14th International Conference on Frontiers in ...

Proceedings

14th International Conference on Frontiers in Handwriting Recognition ICFHR 2014

1-4 September 2014 Hersonissos, Crete Island, Greece

IEEE Computer Society Technical & Conference Activities Board T&C Board Vice President Cecilia Metra Università di Bologna, Italy

IEEE Computer Society Staff Evan Butterfield, Director of Products and Services Lynne Harris, CMP, Senior Manager, Conference Support Services Patrick Kellenberger, Supervisor, Conference Publishing Services IEEE Computer Society Publications The world-renowned IEEE Computer Society publishes, promotes, and distributes a wide variety of authoritative computer science and engineering texts. These books are available from most retail outlets. Visit the CS Store at http://www.computer.org/portal/site/store/index.jsp for a list of products.

IEEE Computer Society Conference Publishing Services (CPS) The IEEE Computer Society produces conference publications for more than 300 acclaimed international conferences each year in a variety of formats, including books, CD-ROMs, USB Drives, and on-line publications. For information about the IEEE Computer Society’s Conference Publishing Services (CPS), please e-mail: [email protected] or telephone +1-714-821-8380. Fax +1-714-761-1784. Additional information about Conference Publishing Services (CPS) can be accessed from our web site at: http://www.computer.org/cps Revised: 18 January 2012

CPS Online is our innovative online collaborative conference publishing system designed to speed the delivery of price quotations and provide conferences with real-time access to all of a project's publication materials during production, including the final papers. The CPS Online workspace gives a conference the opportunity to upload files through any Web browser, check status and scheduling on their project, make changes to the Table of Contents and Front Matter, approve editorial changes and proofs, and communicate with their CPS editor through discussion forums, chat tools, commenting tools and e-mail. The following is the URL link to the CPS Online Publishing Inquiry Form:

http://www.computer.org/portal/web/cscps/quote

Copyright © 2014 by The Institute of Electrical and Electronics Engineers, Inc. All rights reserved. Copyright and Reprint Permissions: Abstracting is permitted with credit to the source. Libraries may photocopy beyond the limits of US copyright law, for private use of patrons, those articles in this volume that carry a code at the bottom of the first page, provided that the per-copy fee indicated in the code is paid through the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923. Other copying, reprint, or republication requests should be addressed to: IEEE Copyrights Manager, IEEE Service Center, 445 Hoes Lane, P.O. Box 133, Piscataway, NJ 08855-1331. The papers in this book comprise the proceedings of the meeting mentioned on the cover and title page. They reflect the authors’ opinions and, in the interests of timely dissemination, are published as presented and without change. Their inclusion in this publication does not necessarily constitute endorsement by the editors, the IEEE Computer Society, or the Institute of Electrical and Electronics Engineers, Inc. IEEE Computer Society Order Number P5170 BMS Part # CFP14311-PRT ISBN # 978-1-4799-4334-0 ISSN # 2167-6445 Additional copies may be ordered from: IEEE Computer Society Customer Service Center 10662 Los Vaqueros Circle P.O. Box 3014 Los Alamitos, CA 90720-1314 Tel: + 1 800 272 6657 Fax: + 1 714 821 4641 http://computer.org/cspress [email protected]

IEEE Service Center 445 Hoes Lane P.O. Box 1331 Piscataway, NJ 08855-1331 Tel: + 1 732 981 0060 Fax: + 1 732 981 9667 http://shop.ieee.org/store/ [email protected]

IEEE Computer Society Asia/Pacific Office Watanabe Bldg., 1-4-2 Minami-Aoyama Minato-ku, Tokyo 107-0062 JAPAN Tel: + 81 3 3408 3118 Fax: + 81 3 3408 3553 [email protected]

Individual paper REPRINTS may be ordered at:

Editorial production by Juan E. Guerrero Cover art production by Mark J. Bartosik Printed in the United States of America by Applied Digital Imaging

IEEE Com puter Society

Conference Publishing Services (CPS) http://www.computer.org/cps

14th International Conference on Frontiers in Handwriting Recognition

ICFHR 2014 Table of Contents Message from Conference and Program Chairs...................................................................................xvii Conference Organization..........................................................................................................................xix Program Committee...................................................................................................................................xx Additional Reviewers................................................................................................................................xxi Invited Talks..............................................................................................................................................xxii

Session 1 – Word Spotting A Simple and Fast Word Spotting Method ....................................................................................................3 Alon Kovalchuk, Lior Wolf, and Nachum Dershowitz Segmentation-Based Historical Handwritten Word Spotting Using Document-Specific Local Features ...............................................................................................................9 Konstantinos Zagoris, Ioannis Pratikakis, and Basilis Gatos An Historical Handwritten Arabic Dataset for Segmentation-Free Word Spotting - HADARA80P ..............................................................................................................................15 Werner Pantke, Martin Dennhardt, Daniel Fecker, Volker Märgner, and Tim Fingscheidt

Session 2 – Document Image Pre-processing Techniques Text/Non-text Classification in Online Handwritten Documents with Recurrent Neural Networks .........................................................................................................................................23 Truyen Van Phan and Masaki Nakagawa Handwritten/Printed Text Separation Using Pseudo-Lines for Contextual Re-labeling ..................................................................................................................................................29 Ahmad Montaser Awal, Abdel Belaïd, and Vincent Poulain D'Andecy

v

Document Writer Analysis with Rejection for Historical Arabic Manuscripts ............................................743 Daniel Fecker, Abedelkadir Asi, Werner Pantke, Volker Märgner, Jihad El-Sana, and Tim Fingscheidt Graph Based Re-ranking Method with Application to Handwritten Digits .................................................749 Foteini Fotopoulou and George Economou Multiple Training - One Test Methodology for Handwritten Word-Script Identification ..............................................................................................................................................754 Miguel A. Ferrer, Aythami Morales, Nayara Rodríguez, and Umapada Pal An Application of LBF Energy in Image/Video Frame Text Detection ......................................................760 V.N. Manjunath Aradhya and M.S. Pavithra Ground-Truth and Metric for the Evaluation of Arabic Handwritten Character Segmentation ............................................................................................................................................766 Yousef Elarian, Abdelmalek Zidouri, and Wasfi Al-Khatib Recognizing Glagolitic Characters in Degraded Historical Documents ....................................................771 Sajid Saleem, Fabian Hollaus, Markus Diem, and Robert Sablatnig

