorithm (GA). Exhaustive search is not suitable for large database as it is a time consuming process. SFS and SBS operations are good because all features in database are evaluated at the same time. However once a feature is removed it does not have chance of being selected again. The selected features will remain in the selection even after new features selected, resulting some redundancies in some cases. GA differs from this type of feature selection method due to the capability to evolve new features from the selected features and a vast exploration of search space for new fitter solutions. The evolving process is made possible using GA operators which are selection, cross-over and mutations. The process will continue until the best solutions are found or the maximum number of iterations is met.</p>
<p>De Stefano et al. (
<xref ref-type="bibr" rid="CR9">2002</xref>
) have investigated the possibility of using a pre-classification technique for reducing the number of classes. They have used genetic programming for pre classification. Parkins and Nandi (
<xref ref-type="bibr" rid="CR30">2004</xref>
) have used Genetic Algorithm based feature selection and Neural Network as a classifier for handwritten digit recognition. In Saurabh and Singh (
<xref ref-type="bibr" rid="CR35">2011</xref>
) performance of feed-forward Neural Networks for recognition of handwritten English alphabet has been evaluated with three different soft computing techniques namely back-propagation, evolutionary algorithm and hybrid evolutionary algorithm.</p>
<p>The authors have presented a GA based feature selection algorithm for detecting feature subsets, where the samples belonging to different classes are properly classified (De Stefano et al.
<xref ref-type="bibr" rid="CR10">2013</xref>
). Das et al. (
<xref ref-type="bibr" rid="CR8">2012</xref>
) have applied GA to extract the optimal feature set that has the best discriminating features to recognise the handwritten Bangla digits. To recognise handwritten Farsi/Arabic digit, MLPNN (Multilayer Perceptron Neural Network) with back-propagation as classifier has been used (Shayegan and Chan
<xref rid="CR26" ref-type="bibr">2012</xref>
). By employing 1 and 2D spectrum diagram for standard deviation and minimum to maximum distribution, an optimal subset of initial feature subset was selected automatically.</p>
<p>The authors have proposed a fast and efficient solution based on Genetic Algorithm to solve the feature selection problems (Chouaib et al.
<xref ref-type="bibr" rid="CR6">2012</xref>
; Lanzi
<xref ref-type="bibr" rid="CR20">1997</xref>
). Lp-norms generalized principle component analysis (PCA) for dimensionality reduction has been proposed in (Liang et al.
<xref ref-type="bibr" rid="CR24">2013</xref>
).</p>
<p>To recognise Chinese character a Genetic Algorithm based feature selection method with Support Vector Machine as a classifier has been proposed (Feng et al.
<xref ref-type="bibr" rid="CR11">2004</xref>
). A multi objective Genetic Algorithm has been used for feature selection for handwriting recognition (Oliveira et al.
<xref ref-type="bibr" rid="CR29">2002</xref>
). The authors have proposed a new feature selection method for multi class problem based on recursive feature elimination in least square Support Vector Machine (LS-SVM) (Zeng and Chen
<xref ref-type="bibr" rid="CR40">2009</xref>
). In another work the authors have proposed a simple multilayer cluster neutral network with five independent sub networks for off-line recognition of totally unconstrained handwritten numeral and Genetic Algorithm have been used to avoid problems of finding local minima in training the multilayer cluster neural network with gradient descent techniques (Lee and Kim
<xref ref-type="bibr" rid="CR23">1994</xref>
; Lee
<xref ref-type="bibr" rid="CR21">1995</xref>
,
<xref ref-type="bibr" rid="CR22">1996</xref>
). Comparison of five feature selection methods using Neural Network as a classifier for handwritten character recognition has been presented in Chung and Yoon (
<xref ref-type="bibr" rid="CR7">1997</xref>
). In Cho (
<xref ref-type="bibr" rid="CR5">1999</xref>
) GA determines the optimal weight parameter of the neural network to classify the unconstrained handwritten numerals.</p>
<p>In our work GA has been used for feature selection because of its potential as an optimization technique for complex problems. Advantage of this approach include the ability to accommodate multiple criteria such as number of features and accuracy of classifier as well as the capacity to deal with huge database in order to adequately represent the pattern recognition problem.</p>
<p>The decision making part of the recognition system is the classification stage. The efficiency of the classifier depends on the quality of the features. Different classification techniques for character recognition can be split up into two general approaches (Plamondon and Srihari
<xref ref-type="bibr" rid="CR31">2000</xref>
; Arica and Yarman-Vural
<xref ref-type="bibr" rid="CR2">2001</xref>
).</p>
<sec id="Sec2">
<title>Classical techniques</title>
<p>Template matching, Statistical techniques and Structural techniques</p>
</sec>
<sec id="Sec3">
<title>Soft computing techniques</title>
<p>Artificial Neural Networks (ANNs), Fuzzy logic technique, and Evolutionary computing techniques.</p>
<p>Nowadays Classical techniques are not extensively used in this area of research because the performance of classical techniques are dependent on the amount of data used in the defining the parameters of statistical model and are not flexible enough to adapt to new handwriting constraints. Moreover these techniques can be used only when the input data is uniform with respect to time. Recognition techniques based on Neural Network has been currently taking more attention of researchers. Quick recognition, automatic learning and flexibility once the networks are properly trained are some of the advantages of ANN. Therefore we have chosen Neural Network as a classifier in our recognition system.</p>
</sec>
</sec>
<sec id="Sec4">
<title>Contribution of this work</title>
<p>To improve the performance of the off-line handwritten character recognition system either the performance of the classifier has to be improved or better feature extraction techniques and/or feature selection techniques need to be explored. We have presented a methodology for hybrid feature extraction and a Genetic Algorithm based approach for optimal selection of feature subset along with an adaptive Multi layer perception (MLP) as pattern classifier. Adaptive nature in the classifier is achieved by implementing a function for selection of best architecture of MLP during the feature selection and classification phases. We have extracted seven feature sets based on moment features, distance-based feature, geometrical feature and local features as discussed in the following section. The effectiveness of the method is tested on the problem of off-line handwritten character recognition to address the problem of diversity in style, size and shape, which can be found in handwriting produced by different writers. All the handwritten data considered here are unconstrained alphabets to avoid the process of segmentation. Selection of features also plays an important role in improving the performance of the system. The novelty of the proposed recognition system is that by hybridization of the feature extraction techniques and randomly selecting the features using GA along with an adaptive MLP Neural Network classifier, the accuracy of the system is improved and the computational time of the system is reduced.</p>
</sec>
<sec id="Sec5">
<title>Proposed recognition system</title>
<p>The objective of the proposed work is to use a GA based feature reduction technique to develop an off-line character recognition system using Adaptive Neural Network as a classifier. The following steps have been adopted in the algorithm for character recognition:
<list list-type="bullet">
<list-item>
<p>Input</p>
</list-item>
<list-item>
<p>Pre-processing</p>
</list-item>
<list-item>
<p>Feature extraction</p>
</list-item>
<list-item>
<p>Feature selection using GA</p>
</list-item>
<list-item>
<p>Artificial Neural Network based classifier.</p>
</list-item>
</list>
</p>
<p>All the above mentioned stages and their interconnection for the proposed recognition system are shown in Fig. 
<xref rid="Fig1" ref-type="fig">1</xref>
.
<fig id="Fig1">
<label>Fig. 1</label>
<caption>
<p>Block diagram of the proposed recognition system</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig1_HTML" id="MO1"></graphic>
</fig>
</p>
<sec id="Sec6">
<title>Input database</title>
<p>For off-line handwritten English alphabets recognition the benchmark dataset used by researchers is CEDAR (Center for Excellence in Document Analysis and Recognition, USA) CDROM-1. The database has been acquired from CEDAR, Buffalo University, USA (
<ext-link ext-link-type="uri" xlink:href="http://www.cedar.buffalo.edu/ilt/research.html">http://www.cedar.buffalo.edu/ilt/research.html</ext-link>
). The bi-tonal images of alphabetic and numeric characters in the database are divided into two groups. One group contains mixed alphabetic and numeric (BINANUMS) and the other group contains numeric characters (BINDIGIS) only. The separation of database into training and testing set is shown in Table 
<xref rid="Tab1" ref-type="table">1</xref>
(Hull
<xref ref-type="bibr" rid="CR18">1995</xref>
). There are 19,145 known alphabet characters in training set and 2183 unknown alphabet characters in testing set.