Competitions ICFHR 2014 Competition on Handwritten Digit String Recognition in Challenging Datasets (HDSRC 2014) ...................................................................................................779 Markus Diem, Stefan Fiel, Florian Kleber, Robert Sablatnig, Jose M. Saavedra, David Contreras, Juan Manuel Barrios, and Luiz S. Oliveira ICFHR2014 Competition on Handwritten Text Recognition on Transcriptorium Datasets (HTRtS) .....................................................................................................................................785 Joan Andreu Sánchez, Verónica Romero, Alejandro H. Toselli, and Enrique Vidal ICFHR 2014 Competition on Recognition of On-Line Handwritten Mathematical Expressions (CROHME 2014) ...........................................................................................791 H. Mouchère, C. Viard-Gaudin, R. Zanibbi, and U. Garain ICFHR2014 Competition on Arabic Writer Identification Using AHTID/MW and KHATT Databases .............................................................................................................................797 Fouad Slimane, Sameh Awaida, Anis Mezghani, Mohammad Tanvir Parvez, Slim Kanoun, Sabri A. Mahmoud, and Volker Märgner ICFHR2014 Competition on Word Recognition from Historical Documents: ANncestry Word REcognition from Segmented Historical Documents (ANWRESH) .............................................................................................................................................803 Jackson Reese, Michael Murdock, Shawn Reid, and Blaine Hamilton ICFHR2014 Competition on Handwritten Document Image Binarization (H-DIBCO 2014) .......................................................................................................................................809 Konstantinos Ntirogiannis, Basilis Gatos, and Ioannis Pratikakis

xv

2014 14th International Conference on Frontiers in Handwriting Recognition

MultipleTraining – One Test methodology for Handwritten Word-Script Identification Miguel A. Ferrer, Aythami Morales, Nayara Rodríguez

Umapada Pal

Instituto Universitario para el Desarrollo Tecnológico y la Innovación en Comunicaciones Universidad de Las Palmas de Gran Canaria, Spain Email: {mferrer,amorales}@idetic.eu

Computer Vision and Pattern Recognition Unit Indian Statistical Institute Kolkata, India E-mail: [email protected]

Abstract—Script identification is an important area in handwriting document image analysis field. The script identification at word level on documents written in multiple scripts is an open challenge for the scientific community and a real concern in countries with multiple official languages, e. g. the country like India. Such documents usually contain two scripts: the most of the document are written in the regional script while some words, acronyms or numbers are written in Roman script. In this case a word or even a character level script identification is required to locate the second script characters in the document. Here the major problem is the few script descriptors available for the script estimation which convey high error rates. The literatures try to address this problem by looking for more efficient descriptors. In this paper we propose a Multiple Training – One Test technique to alleviate this problem. Several classifiers are trained, each one with words of similar amount of information. A scale invariable word information index is defined for this sake. To identify the script of a query word, its word information index is worked out, and its script is identified with the most appropriate classifier. Accuracy improvements has been obtained with this promising technique, especially for the shorten words. Keywords: Handwritten Script Identification, training, Texture descriptors, Document Analysis.

I.

eight different script types (Latin, Chinese, Japanese, Greek, Cyrillic, Hebrew, Sanskrit, and Farsi) with a minimum script recognition error reported of 2.1%. A new step is given in [4] using rotation invariant uniform local binary patterns features. This proposal uses block size texture analysis where the block size guaranties an enough number of descriptors for accurate script characterization. Therefore it is not useful for multi script documents where a single small word could be written in a different script. The line-wise script identifier was proposed in [5]. In this case the script appearance is given by the frequency of the different stroke directions featured by the histogram of the Local Binary Patterns (LBP) codes. A Least Square Support Vector Machine (LS-SVM) is used as classifier. The procedure is evaluated with a freely available database which include ten scripts: Arab, Bangla, Greek, Gujarati, Hindi, Japanese, Kannada, Roman, Russian and Urdu. Another example of line-wise script detection can be found in [6] which proposes to distinguish between Kashmiri and Roman scripts based on knowledge models. Recent papers have addressed the word-wise script identification. For instance [7] tries to identify the script of words based on heuristically script oriented features. Recently, word-wise script recognition has been applied to video frames looking for automatic indexing of such videos. This application faces the low quality problem which convey on biased features with a procedure that include a preprocessing step such as skeleton extraction or superresolution. The script features are on Gabor Filters, Zernike moments and Gradient directional features. [8] Uses the video database with horizontal run statistics and wavelets features for distinguishing five Indian scripts: namely Hindi, Kannada, Roman, Malayalam and Tamil. Both [7] and [8] train their systems with samples obtained from the same database. This paper focuses on word-wise script identification in multi-script documents facing the problem of identifying the script of short words. It is a common problem in India where bi-scripts or tri-scripts documents are very common. They contain usually a local script with a few short words written in Roman script such as acronyms, dates, addresses, numbers, etc. An accurate location of them is required for an eventual posterior automatic processing of the document, i.e. to apply optical character recognition.

Multiple

INTRODUCTION

Worldwide, there are many different languages and many different scripts in which these languages are typeset. In a document image processing system that includes automated document recognition capabilities, previous detection of the script or scripts present in a document is required for choosing the proper optical character recognition. State of the art of script identification extracts attributes from connected components of individual characters such as upward concavities, optical densities, character height densities, top and bottom line profiles, etc. These approaches could be called local approaches as they require the analysis of individual components [1]. Appearance technique based on texture for script identification was proposed at 1998 by Tan [2] using a multichannel Gabor filter with an adaptive GMM. Busch et al [3] extended the work to new texture features as: Gray-Level Co-Occurrence Matrix, Gabor Energy, Wavelet Energy, etc. The proposed algorithm was tested on a database containing 2167-6445/14 $31.00 © 2014 IEEE DOI 10.1109/ICFHR.2014.132

754

The major problem of this application is the few amount of script descriptors than can be extracted from a single short word. Few descriptors convey a biased script stroke statistics which will mislead the script estimation given by the classifier. The script identification literature tries to face this problem looking for more efficient script features. Otherwise, in this paper we propose a Multiple Training – One Test (MTOT) procedure. This procedure tests a query word with the classifier trained with words that include a similar amount of information. With this aim, a measure that estimate the amount of information included in a word for script identification is proposed. It is called word information index (‫)݅݅ݓ‬. Concisely, several classifiers are trained, each one with words with a similar ‫݅݅ݓ‬. The script estimation of a query word is done in two steps: the ‫ ݅݅ݓ‬of the query word is obtained and its script is estimated with the classifiers trained for such a ‫ ݅݅ݓ‬. The descriptors used for the script identifications are worked out using the Local Binary Patterns proposed in [5] The rest of the paper is organized as follows: Section II describes the scripts databases; Section III presents the script characterization and identification while the Section IV proposes the multiple training procedure defining the word information index ‫ ݅݅ݓ‬. The Section V is devoted to the experiments and results. The paper closes with the conclusions at Section VI. II.

SCRIPT DATABASE

Figure 1. Example of text line segmentation for Roman and Bangla Scripts

The database includes three different scripts: Bangla, Persian and Roman. The Bangla handwriting contains handwritten documents from 26 different writers. The number of lines per writer is between 10 and 28 with a total of 515 lines. The mode of line per writer is 28. The Persian includes 42 writers. After segmenting the lines we obtained 537 lines being the minimum number of lines per writer 2 and the maximum 24. The mode of lines per writer is 24. The Roman dataset contains 90 writers in English language with a total of 857 different lines. The number of lines per writer ranges from 2 to 11 being the mode equal to 11 lines per writer. The Bangla and Persian documents have been provided by the Indian Statistical Institute while the Roman document comes from the IAM database, concisely are the first document of each writer [12]. The printer area of the IAM database documents were cropped automatically. Therefore the number of lines per script is balanced and the text of each line is different in order to avoid bias due to text or writer and focus the identifier on the script statistics. Each line contains just one script. The documents were scanned at 150 dpi and converted to black and white by means of an Otsu’s threshold. The lines have been segmented as follows [9]: each connected object of the image has been labeled and its convex hull worked out. The result is dilated horizontally in order to connect the objects belonging to the same line. Then the horizontal histogram is obtained and the line located as the maximums of the histogram. In each line a horizontal line is drawn in order to connect distant object of the same line. The next step is to extract line by line which has been done as follows:

Figure 2. Examples of lines database: Bangla (upper line), Persian (middle line) and Roman (lower line)

1. Select the top object of the dilated lines and workout its horizontal histogram. 2. If its histogram has a single maximum, it is supposed to be a single line and the dilated object is used as mask to segment the line. 3. If the object has several peaks and valleys, it is supposed that several lines have been mixed in the object. To separate them, we follow these two steps: a. The object is horizontally eroded until the top object contain single peak. b. The new top object is dilated to recover the original shape and it is used as mask to segment the top line.

755

TABLE I. NUMBER OF WORDS PER SCRIPT Number of

Bangla

Persian

Roman

Total

lines words

515 3526

537 4605

857 7350

1909 15481 (a) LBP image of the segmented line divided into 4 blocks for histogram calculation

c. The inclination of the mask is calculated and the line is rotated to reduce the line skew. The obtained image is cropped and saved. 4. The top line is deleted and control goes to step 1 until no more line is left for segmentation. An example of the result with Roman and Bangla scripts can be seen at Figure 1, while Figure 2 shows examples of segmented lines from Bangla, Persian and Roman documents. The number of lines extracted by script and database are summarized at Table I. The width of the line is obtained from the horizontal projection. The upper and lower bound of the line are established as the first and last index where the histogram is greater than the 20% of its maximum. This simple algorithm has shown to be effective for the all the scripts as it is shown at Figure 3. The word separation is accomplished by the vertical histogram. When the vertical histogram is zero, a word separation mark is added. The mark is deleted if the zero gap is shorter than a threshold that is proportional to adjacent word lengths. We are aware that this procedure could generate several false separations of word, but it is effective when it separates words of different scripts, e.g. Roman script words inserted in Bangla or Persian documents usually are well separate. Figure 3 shows an example of word segmentation. Table I displays the number of words calculated by script at each database. Although line segmentation and word separation are important steps for script detection, their accuracy is not a critical step. The script features we are using are statistical description of the stroke directions. If a word is separated into two by mistake, the worked out features would be less statistically significant. It may originate mistakes if the word parts are too short. Instead, if two words are joined the feature will be statistically a better description of the script. The only problem is to join two words of different scripts which rarely occur because the space between two words with different scripts usually is wide.

(b) LBP feature vector as block histograms concatenation Figure 4. Example of LBP feature vector for Persian