<table-wrap id="Tab1">
<label>Table 1</label>
<caption>
<p>Number of samples from CEDAR CDROM-1for training and testing</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Capital alphabet</th>
<th align="left">Training set</th>
<th align="left">Testing set</th>
<th align="left">Small alphabet</th>
<th align="left">Training set</th>
<th align="left">Testing set</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">A</td>
<td char="." align="char">1237</td>
<td char="." align="char">162</td>
<td align="left">a</td>
<td char="." align="char">57</td>
<td char="." align="char">68</td>
</tr>
<tr>
<td align="left">B</td>
<td char="." align="char">595</td>
<td char="." align="char">69</td>
<td align="left">b</td>
<td char="." align="char">84</td>
<td char="." align="char">8</td>
</tr>
<tr>
<td align="left">C</td>
<td char="." align="char">588</td>
<td char="." align="char">51</td>
<td align="left">c</td>
<td char="." align="char">211</td>
<td char="." align="char">17</td>
</tr>
<tr>
<td align="left">D</td>
<td char="." align="char">388</td>
<td char="." align="char">49</td>
<td align="left">d</td>
<td char="." align="char">249</td>
<td char="." align="char">26</td>
</tr>
<tr>
<td align="left">E</td>
<td char="." align="char">490</td>
<td char="." align="char">57</td>
<td align="left">e</td>
<td char="." align="char">736</td>
<td char="." align="char">107</td>
</tr>
<tr>
<td align="left">F</td>
<td char="." align="char">287</td>
<td char="." align="char">32</td>
<td align="left">f</td>
<td char="." align="char">120</td>
<td char="." align="char">16</td>
</tr>
<tr>
<td align="left">G</td>
<td char="." align="char">143</td>
<td char="." align="char">19</td>
<td align="left">g</td>
<td char="." align="char">93</td>
<td char="." align="char">14</td>
</tr>
<tr>
<td align="left">H</td>
<td char="." align="char">247</td>
<td char="." align="char">25</td>
<td align="left">h</td>
<td char="." align="char">205</td>
<td char="." align="char">23</td>
</tr>
<tr>
<td align="left">I</td>
<td char="." align="char">490</td>
<td char="." align="char">56</td>
<td align="left">I</td>
<td char="." align="char">803</td>
<td char="." align="char">70</td>
</tr>
<tr>
<td align="left">J</td>
<td char="." align="char">68</td>
<td char="." align="char">15</td>
<td align="left">j</td>
<td char="." align="char">1</td>
<td char="." align="char">2</td>
</tr>
<tr>
<td align="left">K</td>
<td char="." align="char">160</td>
<td char="." align="char">19</td>
<td align="left">k</td>
<td char="." align="char">94</td>
<td char="." align="char">14</td>
</tr>
<tr>
<td align="left">L</td>
<td char="." align="char">563</td>
<td char="." align="char">87</td>
<td align="left">i</td>
<td char="." align="char">684</td>
<td char="." align="char">69</td>
</tr>
<tr>
<td align="left">M</td>
<td char="." align="char">588</td>
<td char="." align="char">59</td>
<td align="left">m</td>
<td char="." align="char">184</td>
<td char="." align="char">21</td>
</tr>
<tr>
<td align="left">N</td>
<td char="." align="char">1022</td>
<td char="." align="char">123</td>
<td align="left">n</td>
<td char="." align="char">469</td>
<td char="." align="char">60</td>
</tr>
<tr>
<td align="left">O</td>
<td char="." align="char">905</td>
<td char="." align="char">102</td>
<td align="left">o</td>
<td char="." align="char">833</td>
<td char="." align="char">84</td>
</tr>
<tr>
<td align="left">P</td>
<td char="." align="char">516</td>
<td char="." align="char">59</td>
<td align="left">p</td>
<td char="." align="char">129</td>
<td char="." align="char">7</td>
</tr>
<tr>
<td align="left">Q</td>
<td char="." align="char">3</td>
<td char="." align="char">2</td>
<td align="left">q</td>
<td char="." align="char">5</td>
<td char="." align="char">2</td>
</tr>
<tr>
<td align="left">R</td>
<td char="." align="char">749</td>
<td char="." align="char">97</td>
<td align="left">r</td>
<td char="." align="char">460</td>
<td char="." align="char">47</td>
</tr>
<tr>
<td align="left">S</td>
<td char="." align="char">834</td>
<td char="." align="char">77</td>
<td align="left">s</td>
<td char="." align="char">490</td>
<td char="." align="char">29</td>
</tr>
<tr>
<td align="left">T</td>
<td char="." align="char">441</td>
<td char="." align="char">54</td>
<td align="left">t</td>
<td char="." align="char">428</td>
<td char="." align="char">46</td>
</tr>
<tr>
<td align="left">U</td>
<td char="." align="char">268</td>
<td char="." align="char">31</td>
<td align="left">u</td>
<td char="." align="char">333</td>
<td char="." align="char">26</td>
</tr>
<tr>
<td align="left">V</td>
<td char="." align="char">201</td>
<td char="." align="char">27</td>
<td align="left">v</td>
<td char="." align="char">127</td>
<td char="." align="char">20</td>
</tr>
<tr>
<td align="left">W</td>
<td char="." align="char">249</td>
<td char="." align="char">30</td>
<td align="left">w</td>
<td char="." align="char">102</td>
<td char="." align="char">15</td>
</tr>
<tr>
<td align="left">X</td>
<td char="." align="char">112</td>
<td char="." align="char">19</td>
<td align="left">x</td>
<td char="." align="char">173</td>
<td char="." align="char">10</td>
</tr>
<tr>
<td align="left">Y</td>
<td char="." align="char">259</td>
<td char="." align="char">39</td>
<td align="left">y</td>
<td char="." align="char">136</td>
<td char="." align="char">15</td>
</tr>
<tr>
<td align="left">Z</td>
<td char="." align="char">24</td>
<td char="." align="char">7</td>
<td align="left">z</td>
<td char="." align="char">15</td>
<td char="." align="char">0</td>
</tr>
<tr>
<td align="left">Total</td>
<td char="." align="char">11,454</td>
<td char="." align="char">1367</td>
<td align="left">Total</td>
<td char="." align="char">7691</td>
<td char="." align="char">816</td>
</tr>
</tbody>
</table>
</table-wrap>
</p>
</sec>
<sec id="Sec7">
<title>Preprocessing</title>
<p>The main goal of the pre-processing is to arrange the information to make the character recognition process simpler. The following pre-processing steps have been applied.</p>
<sec id="Sec8">
<title>Binarization</title>
<p>In order to avoid the problems resulting due to noise and lost information, the gray scale image of up to 256 gray levels is converted into binary matrix. We have used global thresholding method for binarization. If the intensity of the pixel is more than a particular threshold value, it is set to white (represented by 1) and otherwise to black (represented by 0). In this work we have chosen
<bold>mean [mean (image)]</bold>
as a threshold. The threshold value changes as the images change.</p>
</sec>
<sec id="Sec9">
<title>Slant correction</title>
<p>Slant is defined as slope of the general writing trend with respect to the vertical line. The image matrix is divided into upper and lower halves. The centers of gravity of the lower and upper halves are computed and connected. The slope of the connecting line defines the slope of the window (image matrix) (Hanmandlu et al.
<xref ref-type="bibr" rid="CR17">2003</xref>
; Hanmandlu and Murthy
<xref ref-type="bibr" rid="CR15">2007</xref>
).</p>
</sec>
<sec id="Sec10">
<title>Smoothing and noise removal</title>
<p>Exact surrounding region of a character image can be determined by smoothing using Wiener filter. Median filter is applied to further enhance the image quality by removing any leftover noise.</p>
</sec>
<sec id="Sec11">
<title>Normalization</title>
<p>Normalization is required as the size of the character varies from one person to another and also from time to time even when the person is same. Normalization helps in equating the size of the character image (binary matrix) so that features can be extracted on the same footing. The character image is standardized to a window size of 42 × 32 as shown in Fig. 
<xref rid="Fig2" ref-type="fig">2</xref>
.
<fig id="Fig2">
<label>Fig. 2</label>
<caption>
<p>Normalisation of an image into a window size of 42 × 32</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig2_HTML" id="MO3"></graphic>
</fig>
</p>
</sec>
</sec>
<sec id="Sec12">
<title>Feature extraction and feature selection</title>
<p>For any recognition system, feature extraction is an integral part. There are several feature extraction techniques and its selection is the most crucial factor in achieving high accuracy. The right choice of the features from the whole available set is the most important step in any classification process. Different feature selection approaches have been proposed in literature (Trier et al.
<xref ref-type="bibr" rid="CR37">1996</xref>
).</p>
<sec id="Sec13">
<title>Feature extraction</title>
<p>At this stage, the features of the characters that are crucial for classifying them at recognition stage are extracted. We have extracted seven sets of feature vectors. The methods of feature extraction are as follows:</p>
<sec id="Sec14">
<title>Box approach</title>
<p>Box approach is based on spatial division of character image. Horizontal and vertical grid lines of 6 × 4 are superimposed on the character image of size 42 × 32. In this process 24 boxes, each of size 7 × 8 are devised. Some of the boxes will have a portion of the image and others remain empty. However, all boxes are considered for analysis (Hanmandlu et al.
<xref ref-type="bibr" rid="CR17">2003</xref>
; Hanmandlu and Murthy
<xref ref-type="bibr" rid="CR15">2007</xref>
) as shown in Fig. 
<xref rid="Fig3" ref-type="fig">3</xref>
.
<fig id="Fig3">
<label>Fig. 3</label>
<caption>
<p>Diagonal distance feature extraction method</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig3_HTML" id="MO4"></graphic>
</fig>
</p>
<p>A normalized vector distance for each box is computed as:
<disp-formula id="Equ1">
<label>1</label>
<alternatives>
<tex-math id="M1">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma_{b} = \frac{1}{{n_{b} }}\sum\limits_{k = 1}^{{n_{b} }} {d_{k}^{b} }$$\end{document}</tex-math>
<mml:math id="M2" display="block">
<mml:mrow>
<mml:msub>
<mml:mi mathvariant="italic">γ</mml:mi>
<mml:mi>b</mml:mi>
</mml:msub>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mn>1</mml:mn>
<mml:msub>
<mml:mi>n</mml:mi>
<mml:mi>b</mml:mi>
</mml:msub>
</mml:mfrac>
<mml:munderover>
<mml:mo movablelimits="false">∑</mml:mo>
<mml:mrow>
<mml:mi>k</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:msub>
<mml:mi>n</mml:mi>
<mml:mi>b</mml:mi>
</mml:msub>
</mml:munderover>
<mml:msubsup>
<mml:mi>d</mml:mi>
<mml:mrow>
<mml:mi>k</mml:mi>
</mml:mrow>
<mml:mi>b</mml:mi>
</mml:msubsup>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ1.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
<disp-formula id="Equ2">
<label>2</label>
<alternatives>
<tex-math id="M3">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{k}^{b} = (i^{2} + j^{2} )^{{\frac{1}{2}}}$$\end{document}</tex-math>
<mml:math id="M4" display="block">
<mml:mrow>
<mml:msubsup>
<mml:mi>d</mml:mi>
<mml:mrow>
<mml:mi>k</mml:mi>
</mml:mrow>
<mml:mi>b</mml:mi>
</mml:msubsup>
<mml:mo>=</mml:mo>
<mml:msup>
<mml:mrow>
<mml:mo stretchy="false">(</mml:mo>
<mml:msup>
<mml:mi>i</mml:mi>
<mml:mn>2</mml:mn>
</mml:msup>
<mml:mo>+</mml:mo>
<mml:msup>
<mml:mi>j</mml:mi>
<mml:mn>2</mml:mn>
</mml:msup>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
<mml:mfrac>
<mml:mn>1</mml:mn>
<mml:mn>2</mml:mn>
</mml:mfrac>
</mml:msup>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ2.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
where n
<sub>b</sub>
is number of pixels in bth box.</p>
<p>Normalized angle for each box is computed as:
<disp-formula id="Equ3">
<label>3</label>
<alternatives>
<tex-math id="M5">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha_{b} = \frac{1}{{n_{b} }}\sum\limits_{k = 1}^{{n_{b} }} {\theta_{k}^{b} }$$\end{document}</tex-math>
<mml:math id="M6" display="block">
<mml:mrow>
<mml:msub>
<mml:mi mathvariant="italic">α</mml:mi>
<mml:mi>b</mml:mi>
</mml:msub>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mn>1</mml:mn>
<mml:msub>
<mml:mi>n</mml:mi>
<mml:mi>b</mml:mi>
</mml:msub>
</mml:mfrac>
<mml:munderover>
<mml:mo movablelimits="false">∑</mml:mo>
<mml:mrow>
<mml:mi>k</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:msub>
<mml:mi>n</mml:mi>
<mml:mi>b</mml:mi>
</mml:msub>
</mml:munderover>
<mml:msubsup>
<mml:mi mathvariant="italic">θ</mml:mi>
<mml:mrow>
<mml:mi>k</mml:mi>
</mml:mrow>
<mml:mi>b</mml:mi>
</mml:msubsup>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ3.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
where b = 1, 2,…,24. θ
<sub>k</sub>
 = tan
<sup>−1</sup>
(−j/i) for pixel at (i, j).</p>
<p>These 24 pairs constitute the feature set.</p>
</sec>
<sec id="Sec15">
<title>Diagonal distance approach</title>
<p>Every character image of size 42 × 32 pixels is divided into 24 equal zones each of size 7 × 8 pixels. The features are extracted from each zone pixels by moving along the diagonals of its respective 7 × 8 pixels. Each zone has 14 diagonal lines, thus 14 sub-features are obtained from the each zone. These 14 sub-features values are averaged to form a single feature value and placed in the corresponding zone. Finally, 24 features are extracted for each character (Pradeep et al.
<xref ref-type="bibr" rid="CR33">2011</xref>
). The complete process is shown in Fig. 
<xref rid="Fig3" ref-type="fig">3</xref>
.</p>
</sec>
<sec id="Sec16">
<title>Mean</title>
<p>Mean gives an idea of what pixels color to choose to assess the color of the complete image. It is measured by taking sum of all the 1 pixels and dividing them by total number of pixels of each box. Thus 24 features of each image are obtained.</p>
<p>The mean for each box is commutated as:
<disp-formula id="Equ4">
<label>4</label>
<alternatives>
<tex-math id="M7">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Mean = \bar{X} = \frac{{\sum\limits_{i = 1}^{n} {X_{i} } }}{N}$$\end{document}</tex-math>
<mml:math id="M8" display="block">
<mml:mrow>
<mml:mi>M</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>n</mml:mi>
<mml:mo>=</mml:mo>
<mml:mover accent="true">
<mml:mrow>
<mml:mi>X</mml:mi>
</mml:mrow>
<mml:mrow>
<mml:mo stretchy="false">¯</mml:mo>
</mml:mrow>
</mml:mover>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mrow>
<mml:munderover>
<mml:mo movablelimits="false">∑</mml:mo>
<mml:mrow>
<mml:mi>i</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:mi>n</mml:mi>
</mml:munderover>
<mml:msub>
<mml:mi>X</mml:mi>
<mml:mi>i</mml:mi>
</mml:msub>
</mml:mrow>
<mml:mi>N</mml:mi>
</mml:mfrac>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ4.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
where N = Total number of pixels in each box.</p>
</sec>
<sec id="Sec17">
<title>Gradient operations</title>
<p>Image gradient is the variations of pixels in horizontal and vertical directions. Image gradient may be used to extract information from images. The gradients have been calculated by using the following formula.
<disp-formula id="Equ5">
<label>5</label>
<alternatives>
<tex-math id="M9">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Gradient = \nabla F = \frac{\nabla f}{\nabla x}i + \frac{\nabla f}{\nabla y}j$$\end{document}</tex-math>
<mml:math id="M10" display="block">
<mml:mrow>
<mml:mi>G</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>d</mml:mi>
<mml:mi>i</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>n</mml:mi>
<mml:mi>t</mml:mi>
<mml:mo>=</mml:mo>
<mml:mi mathvariant="normal">∇</mml:mi>
<mml:mi>F</mml:mi>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mrow>
<mml:mi mathvariant="normal">∇</mml:mi>
<mml:mi>f</mml:mi>
</mml:mrow>
<mml:mrow>
<mml:mi mathvariant="normal">∇</mml:mi>
<mml:mi>x</mml:mi>
</mml:mrow>
</mml:mfrac>
<mml:mi>i</mml:mi>
<mml:mo>+</mml:mo>
<mml:mfrac>
<mml:mrow>
<mml:mi mathvariant="normal">∇</mml:mi>
<mml:mi>f</mml:mi>
</mml:mrow>
<mml:mrow>
<mml:mi mathvariant="normal">∇</mml:mi>
<mml:mi>y</mml:mi>
</mml:mrow>
</mml:mfrac>
<mml:mi>j</mml:mi>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ5.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
</p>
<p>∆
<italic>f</italic>
/∆
<italic>x</italic>
 = Gradient in the x-direction for pixel at (i, j). ∆
<italic>f</italic>
/∆
<italic>y</italic>
 = Gradient in the y-direction for pixel at (i, j).</p>
<p>Gradient have been measured for each box. This process is sequentially repeated for all the 24 boxes. Thus 48 features are extracted from 24 boxes.</p>
</sec>
<sec id="Sec18">
<title>Standard deviation</title>
<p>Standard deviation (SD) of pixels in each box has been calculated. This process is applied for all 24 boxes to obtain 24 features.
<disp-formula id="Equ6">
<label>6</label>
<alternatives>
<tex-math id="M11">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SD = \sigma = \sqrt {\frac{1}{N}\sum\nolimits_{i}^{N} {(x_{i} - \bar{x})^{2} } }$$\end{document}</tex-math>
<mml:math id="M12" display="block">
<mml:mrow>
<mml:mi>S</mml:mi>
<mml:mi>D</mml:mi>
<mml:mo>=</mml:mo>
<mml:mi mathvariant="italic">σ</mml:mi>
<mml:mo>=</mml:mo>
<mml:msqrt>
<mml:mrow>
<mml:mfrac>
<mml:mn>1</mml:mn>
<mml:mi>N</mml:mi>
</mml:mfrac>
<mml:msubsup>
<mml:mo>∑</mml:mo>
<mml:mrow>
<mml:mi>i</mml:mi>
</mml:mrow>
<mml:mi>N</mml:mi>
</mml:msubsup>
<mml:msup>
<mml:mrow>
<mml:mo stretchy="false">(</mml:mo>
<mml:msub>
<mml:mi>x</mml:mi>
<mml:mi>i</mml:mi>
</mml:msub>
<mml:mo>-</mml:mo>
<mml:mover accent="true">
<mml:mrow>
<mml:mi>x</mml:mi>
</mml:mrow>
<mml:mrow>
<mml:mo stretchy="false">¯</mml:mo>
</mml:mrow>
</mml:mover>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
<mml:mn>2</mml:mn>
</mml:msup>
</mml:mrow>
</mml:msqrt>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ6.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
where N is number of rows and
<inline-formula id="IEq1">
<alternatives>
<tex-math id="M13">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\overline{x}$$\end{document}</tex-math>
<mml:math id="M14">
<mml:mover>
<mml:mi>x</mml:mi>
<mml:mo>¯</mml:mo>
</mml:mover>
</mml:math>
<inline-graphic xlink:href="40064_2016_1775_Article_IEq1.gif"></inline-graphic>
</alternatives>
</inline-formula>
is mean.</p>
</sec>
<sec id="Sec19">
<title>Center of gravity (CG)</title>
<p>Centre of gravity of the pixels in each box is obtained. In this process we get total 48 features for all the 24 boxes
<disp-formula id="Equ7">
<label>7</label>
<alternatives>
<tex-math id="M15">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CG:x_{o} = \frac{{\sum\nolimits_{i = 1}^{n} {\sum\nolimits_{j = 1}^{m} {iB(i,j)} } }}{a},y_{o} = \frac{{\sum\nolimits_{i = 1}^{n} {\sum\nolimits_{j = 1}^{m} {jB(i,j)} } }}{a}$$\end{document}</tex-math>
<mml:math id="M16" display="block">
<mml:mrow>
<mml:mi>C</mml:mi>
<mml:mi>G</mml:mi>
<mml:mo>:</mml:mo>
<mml:msub>
<mml:mi>x</mml:mi>
<mml:mi>o</mml:mi>
</mml:msub>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mrow>
<mml:msubsup>
<mml:mo>∑</mml:mo>
<mml:mrow>
<mml:mi>i</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:mi>n</mml:mi>
</mml:msubsup>
<mml:mrow>
<mml:msubsup>
<mml:mo>∑</mml:mo>
<mml:mrow>
<mml:mi>j</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:mi>m</mml:mi>
</mml:msubsup>
<mml:mrow>
<mml:mi>i</mml:mi>
<mml:mi>B</mml:mi>
<mml:mo stretchy="false">(</mml:mo>
<mml:mi>i</mml:mi>
<mml:mo>,</mml:mo>
<mml:mi>j</mml:mi>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
</mml:mrow>
</mml:mrow>
<mml:mi>a</mml:mi>
</mml:mfrac>
<mml:mo>,</mml:mo>
<mml:msub>
<mml:mi>y</mml:mi>
<mml:mi>o</mml:mi>
</mml:msub>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mrow>
<mml:msubsup>
<mml:mo>∑</mml:mo>
<mml:mrow>
<mml:mi>i</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:mi>n</mml:mi>
</mml:msubsup>
<mml:mrow>
<mml:msubsup>
<mml:mo>∑</mml:mo>
<mml:mrow>
<mml:mi>j</mml:mi>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
</mml:mrow>
<mml:mi>m</mml:mi>
</mml:msubsup>
<mml:mrow>
<mml:mi>j</mml:mi>
<mml:mi>B</mml:mi>
<mml:mo stretchy="false">(</mml:mo>
<mml:mi>i</mml:mi>
<mml:mo>,</mml:mo>
<mml:mi>j</mml:mi>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
</mml:mrow>
</mml:mrow>
<mml:mi>a</mml:mi>
</mml:mfrac>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ7.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
where
<italic>B</italic>
(
<italic>i</italic>
,
<italic>j</italic>
) is Binary Region and is given as:
<inline-formula id="IEq2">
<alternatives>
<tex-math id="M17">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$B\left( {i,j} \right) = 1 if \left( {i,j} \right)\,in \,region\, or\, 0 \,otherwise.$$\end{document}</tex-math>
<mml:math id="M18">
<mml:mrow>
<mml:mi>B</mml:mi>
<mml:mfenced close=")" open="(" separators="">
<mml:mrow>
<mml:mi>i</mml:mi>
<mml:mo>,</mml:mo>
<mml:mi>j</mml:mi>
</mml:mrow>
</mml:mfenced>
<mml:mo>=</mml:mo>
<mml:mn>1</mml:mn>
<mml:mi>i</mml:mi>
<mml:mi>f</mml:mi>