The LBP operator is defined as a gray level invariant texture measure in a local neighborhood. The original LBP operator labels the pixel of an image by thresholding the ͵ ൈ ͵ neighborhood of each pixel and concatenating the results binomially to form a number. Assume that a given image is defined as ‫ܫ‬ሺܼሻ ൌ ‫ܫ‬ሺ‫ݔ‬ǡ ‫ݕ‬ሻ. The LBP operator transforms the input image to ‫ܲܤܮ‬ሺܼሻ ൌ ‫ܲܤܮ‬ሺ‫ݔ‬ǡ ‫ݕ‬ሻ as follows: ‫ܲܤܮ‬ሺܼ௖ ሻ ൌ σ଻௣ୀ଴ ‫ݏ‬൫‫ܫ‬൫ܼ௣ ൯ െ ‫ܫ‬ሺܼ௖ ሻ൯ ή ʹ௣ , ͳ ݈൒Ͳ where ‫ݏ‬ሺ݈ ሻ ൌ ቄ is the unit step function and ‫ܫ‬൫ܼ௣ ൯ Ͳ ݈൏Ͳ is the 8-neighborhood around ‫ܫ‬ሺܼ௖ ሻ.There are other multiple implementations of the LBP algorithm including generalized and rotation invariant versions. However, deformations such as rotation, scale or translation can be considered moderate in script identification. Therefore, although other implementations have been tested, in this application the best results have been obtained using the basic implementation of the ‫ ܲܤܮ‬algorithm. The ‫ܲܤܮ‬ሺܼሻ code matrix contains information about the structure to which the pixel belongs: the stroke edge, stroke corners, stroke ends, etc. As the characteristic of the top, bottom and central area of the lines are different and distinctive of the script, the spatial distribution of local pattern is modeled by dividing the line into a number of adjacent regions and calculating the histogram in each region in order to do not loss the location of the different structures inside the image. After conducting several experiments, testing a range of smaller and greater region sizes, the best performance was obtained when dividing the image into 4 equal horizontal blocks which overlapped by 15%. For each ସ ௜ ൟ௜ୀଵ Ǥ Take block, we calculate the 255 bin histograms ൛݄௅஻௉ into account that the bin corresponding to the background has not been considered. An example can be seen in Figure 4. The histograms of all the blocks are concatenated in order to do not loss the spatial information obtaining the ௜ ห݅ ൌ ͳǡʹǡ ǥ ǡͶൟ of dimension 4ൈ feature vector ‫ܪ‬௅஻௉ ൌ ൛݄௅஻௉ ʹͷͷ ൌ ͳͲʹͲ. In order to obtain scale invariance the feature

III. SCRIPT CHARACTERIZATION Following the procedure suggested at [5], Local Binary Patterns, so called LBP, are used as script features [10]. They can be seen as a unifying approach to the traditionally statistical and structural approaches of texture analysis. Applied to black and white images, an LBP can be considered as the concatenation of the binary gradient directions. The histogram of these micro patterns contains information of the distribution of the edges, spots, and other local figures in an image which can be used as feature for script detection.

756

TABLE II. EXAMPLE OF WORD IMAGES WITH DIFFERENT ‫݅݅ݓ‬

‫݅݅ݓ‬

Bangla

Persian

the classifier trained with word of similar amount of information than the query word. The amount of information in the word for script identification is related with the number of descriptors (‫ܲܤܮ‬ codes) from which the script statistic feature is build (histogram). As the number of ‫ ܲܤܮ‬is word font size dependent, the number of ‫ ܲܤܮ‬codes is normalized by the line width for scale invariant obtaining the so called word information index defined by:

Roman

1 2 5 10

‫ ݅݅ݓ‬ൌ ෍

being ‫ܲܤܮ‬ሺ‫ݔ‬ǡ ‫ݕ‬ሻ de word image transformed to ‫ ܲܤܮ‬codes and ‫ ܹ݁݊݅ܮ‬the line width of the word. Another information theory metrics were checked as Entropy with worse results. To illustrate the relation between the word information index and the word length, Table II show a correspondence between different ‫ ݅݅ݓ‬and an example of word with each ‫݅݅ݓ‬ for Indian, Persian and Roman scripts. Concisely, the training procedure is performed as follows; i) the lines of the training document are segmented; ii) Each line is divided in non-overlapped frames containing the defined ‫ ݅݅ݓ‬value; iii) The classifier is trained with those frames; iv) The process is performed for several ‫ ݅݅ݓ‬values obtaining a classifier per ‫ ݅݅ݓ‬value. For testing, the ‫ ݅݅ݓ‬value of the query word is worked out and its script is estimated with the classifier trained for a similar ‫ ݅݅ݓ‬value. According to this methodology, the classifiers trained for the shorter words use characteristics with less statistical content. This drawback is compensate by the large training sets obtained when the lines are divided in short segments. Instead, the classifier with greatest wii are less trained but use more accurate features.

20 25 30 35 40 45 50 60 80

vector are normalized to area equal to one. So we have different version of representing the script structure density. As classifier we use a Least Square Support Vector Machine (LS-SVM) [11] with a RBF kernel: The LS-SVM makes binary decision, therefore the multi-class classification for script identification has been made by adopting the one-against all techniques. We carried out gridsearch on the hyper-parameters in the 10-fold cross validation for selecting the parameters on the training sequence. The parameters setting that produce the best crossvalidation accuracy were picked. IV.

ሼ‫ܲܤܮ‬ሺ‫ݔ‬ǡ ‫ݕ‬ሻሽȀ‫ܹ݁݊݅ܮ‬

௫ǡ௬

15

V.

EXPERIMENTS

The experiments have been conducted, firstly, to establish the best relation in term of minimum error between the ‫݅݅ݓ‬ of the words used for training and the ‫ ݅݅ݓ‬of the query word. Secondly to evaluate and compare the performance of the word script recognizer when the classifier is trained with the MTOT procedure or classical one. A. MultipleTraining – One Test set up The proposed multiple training procedure estimates the script of a query word using a classifier trained with a set of words with similar ‫݅݅ݓ‬. This section tries to establish, given a query word with a ‫݅݅ݓ‬, which is the best classifier to be used for the word script identification. This set up is used to reduce the number of trained classifiers. The experiment has been done as follows: the 30% of each script lines are used for training while the remaining are left for testing. Each line is divided in non-overlapped frames of different ‫ ݅݅ݓ‬ൌ ሼͳǡʹǡ͵ǡ ǥ ǡͳͲͲሽ. Note that the ‫ ݅݅ݓ‬average of the database lines is 143. For each ‫݅݅ݓ‬, a classifier is trained and used for the testing words of the test database with different ‫݅݅ݓ‬.The results are summarized at Table III in terms of Equal Error Rate (EER).