<mml:mfenced close=")" open="(" separators="">
<mml:mrow>
<mml:mi>i</mml:mi>
<mml:mo>,</mml:mo>
<mml:mi>j</mml:mi>
</mml:mrow>
</mml:mfenced>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>i</mml:mi>
<mml:mi>n</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>r</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>g</mml:mi>
<mml:mi>i</mml:mi>
<mml:mi>o</mml:mi>
<mml:mi>n</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>o</mml:mi>
<mml:mi>r</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mn>0</mml:mn>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>o</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>h</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>w</mml:mi>
<mml:mi>i</mml:mi>
<mml:mi>s</mml:mi>
<mml:mi>e</mml:mi>
<mml:mo>.</mml:mo>
</mml:mrow>
</mml:math>
<inline-graphic xlink:href="40064_2016_1775_Article_IEq2.gif"></inline-graphic>
</alternatives>
</inline-formula>
a = Area.</p>
</sec>
<sec id="Sec20">
<title>Edge detection</title>
<p>Edge detection is performed on each box and finds the sum of values at edge pixel positions. Sobel edge detection function has been used for this step. The magnitude of the gradient of the image is calculated for each pixel position in the image according to the following formula
<disp-formula id="Equ8">
<label>8</label>
<alternatives>
<tex-math id="M19">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Mag(\nabla F) = \sqrt {(F_{x} )^{2} + (F_{y} )^{2} }$$\end{document}</tex-math>
<mml:math id="M20" display="block">
<mml:mrow>
<mml:mi>M</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>g</mml:mi>
<mml:mrow>
<mml:mo stretchy="false">(</mml:mo>
<mml:mi mathvariant="normal">∇</mml:mi>
<mml:mi>F</mml:mi>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
<mml:mo>=</mml:mo>
<mml:msqrt>
<mml:mrow>
<mml:msup>
<mml:mrow>
<mml:mo stretchy="false">(</mml:mo>
<mml:msub>
<mml:mi>F</mml:mi>
<mml:mi>x</mml:mi>
</mml:msub>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
<mml:mn>2</mml:mn>
</mml:msup>
<mml:mo>+</mml:mo>
<mml:msup>
<mml:mrow>
<mml:mo stretchy="false">(</mml:mo>
<mml:msub>
<mml:mi>F</mml:mi>
<mml:mi>y</mml:mi>
</mml:msub>
<mml:mo stretchy="false">)</mml:mo>
</mml:mrow>
<mml:mn>2</mml:mn>
</mml:msup>
</mml:mrow>
</mml:msqrt>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ8.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
where
<italic>F</italic>
<sub>
<italic>x</italic>
</sub>
 = Gradient in x direction;
<italic>F</italic>
<sub>
<italic>y</italic>
</sub>
 = Gradient in y direction. This process is applied for all 24 boxes to get 24 features.</p>
<p>Thus with the help of all the seven feature extraction methods we obtain 240 number of features for each image matrix.</p>
</sec>
<sec id="Sec21">
<title>Hybrid features</title>
<p>We have obtained a set of 240 features from seven different approaches. Instead of using a single approach we have combined all features obtained by different methods. This combination forms hybrid feature set and is stored in a feature vector. This vector contains a total of 240 features set and is used for recognizing the character. Table 
<xref rid="Tab2" ref-type="table">2</xref>
specifies the number of features extracted by using various feature extraction techniques.
<table-wrap id="Tab2">
<label>Table 2</label>
<caption>
<p>Features description</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Feature set</th>
<th align="left">Number of features</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Box features</td>
<td char="." align="char">48</td>
</tr>
<tr>
<td align="left">Diagonal distance</td>
<td char="." align="char">24</td>
</tr>
<tr>
<td align="left">Mean</td>
<td char="." align="char">24</td>
</tr>
<tr>
<td align="left">Gradient operations</td>
<td char="." align="char">48</td>
</tr>
<tr>
<td align="left">Standard deviation</td>
<td char="." align="char">24</td>
</tr>
<tr>
<td align="left">Centre of gravity</td>
<td char="." align="char">48</td>
</tr>
<tr>
<td align="left">Edge detection</td>
<td char="." align="char">24</td>
</tr>
<tr>
<td align="left">Total</td>
<td char="." align="char">240</td>
</tr>
</tbody>
</table>
</table-wrap>
</p>
</sec>
</sec>
<sec id="Sec22">
<title>Feature selection</title>
<p>In this work GA has been used for feature selection. Figure 
<xref rid="Fig4" ref-type="fig">4</xref>
shows the basic steps of feature selection subsystem.GA is implemented on hybrid features to create a subset of best features. Feature selection can be classified into two main approaches: filter and wrapper approaches. In filter approach evaluation is normally done independently of any classifier. We have used wrapper approach for implementing GA to evaluate feature subset. Wrapper approach employs classifier’s predictive accuracy for the evaluation of the subset of features. The classifier used here is ANN, which imparts a hybrid nature to the feature selection process.
<fig id="Fig4">
<label>Fig. 4</label>
<caption>
<p>Flow chart for feature extraction and feature selection sub system</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig4_HTML" id="MO13"></graphic>
</fig>
</p>
<sec id="Sec23">
<title>Procedure to prune the feature set dimensionally</title>
<p>For initial population we have chosen four chromosomes where size of each chromosome is equal to the dimensionality of the feature set, which is equal to 240 as we have extracted 240 features set in the feature extraction step. In a chromosome, 1 represents presence of that particular dimension and 0 represents absence of the same (see Fig. 
<xref rid="Fig5" ref-type="fig">5</xref>
). Bit string encoding of each chromosome has been shown in Fig. 
<xref rid="Fig6" ref-type="fig">6</xref>
.
<fig id="Fig5">
<label>Fig. 5</label>
<caption>
<p>Initial populations</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig5_HTML" id="MO14"></graphic>
</fig>
<fig id="Fig6">
<label>Fig. 6</label>
<caption>
<p>Bit string encoding</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig6_HTML" id="MO15"></graphic>
</fig>
</p>
<p>For each chromosome we have formed the relevant feature set say F1, F2, F3 and F4 (see Table 
<xref rid="Tab3" ref-type="table">3</xref>
).
<table-wrap id="Tab3">
<label>Table 3</label>
<caption>
<p>Relevant features of chromosome</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Chromosome</th>
<th align="left">Relevant feature</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">1</td>
<td align="left">F1</td>
</tr>
<tr>
<td align="left">2</td>
<td align="left">F2</td>
</tr>
<tr>
<td align="left">3</td>
<td align="left">F3</td>
</tr>
<tr>
<td align="left">4</td>
<td align="left">F4</td>
</tr>
</tbody>
</table>
</table-wrap>
</p>
<p>By using these relevant feature set, we performed training of Neural Network and calculated the accuracy which has been considered as the fitness function of that particular chromosome. Then we have calculated the average of fitness values of all the chromosomes. This process continues until convergence. To get the new sets of chromosomes the fittest chromosomes undergo reproduction, cross over and mutation process. Again for each new chromosome we form relevant feature set and train the Neural Network and calculate the accuracy using Eq. (
<xref rid="Equ9" ref-type="">9</xref>
), which is the fitness function. The average of all the fitness function has been calculated. After convergence the relevant feature set from the chromosome (fittest one) has been chosen to prune the feature dimension. The various parameters of GA have been shown in Table 
<xref rid="Tab4" ref-type="table">4</xref>
.
<table-wrap id="Tab4">
<label>Table 4</label>
<caption>
<p>Parameters of GA</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">GA parameters</th>
<th align="left">Value</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Elitism size</td>
<td align="left">1</td>
</tr>
<tr>
<td align="left">Population size</td>
<td align="left">240</td>
</tr>
<tr>
<td align="left">Initial population</td>
<td align="left">4</td>
</tr>
<tr>
<td align="left">Selection method</td>
<td align="left">Roulette-wheel</td>
</tr>
<tr>
<td align="left">Crossover probability</td>
<td align="left">0.8</td>
</tr>
<tr>
<td align="left">Crossover point</td>
<td align="left">Random</td>
</tr>
<tr>
<td align="left">Mutation probability</td>
<td align="left">0.01</td>
</tr>
<tr>
<td align="left">Generation number</td>
<td align="left">Till convergence</td>
</tr>
</tbody>
</table>
</table-wrap>
<disp-formula id="Equ9">
<label>9</label>
<alternatives>
<tex-math id="M21">\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\% Accuracy = Fitness \,function = \frac{Number\,of\,correctly \,recognised \,character}{Total\, number\, of\, character }$$\end{document}</tex-math>
<mml:math id="M22" display="block">
<mml:mrow>
<mml:mo>%</mml:mo>
<mml:mi>A</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>u</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>y</mml:mi>
<mml:mo>=</mml:mo>
<mml:mi>F</mml:mi>
<mml:mi>i</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>n</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>s</mml:mi>
<mml:mi>s</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>f</mml:mi>
<mml:mi>u</mml:mi>
<mml:mi>n</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>i</mml:mi>
<mml:mi>o</mml:mi>
<mml:mi>n</mml:mi>
<mml:mo>=</mml:mo>
<mml:mfrac>
<mml:mrow>
<mml:mi>N</mml:mi>
<mml:mi>u</mml:mi>
<mml:mi>m</mml:mi>
<mml:mi>b</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>r</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>o</mml:mi>
<mml:mi>f</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>c</mml:mi>
<mml:mi>o</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>l</mml:mi>
<mml:mi>y</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>r</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>o</mml:mi>
<mml:mi>g</mml:mi>
<mml:mi>n</mml:mi>
<mml:mi>i</mml:mi>
<mml:mi>s</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>d</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>c</mml:mi>
<mml:mi>h</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>r</mml:mi>
</mml:mrow>
<mml:mrow>
<mml:mi>T</mml:mi>
<mml:mi>o</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>l</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>n</mml:mi>
<mml:mi>u</mml:mi>
<mml:mi>m</mml:mi>
<mml:mi>b</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>r</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>o</mml:mi>
<mml:mi>f</mml:mi>
<mml:mspace width="0.166667em"></mml:mspace>
<mml:mi>c</mml:mi>
<mml:mi>h</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>r</mml:mi>
<mml:mi>a</mml:mi>
<mml:mi>c</mml:mi>
<mml:mi>t</mml:mi>
<mml:mi>e</mml:mi>
<mml:mi>r</mml:mi>
</mml:mrow>
</mml:mfrac>
</mml:mrow>
</mml:math>
<graphic xlink:href="40064_2016_1775_Article_Equ9.gif" position="anchor"></graphic>
</alternatives>
</disp-formula>
</p>
<p>Following are the steps for training of Neural Network:
<list list-type="bullet">
<list-item>
<p>Extract the features</p>
</list-item>
<list-item>
<p>Pruning the feature dimension</p>
</list-item>
<list-item>
<p>Split the feature sets into training and validation sets.</p>
</list-item>
<list-item>
<p>Choose the network configuration which gives largest validation accuracy.</p>
</list-item>
<list-item>
<p>In our case the model of the neural network has two hidden layers having 100 and 90 neurons respectively.</p>
</list-item>
<list-item>
<p>Store the model.</p>
</list-item>
</list>
</p>
</sec>
</sec>
</sec>
</sec>
<sec id="Sec24">
<title>Classification and recognition</title>
<p>ANN in recent years has proved to be an advance tool in solving classification or pattern recognition problems.ANN is an efficient information processing system which resembles a biological neural network. ANNs are robust and fault tolerant in nature. They process information in parallel and they learn by examples. In this work, to recognize handwritten English character a multilayer perceptron has been used. The architecture used for Neural Network is arranged in layers and so the model is termed as multilayer perceptron. It consists of a layer of input nodes, one or more layers of hidden nodes, and one layer of output nodes. We have designed a function for the ANN classifier which provides us with optimization in terms of selecting the best structure for MLP. Figure 
<xref rid="Fig7" ref-type="fig">7</xref>
presents the pseudo code for selecting best structured MLP. The best structure includes number of hidden layers and number of neurons present in hidden layers. The best structure is found by calculating the accuracy on validation set in each iteration and storing the best one each time. The learning algorithm used in the network is back-propagation gradient descent algorithm.
<fig id="Fig7">
<label>Fig. 7</label>
<caption>
<p>Pseudo code for selection of best architecture MLP</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig7_HTML" id="MO17"></graphic>
</fig>
</p>
<p>Following are the steps for testing of Neural Network:
<list list-type="bullet">
<list-item>
<p>Extract the features.</p>
</list-item>
<list-item>
<p>Pruning the feature dimension</p>
</list-item>
<list-item>
<p>Restore the model.</p>
</list-item>
<list-item>
<p>Get the accuracy on test set.</p>
</list-item>
</list>
</p>
<sec id="Sec25">
<title>Experimental set up</title>
<p>During the phase of feature selection process a multi layered feed forward back-propagation neural network having two hidden layers with architecture 240-100-90-6 is used. The length of the feature vector decides the number of neurons in the input layer. The recognition system consists of a three-layered feed forward NN (Hanmandlu et al.
<xref ref-type="bibr" rid="CR17">2003</xref>
; Hanmandlu and Murthy
<xref ref-type="bibr" rid="CR15">2007</xref>
; Pradeep et al.
<xref ref-type="bibr" rid="CR33">2011</xref>
) with an input of 240 features extracted using:
<list list-type="bullet">
<list-item>
<p>Box method</p>
</list-item>
<list-item>
<p>Diagonal distance approach</p>
</list-item>
<list-item>
<p>Mean</p>
</list-item>
<list-item>
<p>Gradient operations approach</p>
</list-item>
<list-item>
<p>Standard deviation</p>
</list-item>
<list-item>
<p>Centre of gravity</p>
</list-item>
<list-item>
<p>Edge detection</p>
</list-item>
</list>
</p>
<p>There are 52 characters (26 for small and 26 for capital alphabets) to be classified. Hence 6 bit variable is required for the output. Thus the number of output neurons is six. Log sigmoidal (logsig) is the activation function for neurons in hidden and output layers.</p>
<p>It is very difficult to choose the number of hidden layers and number of neurons in hidden layer in the architecture of neural network. Generally, for most of the applications, one hidden layer is enough, but the best way to choose the number of neurons and number of hidden layer is experimentation. Typically number of neurons in a hidden layer is determined by a combination of previous expertise, amount of data available, dimensionality, complexity of the problem, trial and error or validation on an additional training dataset (Polikar
<xref ref-type="bibr" rid="CR32">2006</xref>
). In our work different Neural Networks have been trained with different numbers of hidden layer and hidden neurons and measure the performance of those networks as explained in Fig. 
<xref rid="Fig7" ref-type="fig">7</xref>
.</p>
<p>We retained the configuration that yield the best performing network. The accuracy on validation set saturated for the configuration chosen. Table 
<xref rid="Tab5" ref-type="table">5</xref>
shows the best MLP architecture for the feature selection and Table 
<xref rid="Tab6" ref-type="table">6</xref>
shows the training parameters for the feature selection that produces the best results in our case. Figure 
<xref rid="Fig8" ref-type="fig">8</xref>
shows the structure of feed-forward neural network.
<table-wrap id="Tab5">
<label>Table 5</label>
<caption>
<p>MLP architecture for feature selection</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">No. of input nodes</th>
<th align="left">No. of hidden layers</th>
<th align="left">No. of nodes in hidden layer 1</th>
<th align="left">No. of nodes in hidden layer 2</th>
<th align="left">No of output nodes</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">240</td>
<td align="left">2</td>
<td align="left">100</td>
<td align="left">90</td>
<td align="left">6</td>
</tr>
</tbody>
</table>
</table-wrap>
<table-wrap id="Tab6">
<label>Table 6</label>
<caption>
<p>The training parameters of MLP for feature selection</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Training algorithm</th>
<th align="left">Performance function</th>
<th align="left">Training epochs</th>
<th align="left">Learning rate</th>
<th align="left">Momentum</th>
<th align="left">Training goal</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">traingdx</td>
<td align="left">Mean square error</td>
<td align="left">1000 or till convergence</td>
<td align="left">Adaptive</td>
<td align="left">0.9</td>
<td align="left">0.0001</td>
</tr>
</tbody>
</table>
</table-wrap>
<fig id="Fig8">
<label>Fig. 8</label>
<caption>
<p>Three layer neural network</p>
</caption>
<graphic xlink:href="40064_2016_1775_Fig8_HTML" id="MO18"></graphic>
</fig>
</p>
<p>The structure of the MLP is adaptive in nature. During the classification phase of the proposed work the number of input layer neurons have been changed to 76 as the number of features have been reduced to 76 after the implementation of GA and rest of the architecture of the MLP remains same.</p>
</sec>
</sec>
<sec id="Sec26">
<title>Result and discussion</title>
<p>The architecture of the neural network consists of an input layer, two hidden layers and an output layer. The number of neurons are 240, 100, 90 and 6 respectively for input, two hidden and output layer respectively during the feature selection and the number of input neurons during classification stage is 76. The neural network is trained using gradient descent back-propagation algorithm with momentum and adaptive learning rate. The log sigmiodal activation function is used to calculate the output. 19,145 known dataset of alphabet is used to train the network. After training the network, the system was tested using 2183 unknown alphabet dataset. In this proposed recognition system seven different approaches of feature extraction have been used. Feature selection is mostly done by heuristic or by intuition for specific type of character recognition application. The results are summarised in Tables 
<xref rid="Tab7" ref-type="table">7</xref>
,
<xref rid="Tab8" ref-type="table">8</xref>
,
<xref rid="Tab9" ref-type="table">9</xref>
,
<xref rid="Tab10" ref-type="table">10</xref>
,
<xref rid="Tab11" ref-type="table">11</xref>
and
<xref rid="Tab12" ref-type="table">12</xref>
. As seen from the Table 
<xref rid="Tab7" ref-type="table">7</xref>
, the number of features has been reduced to 76 by using GA. Table 
<xref rid="Tab8" ref-type="table">8</xref>
shows the final features combination that provides the best results. It has been observed in practice that the feature selection process results in reduced features with slightly degradation in performance (Kim and Kim
<xref ref-type="bibr" rid="CR14">2000</xref>
). Hence the process of feature selection is applied where efficiency in terms of speed along with space requirement are important, despite the recognition accuracy been deteriorated. This specially happens in case of small features dimensional problem where as for larger dimensional feature space accuracy improves. Table 
<xref rid="Tab9" ref-type="table">9</xref>
shows the accuracy of the proposed system is 94.65 % for capital alphabet and 91.11 % for small alphabet and it also shows that the accuracy of the proposed system is greater than the original one, where all the features have been considered.