MULTIPLE TRAINING – ONE TEST CLASSIFIER

For word-wise script identification the major challenge is dealing with the small size of the word image which conveys a few descriptors to calculate the script statistics. Therefore, the feature estimator is strongly biased by the training words and the identification error depends of such training process. To alleviate this drawback we propose a multi training procedure. The basic idea is to train several classifiers and at the test time to select the more appropriate. The chosen criterion for classifier selection is the amount of information in the word. Therefore, several classifiers are trained, each one with words of similar amount of information. In this way each classifier learns from examples with the same characteristics without mixing accurate and biased features. At the test step, the script of the query word is estimated with

757

TABLE III. EER OF THE WORD SCRIPT IDENTIFIER FOR DIFFERENT TRAINING AND TESTING ‫݅݅ݓ‬. THE DARK CELLS HIGHLIGHT THE CHOSEN CLASIFFIERS

As can be seen, the Equal Error Rate depends on the relation between the ‫ ݅݅ݓ‬of the classifier and the ‫ ݅݅ݓ‬of the query word. Analyzing Table III, in order to reduce the number of classifier to be trained, the next rule is proposed: Switch (‫݅݅ݓ‬of the query/test word) case ‫ ݅݅ݓ‬൑ ͳ Select classifier trained with case ͳ ൏ ‫ ݅݅ݓ‬൑ Ͷ Select classifier trained with case Ͷ ൏ ‫ ݅݅ݓ‬൑ ͺ Select classifier trained with case ͻ ൏ ‫ ݅݅ݓ‬൑ ͳͺ Select classifier trained with case ͳͺ ൏ ‫ ݅݅ݓ‬൑ ʹͺ Select classifier trained with case ʹͺ ൏ ‫ ݅݅ݓ‬൑ ͵ͺ Select classifier trained with case 38൏ ‫ ݅݅ݓ‬൑ ͻͲ Select classifier trained with otherwise Select classifier trained with end

TABLE IV. SCRIPT IDENTIFICATION ACCURACY

Script identification of bi-script documents with Bangla Persian Bangla and and and Roman Roman Persian Exp.1: line-based 94.65% 82.55% 94.18% Exp. 2: word-based 95.00% 90.01% 96.78 % Exp. 3: MTOT-based 95.03% 90.73% 97.18% Tri-Script Identification: Bangla, Persian and Roman Exp.1: line-based 82.49% Exp. 2: word-based 89.32% Exp. 3: MTOT-based 89.89%

‫ ݅݅ݓ‬ൌ ͳ ‫ ݅݅ݓ‬ൌ ʹ ‫ ݅݅ݓ‬ൌ ͷ ‫ ݅݅ݓ‬ൌ ͳͲ ‫ ݅݅ݓ‬ൌ ʹͲ

TABLE V. NUMBER OF TIMES EACH TRAINED CLASSIFIER IS SELECTED DURING THE TEST

‫ ݅݅ݓ‬ൌ ͵Ͳ

‫݅݅ݓ‬ 1 2 5 10 20 30 40 60

‫ ݅݅ݓ‬ൌ ͶͲ ‫ ݅݅ݓ‬ൌ ͸Ͳ

In this case, training eight classifiers cover the most of the input words with a low EER. B. Applyig MTOT Three experiments have been designed to assess the multi training procedure performance compared with the classical ones. x Experiment 1 – line-based script identification: Train the classifier with the full training lines x Experiment 2 – word-based script identification: Segment the words of the training lines, and train the classifier with the words x Experiment 3– MTOT-based script identification: Train eight classifiers with frames of ‫ ݅݅ݓ‬equal to 1, 2, 5, 10, 20, 30, 40 and 60 respectively. At all the cases, the test is performed with the segmented words of the text lines.

Bangla 0% 0.17% 1.05% 20.87% 31.54% 21.58% 24.25% 0.54%%

Persian 0% 0.24% 9.45% 32.85% 23.30% 14.72% 16.57% 2.87%

Roman 0% 0.11% 4.14% 24.09% 27.82% 18.46% 24.37% 1.01%

TABLE VI. TRI-SCRIPT IDENTIFICATION ACCURACY FOR EACH  ‫ ݅݅ݓ‬AND EXPERIMENT

‫݅݅ݓ‬ 5 10 20 30 40 60

758

Experiment training based on Line Word MTOT 50.63% 59.27% 72.92% 68.69% 79.08% 86.31% 84.46% 92.40% 93.05% 91.09% 95.16% 94.91% 96.68% 96.55% 95.79% 99.49% 98.59% 99.49%

For each experiment, Table IV shows the script identification accuracy of discriminating scripts by pairs and trios. It can be seen that the MTOT procedure clearly outperforms the line-based and word-based script identification. For the sake of better understanding of the procedure, Table V shows the percentages of times each classifier is selected. It shows that most of the words segmented by our algorithm are classified using a ‫ ݅݅ݓ‬equal to 10 and 40.Table VI highlights that the improvement of the MTOT based method is due to the improvement of the performance with the shortest words. Note that ‫ ݅݅ݓ‬൏ ͳͷ could be considered identification at character level. Although they are not predominant in our dataset, in multilingual documents the short words appear more frequently in form of acronyms and numbers. In this case the advantage of the MTOT is significant. Computational requirement are obviously increased in the training step since instead of training just one model we train several ones. The major computational problem is training with low ‫ ݅݅ݓ‬words. Dividing the lines in short words means a greater number of training segments. By ones side it is good because the classifier considers more information which improves the performance but it takes longer time to training. In our case, for ‫ ݅݅ݓ‬ൌ ͳǡ ʹƒ†ͷ we have trained with the 40%, 60% and 80% of the segments chosen randomly. With these numbers the whole multiple training of the entire database is performed in 8 hours with an Intel i7 computer. The first 4 classifiers ( ‫ ݅݅ݓ‬ൌ ͳǡ ʹǡ ͷƒ†ͳͲ) takes the 80% of the time. Obviously, as the test only require to work out the ‫ ݅݅ݓ‬, both steps of the MTOT: work out ‫ ݅݅ݓ‬and classify, it requires the same computational time than usual approaches.

performance has improved with respect to train the classifiers with line-based or word-based script descriptors. Future works are addressed to improve the word information index definition and testing this procedure with different descriptors and databases. ACKNOWLEDGMENTS This study was funded by the Spanish government’s MCINN TEC2012-38630-C04-02 research project. REFERENCES [1]

D. Ghosh, T. Dube, A.P. Shivaprasad, "Script Recognition—A Review," IEEE Transactions on Pattern Analysis and Machine Intelligence,vol.32, no.12, pp.2142,2161, December 2010. [2] T.N. Tan, “Rotation invariant texture features and their use in automatic script identification",IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no.7, pp.751-756, July 1998. [3] Busch, W.W. Boles, S. Sridharan, "Texture for script identification", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.27, no.11, pp.1720-1732, November 2005. [4] JianJia Pan, YuanYan Tang, "A rotation-robust script identification based on BEMD and LBP", 2011 International Conference on Wavelet Analysis and Pattern Recognition, pp.165-170, 10-13 July 2011. [5] M.A. Ferrer, A. Morales, U. Pal, "LBP Based Line-Wise Script Identification", 12th International Conference on Document Analysis and Recognition, pp.369,373, 25-28 August 2013. [6] Bashir, Rumaan; Quadri, Smk, "Identification of Kashmiri script in a bilingual document image", IEEE Second International Conference on Image Information Processing, pp.575,579, 9-11 December 2013. [7] Das, M.S.; Rani, D.S.; Reddy, C. R K, "Heuristic based script identification from multilingual text documents",1st International Conference on Recent Advances in Information Technology, pp.487,492, 15-17 March 2012. [8] S.A. Angadi, M.M. Kodabagi, "A fuzzy approach for word level script identification of text in low resolution display board images using wavelet features", International Conference on Advances in Computing, Communications and Informatics, pp.1804,1811, 22-25 August 2013. [9] Z. Razak, K. Zulkiflee, M. Y. I. Idris, E. M. Tamil, M. N. Mohamed Noor, R. Salleh, M.Yaakob, Z. M. Yusof and M. Yaacob, “Off-line Handwriting Text Line Segmentation: A Review”, in International Journal of Computer Science and Network Security, v. 8, no.7, pp. 12-20, July 2008. [10] T. Mäenpää, M. Pietikäinen, “Texture Analysis with local binary Patterns”, in C.H. Chen, P.S.P. Wang (eds.): Handbook of Pattern Recognition and Computer Vision, 3rd edn. World Scientific, pp. 197216, 2005. [11] J. A. K. Suykens, T. V. Gestel, J. D. Brabanter, B. D. Moor, J. Vandewalle, Least Squares Support Vector Machines, World Scientific Publishing Co., Pte, Ltd, Singapore, 2002. [12] U. Marti and H. Bunke, “A full English sentence database for off-line handwriting recognition”, in Proceedings of the 5th International. Conference on Document Analysis and Recognition, pages 705 - 708, 1999.

VI. CONCLUSION This paper faced the problem of improving the script identification performance in the case of short words in multi script handwritten documents. A short word provides a few script descriptors producing a biased script feature which often mislead the classifier script estimation. This paper proposes to alleviate this problem by a Multiple Training – One Test technique. Different classifiers are trained for several word lengths. To avoid scale variability, a word information index is proposed to measure the word length. The script of each query word is estimated by the classifier trained with the word of similar amount of information. This procedure has been tested with a script identifier based on the histogram of Local Binary Patterns (LBP) for stroke direction characterization and Least Square Support Vector Machine (LS-SVM) as classifier. The

759

Author Index Adak, Chandranath........................................... 643 Adam, S. .......................................................... 411 Ahmad, Irfan...................................................... 537 Akao, Yoshinori................................................. 110 Alaei, Alireza..................................................... 216 Alewijnse, Linda.................................................. 73 Alimi, Adel M. ........................................... 335, 696 Al-Khatib, Wasfi................................................. 766 Almazán, Jon.................................................... 228 Ameri, Mohammad Reza.................................. 512 Anquetil, Eric............................................. 259, 720 Aouadi, N. ........................................................ 452 Aradhya, V.N. Manjunath.................................. 760 Arruda, A.W.A. ................................................. 615 Arvanitopoulos, Nikolaos.................................. 726 Asi, Abedelkadir........................................ 140, 743 Awaida, Sameh................................................. 797 Awal, Ahmad Montaser....................................... 29 Ayyalasomayajula, Kalyan Ram....................... 523 Bakshi, Ainesh.................................................. 104 Balakrishnan, Kannan....................................... 199 Banerjee, Purnendu.......................................... 627 Baral, S. ........................................................... 458 Barbosa, Ricardo da Silva................................ 517 Barbuzzi, D. ...................................................... 169 Barkoula, K. ........................................................ 55 Barlas, P. .......................................................... 411 Barrios, Juan Manuel........................................ 779 Belaïd, A. .......................................................... 452 Belaid, Abdel..................................................... 146 Belaïd, Abdel............................................... 29, 678 Belanger, David................................................. 555 Benzeghiba, Mohamed Faouzi......................... 297 Bhateja, Ashok K. ............................................... 79 Bhattacharya, S. ............................................... 458 Bhattacharya, U. ...................................... 250, 458 Bhattacharya, Ujjwal................................. 240, 627 Bluche, Théodore.............................. 285, 297, 667 Blumenstein, Michael........................................ 271 Bosch, Vicente.................................................. 690