<table-wrap id="Tab7">
<label>Table 7</label>
<caption>
<p>Number of features</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Networks</th>
<th align="left">Number of features</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Adaptive MLP classifier without feature reduction</td>
<td align="left">240</td>
</tr>
<tr>
<td align="left">Adaptive MLP classifier with feature reduction</td>
<td align="left">76</td>
</tr>
</tbody>
</table>
</table-wrap>
<table-wrap id="Tab8">
<label>Table 8</label>
<caption>
<p>Final feature combination for best results</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Feature set</th>
<th align="left">Number of features</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Box features</td>
<td char="." align="char">23</td>
</tr>
<tr>
<td align="left">Diagonal distance</td>
<td char="." align="char">7</td>
</tr>
<tr>
<td align="left">Mean</td>
<td char="." align="char">9</td>
</tr>
<tr>
<td align="left">Gradient operations</td>
<td char="." align="char">13</td>
</tr>
<tr>
<td align="left">Standard deviation</td>
<td char="." align="char">12</td>
</tr>
<tr>
<td align="left">Centre of gravity</td>
<td char="." align="char">8</td>
</tr>
<tr>
<td align="left">Edge detection</td>
<td char="." align="char">4</td>
</tr>
</tbody>
</table>
</table-wrap>
<table-wrap id="Tab9">
<label>Table 9</label>
<caption>
<p>% Accuracy</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Networks</th>
<th align="left">% Accuracy for capital alphabet</th>
<th align="left">% Accuracy for small alphabet</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Adaptive MLP classifier without feature reduction</td>
<td char="." align="char">91.56</td>
<td char="." align="char">87.49</td>
</tr>
<tr>
<td align="left">Adaptive MLP classifier with feature reduction</td>
<td char="." align="char">94.65</td>
<td char="." align="char">91.11</td>
</tr>
</tbody>
</table>
</table-wrap>
<table-wrap id="Tab10">
<label>Table 10</label>
<caption>
<p>Computational time</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Networks</th>
<th align="left">Testing time (m sec)</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Adaptive MLP classifier without feature reduction</td>
<td align="left">43</td>
</tr>
<tr>
<td align="left">Adaptive MLP classifier with feature reduction</td>
<td align="left">25</td>
</tr>
</tbody>
</table>
</table-wrap>
<table-wrap id="Tab11">
<label>Table 11</label>
<caption>
<p>Comparisons with other methods without feature selection</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Authors</th>
<th align="left">Dataset</th>
<th align="left">Feature extraction method</th>
<th align="left">Classifier</th>
<th align="left">% Accuracy capital alphabet</th>
<th align="left">% Accuracy small alphabet</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left">Murthy and Hanmandlu (
<xref ref-type="bibr" rid="CR34">2011</xref>
)</td>
<td align="left">Samples From Matlab</td>
<td align="left">Directional features</td>
<td align="left">Fuzzy logic</td>
<td char="." align="char">83.85</td>
<td align="left"></td>
</tr>
<tr>
<td align="left"> Nasien et al. (
<xref ref-type="bibr" rid="CR28">2010</xref>
)</td>
<td align="left">NIST</td>
<td align="left">Freeman chain code</td>
<td align="left">Support vector machine</td>
<td char="." align="char">88.46</td>
<td char="." align="char">86.00</td>
</tr>
<tr>
<td align="left">Ali et al. (
<xref ref-type="bibr" rid="CR1">2010</xref>
)</td>
<td align="left">–</td>
<td align="left">Wavelet compression</td>
<td align="left">Euclidean distance</td>
<td char="." align="char">89.68</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">Hanmandlu et al. (
<xref ref-type="bibr" rid="CR16">1999</xref>
)</td>
<td align="left">–</td>
<td align="left">Fusion method</td>
<td align="left">Neural network</td>
<td char="." align="char">86.00</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">Singh and Hewitt (
<xref ref-type="bibr" rid="CR36">2000</xref>
)</td>
<td align="left">CEDAR</td>
<td align="left">Hough transform</td>
<td align="left">LDA
<break></break>
Nearest neighbour</td>
<td char="." align="char">67.3
<break></break>
63.5</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">Vamvakas et al. (
<xref ref-type="bibr" rid="CR38">2009</xref>
)</td>
<td align="left">CEDAR</td>
<td align="left">Structural features</td>
<td align="left">SVM</td>
<td char="." align="char">80.19</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">Ganpathy and Liew (
<xref ref-type="bibr" rid="CR12">2008</xref>
)</td>
<td align="left">Self Created</td>
<td align="left">–</td>
<td align="left">Multiscale neural network</td>
<td char="." align="char">87.22</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">Yuan et al. (
<xref ref-type="bibr" rid="CR39">2012</xref>
)</td>
<td align="left">UNIPEN</td>
<td align="left"></td>
<td align="left">Convolution Neural network</td>
<td char="." align="char">93.7</td>
<td char="." align="char">90.2</td>
</tr>
<tr>
<td align="left">Our approach</td>
<td align="left">CEDAR</td>
<td align="left">Hybrid features</td>
<td align="left">Neural network</td>
<td char="." align="char">91.56</td>
<td char="." align="char">87.49</td>
</tr>
</tbody>
</table>
</table-wrap>
<table-wrap id="Tab12">
<label>Table 12</label>
<caption>
<p>Comparison with other methods using feature selection</p>
</caption>
<table frame="hsides" rules="groups">
<thead>
<tr>
<th align="left">Authors</th>
<th align="left">Dataset</th>
<th align="left">Feature extraction method</th>
<th align="left">Feature selection</th>
<th align="left">Classifier</th>
<th align="left">% Accuracy capital alphabet</th>
<th align="left">% Accuracy small alphabet</th>
</tr>
</thead>
<tbody>
<tr>
<td align="left" rowspan="2">Chung and Yoon (
<xref ref-type="bibr" rid="CR7">1997</xref>
)</td>
<td align="left" rowspan="2">NIST</td>
<td align="left">Gradient</td>
<td align="left" rowspan="2">PCA</td>
<td align="left" rowspan="2">Neural network</td>
<td char="." align="char">93.31</td>
<td align="left">–</td>
</tr>
<tr>
<td align="left">UDLRH</td>
<td char="." align="char">90.85</td>
<td align="left">–</td>
</tr>
<tr>
<td align="left" rowspan="3">De Stefano et al. (
<xref ref-type="bibr" rid="CR10">2013</xref>
)</td>
<td align="left" rowspan="3">NIST</td>
<td align="left" rowspan="3">Hybrid features</td>
<td align="left" rowspan="3">GA</td>
<td align="left">SVM</td>
<td char="." align="char">57.88</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">MLP</td>
<td char="." align="char">62.57</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">KNN</td>
<td char="." align="char">61.45</td>
<td align="left"></td>
</tr>
<tr>
<td align="left">Proposed method</td>
<td align="left">CEDAR</td>
<td align="left">Hybrid features</td>
<td align="left">GA</td>
<td align="left">MLP NN</td>
<td char="." align="char">94.65</td>
<td align="left">91.11</td>
</tr>
</tbody>
</table>
</table-wrap>
</p>
<p>The proposed system performs well as it uses less number of features and implements an adaptive MLPNN classifier. Feature selection using GA is more proficient for features with larger dimension (240 in our case) than smaller dimension. This shows that redundant features are negatively contributing in accuracy of the classifiers.</p>
<p>As shown in Table 
<xref rid="Tab10" ref-type="table">10</xref>
, the computational time to recognize the test samples of handwritten character recognition has been reduced from 43 to 25 ms as the number of features reduces from 240 to 76. And also the proposed system requires less storage space. Reduction in features dimension is important as it improves the recognition speed (Kim and Kim
<xref ref-type="bibr" rid="CR14">2000</xref>
).</p>
<p>Comparisons of our handwritten character recognition method without using feature selection methodology with other existing methods which have not used feature selection have been shown in Table 
<xref rid="Tab11" ref-type="table">11</xref>
. Here an attempt is made to attain high accuracy using an adaptive MLPNN classifier. It is clear from the results that our method outperforms the other state of art methods with an accuracy of 91.56 and 87.49 % respectively for capital alphabet and small alphabet respectively except the work presented by Yuan et al. (
<xref ref-type="bibr" rid="CR39">2012</xref>
). The details and number of training and testing set are not clear in their work but in our case we have considered whole training and testing dataset in the experiment. This superior adaptive structured MLP performance is due to the superior generalization ability of MLP in high dimensional space.</p>
<p>Table 
<xref rid="Tab12" ref-type="table">12</xref>
gives the comparative analysis of the works which are using different feature selections techniques. It is evident from the results that the proposed method which implies Genetic Algorithm for feature selection gives better results compared to other existing methodologies.</p>
</sec>
<sec id="Sec27">
<title>Conclusions</title>
<p>This paper presents hybrid feature extraction and GA based feature selection for off-line handwritten character recognition by using adaptive MLPNN classifier for achieving an improved overall performance on real world recognition problems. Seven approaches of feature extraction namely, box method, diagonal distance method, mean and gradient operation, standard deviation, centre of gravity and edge detection have been used to develop an off-line character recognition system. Two different recognition networks are built namely Adaptive MLP classifier without feature reduction and Adaptive MLP classifier with feature reduction. The network is trained and tested on the CEDAR CDROM-1 dataset. It can be concluded from the experimental results that the network which uses GA based feature selection method improves over all performance of the recognition system in terms of speed and storage requirement. It has also been verified that the proposed adaptive MLP Neural Network works as a better classifier and provides better accuracy.</p>
</sec>
</body>
<back>
<ack>
<title>Authors’ contributions</title>
<p>All the research work and experimental analysis has been done by first author and the layout and formatting of the manuscript has been done by co-author. Both authors read and approved the final manuscript.</p>
<sec id="FPar1">
<title>Acknowledgements</title>
<p>We are very thankful to CEDAR, USA for their help as and when required. We would also like to thank Dr. Rangaraja for giving a guide line in finding and using the database of CEDAR CDROM-1.</p>
</sec>
<sec id="FPar2">
<title>Competing interests</title>
<p>The authors declare that they have no competing interests.</p>
</sec>
</ack>
<ref-list id="Bib1">
<title>References</title>
<ref id="CR1">
<mixed-citation publication-type="other">Ali H, Ali Y, Hossain E, Sultana S (2010) Character recognition using wavelet compression. In: Proceedings of 13th international conference on computer and information technology (ICCIT), Bangladesh, pp 452–457, 23–25 December 2010</mixed-citation>
</ref>
<ref id="CR2">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Arica</surname>
<given-names>N</given-names>
</name>
<name>
<surname>Yarman-Vural</surname>
<given-names>FT</given-names>
</name>
</person-group>
<article-title>An overview of character recognition focused on off-line handwriting</article-title>
<source>IEEE Trans Syst Man Cybern Part C Appl Rev</source>
<year>2001</year>
<volume>31</volume>
<issue>2</issue>
<fpage>216</fpage>
<lpage>233</lpage>
<pub-id pub-id-type="doi">10.1109/5326.941845</pub-id>
</element-citation>
</ref>
<ref id="CR3">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Avi-Itzhak</surname>
<given-names>HI</given-names>
</name>
<name>
<surname>Diep</surname>
<given-names>TA</given-names>
</name>
<name>
<surname>Garland</surname>
<given-names>H</given-names>
</name>
</person-group>
<article-title>High accuracy optical character recognition using neural network with centroid dithering</article-title>
<source>IEEE Trans Pattern Anal Mach Intell</source>
<year>1995</year>
<volume>17</volume>
<issue>2</issue>
<fpage>218</fpage>
<lpage>223</lpage>
<pub-id pub-id-type="doi">10.1109/34.368165</pub-id>
</element-citation>
</ref>
<ref id="CR4">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Chen</surname>
<given-names>MY</given-names>
</name>
<name>
<surname>Kundu</surname>
<given-names>A</given-names>
</name>
<name>
<surname>Zhou</surname>
<given-names>J</given-names>
</name>
</person-group>
<article-title>Off-line handwritten word recognition using a hidden Markov model type stochastic network</article-title>
<source>IEEE Trans Pattern Anal Mach Intell</source>
<year>1994</year>
<volume>16</volume>
<fpage>481</fpage>
<lpage>496</lpage>
<pub-id pub-id-type="doi">10.1109/34.291449</pub-id>
</element-citation>
</ref>
<ref id="CR5">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Cho</surname>
<given-names>SB</given-names>
</name>
</person-group>
<article-title>Pattern recognition with neural network combined by genetic algorithm</article-title>
<source>Fuzzy Sets Syst</source>
<year>1999</year>
<volume>13</volume>
<fpage>339</fpage>
<lpage>347</lpage>
<pub-id pub-id-type="doi">10.1016/S0165-0114(98)00232-2</pub-id>
</element-citation>
</ref>
<ref id="CR6">
<mixed-citation publication-type="other">Chouaib H, Cloppet F, Vincent N (2012) Fast feature selection for handwritten digit recognition. In: International conference on frontiers in handwriting recognition, ICFHR Italy, pp 485–490</mixed-citation>
</ref>
<ref id="CR7">
<mixed-citation publication-type="other">Chung K, Yoon J (1997) Performance comparison of several feature selection methods based on node pruning in handwritten character recognition. In: Proceedings of fourth international conference on document analysis and recognition, Ulm, Germany, pp 11–15</mixed-citation>
</ref>
<ref id="CR8">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Das</surname>
<given-names>N</given-names>
</name>
<name>
<surname>Sarkar</surname>
<given-names>R</given-names>
</name>
<name>
<surname>Basu</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Kundu</surname>
<given-names>M</given-names>
</name>
<name>
<surname>Nasipuri</surname>
<given-names>M</given-names>
</name>
<name>
<surname>Basu</surname>
<given-names>KD</given-names>
</name>
</person-group>
<article-title>A genetic algorithm based region sampling for selection of local features in handwritten digit recognition application</article-title>
<source>Appl Soft Comput</source>
<year>2012</year>
<volume>12</volume>
<fpage>1592</fpage>
<lpage>1606</lpage>
<pub-id pub-id-type="doi">10.1016/j.asoc.2011.11.030</pub-id>
</element-citation>
</ref>
<ref id="CR9">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>De Stefano</surname>
<given-names>C</given-names>
</name>
<name>
<surname>Della Cioppa</surname>
<given-names>A</given-names>
</name>
<name>
<surname>Marcelli</surname>
<given-names>A</given-names>
</name>
</person-group>
<article-title>Character pre-classification based on genetic programming</article-title>
<source>Pattern Recogn Lett</source>
<year>2002</year>
<volume>23</volume>
<fpage>1439</fpage>
<lpage>1448</lpage>
<pub-id pub-id-type="doi">10.1016/S0167-8655(02)00104-6</pub-id>
</element-citation>
</ref>
<ref id="CR10">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>De Stefano</surname>
<given-names>C</given-names>
</name>
<name>
<surname>Fontanella</surname>
<given-names>F</given-names>
</name>
<name>
<surname>Marrocco</surname>
<given-names>C</given-names>
</name>
<name>
<surname>Scotto di Freca</surname>
<given-names>A</given-names>
</name>
</person-group>
<article-title>A GA-based feature selection approach with an application to handwritten character recognition</article-title>
<source>Pattern Recogn Lett</source>
<year>2013</year>
<volume>35</volume>
<fpage>130</fpage>
<lpage>141</lpage>
<pub-id pub-id-type="doi">10.1016/j.patrec.2013.01.026</pub-id>
</element-citation>
</ref>
<ref id="CR11">
<mixed-citation publication-type="other">Feng J, Yang Y, Wang H, Wang X-M (2004) Feature selection based on genetic algorithms and support vector machines for handwritten similar Chinese characters recognition. In: Proceedings of the third international conference on machine learning and cybernetics, Shanghai. pp 3600–3605, 26–29 August 2004</mixed-citation>
</ref>
<ref id="CR12">
<mixed-citation publication-type="other">Ganpathy V, Liew KL (2008) Handwritten character recognition using multiscale neural network training technique. In: Proceedings of world academy of science engineering and technology PWASET, vol 29, pp 32–37</mixed-citation>
</ref>
<ref id="CR13">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Govindan</surname>
<given-names>K</given-names>
</name>
<name>
<surname>Shivaprasad</surname>
<given-names>AP</given-names>
</name>
</person-group>
<article-title>Character recognition—a review</article-title>
<source>Pattern Recogn</source>
<year>1990</year>
<volume>23</volume>
<issue>7</issue>
<fpage>671</fpage>
<lpage>683</lpage>
<pub-id pub-id-type="doi">10.1016/0031-3203(90)90091-X</pub-id>
</element-citation>
</ref>
<ref id="CR14">
<mixed-citation publication-type="other">Kim G, Kim S (2000) Feature selection using genetic algorithm for handwritten character recognition. In: Proceedings of the seventh International conference on frontiers in handwriting recognition, Amsterdam, pp 103–112, 11–13 September 2000</mixed-citation>
</ref>
<ref id="CR15">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Hanmandlu</surname>
<given-names>M</given-names>
</name>
<name>
<surname>Murthy</surname>
<given-names>OVR</given-names>
</name>
</person-group>
<article-title>Fuzzy model based recognition of handwritten numerals</article-title>
<source>Pattern Recogn</source>
<year>2007</year>
<volume>40</volume>
<fpage>1840</fpage>
<lpage>1854</lpage>
<pub-id pub-id-type="doi">10.1016/j.patcog.2006.08.014</pub-id>
</element-citation>
</ref>
<ref id="CR16">
<mixed-citation publication-type="other">Hanmandlu M, Mohan KRM, Kumar H (1999) Neural based handwritten character recognition. In: Proceeding of fifth international conference on document analysis and recognition, ICDAR-99, Bangalore, pp 241–244, 20–22 September 1999</mixed-citation>
</ref>
<ref id="CR17">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Hanmandlu</surname>
<given-names>M</given-names>
</name>
<name>
<surname>Mohan</surname>
<given-names>KRM</given-names>
</name>
<name>
<surname>Chakraborty</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Goyal</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Choudhury</surname>
<given-names>DR</given-names>
</name>
</person-group>
<article-title>Unconstrained handwritten character recognition based on fuzzy logic</article-title>
<source>Pattern Recogn</source>
<year>2003</year>
<volume>36</volume>
<fpage>603</fpage>
<lpage>623</lpage>
<pub-id pub-id-type="doi">10.1016/S0031-3203(02)00069-9</pub-id>
</element-citation>
</ref>
<ref id="CR18">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Hull</surname>
<given-names>JJ</given-names>
</name>
</person-group>
<article-title>A database for handwritten text recognition research</article-title>
<source>IEEE Trans Pattern Anal Mach Intell</source>
<year>1995</year>
<volume>16</volume>
<issue>5</issue>
<fpage>550</fpage>
<lpage>554</lpage>
<pub-id pub-id-type="doi">10.1109/34.291440</pub-id>
</element-citation>
</ref>
<ref id="CR19">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Katiyar</surname>
<given-names>G</given-names>
</name>
<name>
<surname>Mehfuz</surname>
<given-names>S</given-names>
</name>
</person-group>
<article-title>Evolutionary computing techniques in off-line handwritten character recognition: a review</article-title>
<source>UACEE Int J Comput Sci Appl</source>
<year>2012</year>
<volume>1</volume>
<fpage>133</fpage>
<lpage>137</lpage>
</element-citation>
</ref>
<ref id="CR20">
<mixed-citation publication-type="other">Lanzi PL (1997) Fast feature selection with genetic algorithms: a filter approach. In: Proceeding of IEEE international conference on evolutionary computation, Indianapolis, IN. pp 537–540</mixed-citation>