Bouillon, Manuel................................................ 720 Brès, Stéphane................................................. 363 Bresler, Martin................................................... 563 Brun, Anders............................................. 523, 732 Bui, Tien D. .............................................. 512, 575 Burgers, Jan...................................................... 265 Camara, Antonio Carlos A. .............................. 517 Cao, Huaigu.............................................. 134, 555 Cappuzzo, Andrea Giuseppe............................ 440 Carayannis, George.......................................... 164 Castro-Bleda, María José................................. 633 Chakraborty, A. ................................................ 458 Chanda, Sukalpa............................................... 405 Chatelain, C. .................................................... 411 Chaudhuri, B.B. ................................................ 375 Chaudhuri, Bidyut B. ........................ 199, 627, 643 Chaudhury, Santanu........................................... 79 Chen, Gang....................................................... 714 Chen, Jinying.................................................... 134 Chen, Kai.................................................... 87, 488 Chen, Youxin..................................................... 246 Chen, Zhaoxin................................................... 259 Cherfa, Yazid.................................................... 655 Cheriet, Mohamed............................ 506, 655, 702 Chherawala, Youssouf...................................... 506 Chibani, Youcef................................... 93, 446, 702 Chowdhury, A. Roy........................................... 250 Coetzer, Johannes............................................ 434 Cohen, Rafi....................................................... 140 Contreras, David............................................... 779 Coüasnon, Bertrand............................................ 35 Cuccovillo, V. ..................................................... 67 Dai, Li-Rong...................................................... 311 D'Andecy, Vincent Poulain.................................. 29 Das, Debleena.................................................. 405 Davila, Kenny.................................................... 323 de Does, Jesse................................................. 349 De Stefano, Claudio.......................................... 569 Dengel, Andreas............................................... 621 Dennhardt, Martin............................................... 15

820

Author Index Depuydt, Katrien............................................... 349 Dershowitz, Nachum............................................. 3 Dey, Prasenjit.................................................... 661 Diamantatos, Paraskevas................................. 649 Diaz-Cabrera, Moises................................. 61, 482 Diem, Markus............................................ 771, 779 Ding, Xiaoqing................................................... 246 Dinstein, Itshak.................................................. 140 Djeddi, Chawki............................................ 93, 446 Doermann, David.............................................. 583 Doetsch, Patrick........................ 279, 343, 494, 549 Drira, Fadoua.................................................... 696 Du, Jun...................................................... 303, 311 Duong, Chi Nhan............................................... 575 Dutta, D. ........................................................... 250 Economou, G. .................................................... 55 Economou, George........................................... 749 Eglin, Veronique................................................ 673 Églin, Véronique................................................ 363 El Abed, Haikal.................................................... 93 Elarian, Yousef.................................................. 766 El-Sana, Jihad................................... 140, 387, 743 Ergina, Kavallieratou......................................... 589 Eskander, George S. ........................................ 187 España-Boquera, Salvador............................... 633 Fan, Wei............................................................ 291 Fecker, Daniel............................................. 15, 743 Fernández-Mota, David............................ 228, 476 Ferrer, Miguel A. ................................ 61, 482, 754 Fiel, Stefan........................................................ 779 Fingscheidt, Tim.......................................... 15, 743 Fink, Gernot A. ......................................... 470, 537 Fischer, Andreas....................................... 222, 512 Fontana, Fabio.................................................. 440 Fontanella, Francesco....................................... 569 Fornés, Alicia............................................ 228, 476 Fotopoulos, S. .................................................... 55 Fotopoulou, Foteini........................................... 749 Freca, Alessandra Scotto di.............................. 569 Frinken, Volkmar............................... 128, 393, 621

Fu, Chi-Wing..................................................... 122 Galbally, Javier.................................................. 482 Garain, U. ......................................................... 791 Garcia, Christophe............................................ 696 Gatos, Basilis.......................... 9, 41, 464, 809, 814 Gattal, Abdeljalil.......................................... 93, 446 Gerogiannis, Demetrios P. ............................... 399 Ghanmi, Nabil................................................... 146 Giotis, Angelos P. ............................................. 399 Goel, Kratarth.................................................... 104 Gómez, Juan C. ................................................. 73 Gomez-Barrero, Marta...................................... 482 Gomez-Gil, Pilar................................................ 649 Govindaraju, Venu............................................ 357 Granger, Eric..................................................... 187 Griechisch, Erika............................................... 738 Grzeszick, René................................................ 470 Haddad, Lobna.................................................. 335 Hadjadj, Zineb................................................... 655 Haji, Medhi........................................................ 512 Hamdani, Mahdi................................................ 494 Hamdani, Tarek M. ........................................... 335 Hamilton, Blaine................................................ 803 Hangarge, Mallikarjun....................................... 375 Hanson, Lisa..................................................... 601 Hao, Yuechan................................................... 329 Hati, Anirban Jyoti............................................. 595 He, Sheng......................................................... 265 He, Yuan........................................................... 291 Hebert, D. ......................................................... 411 Hennebert, Jean................................................ 488 Higashikawa, Yoshiyasu................................... 110 Hirata, Nina S.T. ............................................... 500 Hlavác, Václav.................................................. 563 Hollaus, Fabian................................................. 771 Holtkamp, Michiel.............................................. 175 Hu, Jin-Shui....................................................... 311 Hui, Siu Cheung................................................ 122 Impedovo, D. .............................................. 67, 639 Impedovo, S. .................................................... 169

821

Author Index Inatani, Soichiro................................................ 684 Ingold, Rolf.......................................... 87, 423, 488 Jaiem, Faten Kallel............................................ 673 Jain, Rajiv.......................................................... 583 John, Jomy........................................................ 199 Julca–Aguilar, Frank......................................... 500 Kacem, A. ......................................................... 452 Kacem, Afef....................................................... 678 Kakisako, Ryosuke............................................ 128 Kanoun, Slim............................................. 673, 797 Karabaa, Faik.................................................... 175 Kassis, Majeed.................................................. 387 Katsouros, Vassilis............................................ 164 Kavallieratou, Ergina......................................... 649 Kedem, Klara.................................................... 140 Kermorvant, Christopher........... 158, 285, 297, 667 Khayyat, Muna.................................................. 152 Khémiri, Akram.................................................. 678 Kimura, Fumitaka...................................... 405, 661 Kleber, Florian................................................... 779 Knibbe, Maxime................................................ 297 Kour, George..................................................... 417 Kovalchuk, Alon.................................................... 3 Kozielski, Michal............................... 279, 343, 549 Kumar, Gaurav.................................................. 357 Lam, Louisa....................................................... 152 Lebourgeois, Frank........................................... 696 Lemaitre, Aurélie................................................. 35 Leydier, Yann.................................................... 363 Li, Nan....................................................... 134, 555 Li, Peiyu............................................................ 720 Li, Xin................................................................ 246 Lins, Rafael Dueire............................................ 517 Lira, Edson da F. de.......................................... 517 Liu, Changsong................................................. 246 Liu, Cheng-Lin................................................... 193 Liwicki, Marcus.................... 87, 423, 488, 621, 738 Lladós, Josep............................................ 228, 476 Louloudis, Georgios............................ 41, 464, 814 Louradour, Jérôme.................................... 285, 297