</ref>
<ref id="CR21">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Lee</surname>
<given-names>SW</given-names>
</name>
</person-group>
<article-title>Multi cluster neural network for totally unconstrained handwritten numeral recognition</article-title>
<source>Neural Netw</source>
<year>1995</year>
<volume>8</volume>
<issue>5</issue>
<fpage>783</fpage>
<lpage>792</lpage>
<pub-id pub-id-type="doi">10.1016/0893-6080(95)00020-Z</pub-id>
</element-citation>
</ref>
<ref id="CR22">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Lee</surname>
<given-names>S-W</given-names>
</name>
</person-group>
<article-title>Off-line recognition of totally unconstrained handwritten numerals using multilayer cluster neural network</article-title>
<source>IEEE Trans Pattern Anal Mach Intell</source>
<year>1996</year>
<volume>18</volume>
<issue>6</issue>
<fpage>648</fpage>
<lpage>652</lpage>
<pub-id pub-id-type="doi">10.1109/34.506415</pub-id>
</element-citation>
</ref>
<ref id="CR23">
<mixed-citation publication-type="other">Lee S-W, Kim YJ (1994) Off-line recognition of totally unconstrained handwritten numerals using multilayer cluster neural network. In: Proceedings of 12th IAPR international conference on computer vision and image processing. Jerusalem, pp 507–509</mixed-citation>
</ref>
<ref id="CR24">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Liang</surname>
<given-names>Z</given-names>
</name>
<name>
<surname>Xia</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Zhou</surname>
<given-names>Y</given-names>
</name>
<name>
<surname>Zhang</surname>
<given-names>L</given-names>
</name>
<name>
<surname>Li</surname>
<given-names>Y</given-names>
</name>
</person-group>
<article-title>Feature extraction based on Lp-norm generalized principal component analysis</article-title>
<source>Pattern Recognit Lett</source>
<year>2013</year>
<volume>34</volume>
<fpage>1037</fpage>
<lpage>1045</lpage>
<pub-id pub-id-type="doi">10.1016/j.patrec.2013.01.030</pub-id>
</element-citation>
</ref>
<ref id="CR25">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Mehfuz</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Katiyar</surname>
<given-names>G</given-names>
</name>
</person-group>
<article-title>Intelligent system for off-line handwritten character recognition: a review</article-title>
<source>Int J Emerg Technol Adv Eng</source>
<year>2012</year>
<volume>2</volume>
<fpage>538</fpage>
<lpage>543</lpage>
</element-citation>
</ref>
<ref id="CR26">
<mixed-citation publication-type="other">Shayegan MA, Chan CS (2012) A new approach to feature selection in handwritten farsi/arabic character recognition. In: International conference on advanced computer science applications and technologies (ACSAT), Kuala Lumpur, pp 506–511</mixed-citation>
</ref>
<ref id="CR27">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Mori</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Suen</surname>
<given-names>CY</given-names>
</name>
<name>
<surname>Yamamoto</surname>
<given-names>K</given-names>
</name>
</person-group>
<article-title>Historical review of OCR research and development</article-title>
<source>Proc IEEE</source>
<year>1992</year>
<volume>80</volume>
<fpage>1029</fpage>
<lpage>1057</lpage>
<pub-id pub-id-type="doi">10.1109/5.156468</pub-id>
</element-citation>
</ref>
<ref id="CR28">
<mixed-citation publication-type="other">Nasien D, Haron H, Yuhaniz SS (2010) Support Vector Machine (SVM) for English handwritten character recognition. In: Second international conference on computer engineering and applications, pp 249–252</mixed-citation>
</ref>
<ref id="CR29">
<mixed-citation publication-type="other">Oliveira LSR, Saburin R, Bortolozzi F, Suen CY (2002) Feature selection using multi-objective genetic algorithms for handwritten digit recognition. IEEE, pp 568–571</mixed-citation>
</ref>
<ref id="CR30">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Parkins</surname>
<given-names>AD</given-names>
</name>
<name>
<surname>Nandi</surname>
<given-names>AK</given-names>
</name>
</person-group>
<article-title>Genetic programming techniques for handwritten digit recognition</article-title>
<source>Signal Process</source>
<year>2004</year>
<volume>84</volume>
<fpage>2345</fpage>
<lpage>2365</lpage>
<pub-id pub-id-type="doi">10.1016/j.sigpro.2004.07.027</pub-id>
</element-citation>
</ref>
<ref id="CR31">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Plamondon</surname>
<given-names>R</given-names>
</name>
<name>
<surname>Srihari</surname>
<given-names>SN</given-names>
</name>
</person-group>
<article-title>On-line and off-line handwriting recognition: a comprehensive survey</article-title>
<source>IEEE Trans Pattern Anal Mach Intell</source>
<year>2000</year>
<volume>22</volume>
<issue>1</issue>
<fpage>63</fpage>
<lpage>84</lpage>
<pub-id pub-id-type="doi">10.1109/34.824821</pub-id>
</element-citation>
</ref>
<ref id="CR32">
<mixed-citation publication-type="other">Polikar R (2006) Pattern recognition. Rowan University Glassboro, New Jersey, Wiley encyclopaedia of biomedical engineering. Wiley, New York, pp 1–22</mixed-citation>
</ref>
<ref id="CR33">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Pradeep</surname>
<given-names>J</given-names>
</name>
<name>
<surname>Srinivasn</surname>
<given-names>E</given-names>
</name>
<name>
<surname>Himavathi</surname>
<given-names>S</given-names>
</name>
</person-group>
<article-title>Diagonal based feature extraction for handwritten alphabet recognition using neural network</article-title>
<source>Int J Comput Sci Technol</source>
<year>2011</year>
<volume>3</volume>
<issue>1</issue>
<fpage>27</fpage>
<lpage>38</lpage>
</element-citation>
</ref>
<ref id="CR34">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Murthy</surname>
<given-names>OVR</given-names>
</name>
<name>
<surname>Hanmandlu</surname>
<given-names>M</given-names>
</name>
</person-group>
<article-title>Interactive fuzzy model based recognition of handwritten characters</article-title>
<source>J Pattern Recognit Res</source>
<year>2011</year>
<volume>2</volume>
<fpage>154</fpage>
<lpage>165</lpage>
<pub-id pub-id-type="doi">10.13176/11.245</pub-id>
</element-citation>
</ref>
<ref id="CR35">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Shrivastava</surname>
<given-names>S</given-names>
</name>
<name>
<surname>Singh</surname>
<given-names>MP</given-names>
</name>
</person-group>
<article-title>Performance evaluation of feed-forward neural network with soft computing techniques for hand written English alphabets</article-title>
<source>Appl Soft Comput</source>
<year>2011</year>
<volume>11</volume>
<fpage>1156</fpage>
<lpage>1182</lpage>
<pub-id pub-id-type="doi">10.1016/j.asoc.2010.02.015</pub-id>
</element-citation>
</ref>
<ref id="CR36">
<mixed-citation publication-type="other">Singh S, Hewitt (2000) Cursive digit and character recognition on cedar database. In: Proceedings of 15th international conference on pattern recognition, Barcelona, 3–8 September 2000, vol 2. IEEE Press, New York, pp 569–572</mixed-citation>
</ref>
<ref id="CR37">
<element-citation publication-type="journal">
<person-group person-group-type="author">
<name>
<surname>Trier</surname>
<given-names>OD</given-names>
</name>
<name>
<surname>Jain</surname>
<given-names>AK</given-names>
</name>
<name>
<surname>Taxt</surname>
<given-names>J</given-names>
</name>
</person-group>
<article-title>feature extraction methods for character recognition: a survey</article-title>
<source>Pattern Recogn</source>
<year>1996</year>
<volume>29</volume>
<issue>4</issue>
<fpage>641</fpage>
<lpage>662</lpage>
<pub-id pub-id-type="doi">10.1016/0031-3203(95)00118-2</pub-id>
</element-citation>
</ref>
<ref id="CR38">
<mixed-citation publication-type="other">Vamvakas G, Gatos B, Perantonis SJ (2009) A novel feature extraction and classification methodology for the recognition of historical documents. In: IEEE 10th International conference on document analysis and recognition, pp 491–495</mixed-citation>
</ref>
<ref id="CR39">
<mixed-citation publication-type="other">Yuan A, Bai G, Jiao L, Liu Y (2012) Offline handwritten english character recognition based on convolution neural network. In: 10th IAPR International workshop on document analysis systems</mixed-citation>
</ref>
<ref id="CR40">
<mixed-citation publication-type="other">Zeng X, Chen Y-W (2009) Feature selection using recursive feature elimination for handwritten digit recognition. In: Proceedings of fifth international conference on intelligent information hiding and multimedia signal processing. pp 1205–1208</mixed-citation>
</ref>
</ref-list>
</back>
</pmc>
<affiliations>
<list>
<country>
<li>Inde</li>
</country>
</list>
<tree>
<country name="Inde">
<noRegion>
<name sortKey="Katiyar, Gauri" sort="Katiyar, Gauri" uniqKey="Katiyar G" first="Gauri" last="Katiyar">Gauri Katiyar</name>
</noRegion>
<name sortKey="Mehfuz, Shabana" sort="Mehfuz, Shabana" uniqKey="Mehfuz S" first="Shabana" last="Mehfuz">Shabana Mehfuz</name>
</country>
</tree>
</affiliations>
</record>

Pour manipuler ce document sous Unix (Dilib)

EXPLOR_STEP=$WICRI_ROOT/Ticri/CIDE/explor/OcrV1/Data/Ncbi/Merge

HfdSelect -h $EXPLOR_STEP/biblio.hfd -nk 000258 | SxmlIndent | more

Ou

HfdSelect -h $EXPLOR_AREA/Data/Ncbi/Merge/biblio.hfd -nk 000258 | SxmlIndent | more

Pour mettre un lien sur cette page dans le réseau Wicri

{{Explor lien
   |wiki=    Ticri/CIDE
   |area=    OcrV1
   |flux=    Ncbi
   |étape=   Merge
   |type=    RBID
   |clé=     PMC:4801835
   |texte=   A hybrid recognition system for off-line handwritten characters
}}

Pour générer des pages wiki

HfdIndexSelect -h $EXPLOR_AREA/Data/Ncbi/Merge/RBID.i   -Sk "pubmed:27066370" \
       | HfdSelect -Kh $EXPLOR_AREA/Data/Ncbi/Merge/biblio.hfd   \
       | NlmPubMed2Wicri -a OcrV1
Wicri

This area was generated with Dilib version V0.6.32.
Data generation: Sat Nov 11 16:53:45 2017. Site generation: Mon Mar 11 23:15:16 2024