Ludi, Stephanie................................................. 323 Ma, Long-Long.................................................. 317 Mahmoud, Sabri A. .................................. 537, 797 Malik, Muhammad Imran.......................... 621, 738 Marcelli, Angelo......................................... 440, 569 Märgner, Volker.......................... 15, 708, 743, 797 Mårtensson, Lasse............................................ 732 Martín-Albo, Daniel........................................... 543 Matsushita, Tomohisa....................................... 369 Matysiak, Martin................................................ 343 Maziane, Abdelkrim........................................... 655 Medjkoune, Sofiane.................................. 205, 500 Mello, C.A.B. .................................................... 615 Messina, Ronaldo..................................... 158, 297 Mezghani, Anis.................................................. 797 Mignone, P. ........................................................ 67 Mitianoudis, Nikolaos........................................ 609 Molinari, Cristiano............................................. 440 Mondal, Tanmoy............................................... 210 Morales, Aythami................................ 61, 482, 754 Mori, Minoru...................................................... 393 Mouchère, H. .................................................... 791 Mouchere, Harold.............................................. 205 Mouchère, Harold...................................... 259, 500 Moysset, Bastien............................... 158, 297, 667 Murdock, Michael.............................................. 803 Nakagawa, Masaki...... 23, 234, 329, 369, 563, 684 Naoi, Satoshi..................................................... 291 Natarajan, Prem........................................ 134, 555 Ney, Hermann........................... 279, 343, 494, 549 Nguyen, Cuong Tuan........................................ 234 Nguyen, Vu....................................................... 271 Nikou, Christophoros......................................... 399 Nobile, Nicola.................................................... 152 Ntirogiannis, Konstantinos................................ 809 Nuhn, Malte....................................................... 549 Ogata, Ryota..................................................... 393 Oliveira, Luiz S. ................................................ 779 Oosten, Jean-Paul Van..................................... 175 O'Reilly, Christian.............................................. 222

822

Author Index Pal, Umapada... 210, 271, 381, 405, 595, 661, 754 Pantke, Werner........................................... 15, 743 Papamarkos, Nikolaos...................................... 609 Papavassiliou, Vassilis...................................... 164 Paquet, T. ......................................................... 411 Paraskevas, Diamantatos................................. 589 Pardeshi, Rajmohan.......................................... 375 Parodi, Marianela................................................ 73 Parui, S.K. ................................................ 250, 458 Parui, Swapan K. ............................................. 240 Parvez, Mohammad Tanvir............................... 797 Parziale, Antonio....................................... 440, 569 Pastor-Pellicer, Joan......................................... 633 Pavithra, M.S. ................................................... 760 Peng, Xujun....................................................... 555 Petitrenaud, Simon............................................ 205 Pham, Vu.......................................................... 285 Phan, Truyen Van............................... 23, 563, 684 Pirlo, G. .............................................. 67, 169, 639 Plamondon, Réjean.................................. 222, 543 Poirriez, Baptiste................................................. 35 Ponson, Dominique........................................... 512 Pramod, K.V. .................................................... 199 Pratikakis, Ioannis................................. 9, 809, 814 Prusa, Daniel..................................................... 563 Puigcerver, Joan............................................... 181 Quach, Kha Gia................................................. 575 Ragot, Nicolas................................................... 210 Ramel, Jean-Yves............................................. 210 Reese, Jackson................................................. 803 Reid, Shawn...................................................... 803 Riba, Pau.................................................. 228, 476 Richard, Grégoire.............................................. 720 Rizzo, Anna Paola............................................. 440 Rodríguez, Nayara............................................ 754 Romero, Verónica............................................. 785 Rothacker, Leonard........................................... 470 Roy, Partha Pratim.................... 216, 506, 595, 661 Roy, Sangheeta................................................ 661 Russo, M. ......................................................... 639

Saabne, Raid.................................................... 417 Saavedra, Jose M. ........................................... 779 Sablatnig, Robert...................................... 771, 779 Sabourin, Robert............................................... 187 Saleem, Sajid.................................................... 771 Sammara, Petros.............................................. 265 Sánchez, Joan Andreu...................................... 785 Santoro, Adolfo................................................. 440 Santosh, K.C. ................................................... 375 Savaria, Yvon.................................................... 222 Saxena, P.K. ...................................................... 79 Schlöter, Ralf..................................................... 343 Schomaker, Lambert......................... 175, 265, 531 Sehad, Abdenour.............................................. 702 Seuret, Mathias................................................. 423 Shaw, Bikash.................................................... 240 Shibata, Tomoyuki............................................ 116 Siddiqi, Imran.............................................. 93, 446 Simistira, Fotini.................................................. 164 Slimane, Fouad......................................... 708, 797 Souici-Meslati, Labiba......................................... 93 Srihari, Sargur N. ..................................... 601, 714 Stamatopoulos, Nikolaos.................... 41, 464, 814 Stefanos, Gritzalis............................................. 589 Stutzmann, Dominique...................................... 363 Subramanian, Krishna....................................... 555 Suen, Ching Y. ................................................. 152 Sun, Jun............................................................ 291 Surinta, Olarik................................................... 175 Süsstrunk, Sabine............................................. 726 Suwanwiwat, Hemmaphan............................... 271 Swanepoel, Jacques......................................... 434 Szocs, Barna..................................................... 429 Tan, Chew Lim............................................ 98, 381 Tang, Peng........................................................ 122 Tanha, Jafar...................................................... 349 Tonouchi, Yojiro................................................ 116 Toselli, Alejandro H. ................................... 47, 785 Toselli, Alejandro Héctor........................... 181, 690 Uchida, Seiichi.................................. 128, 393, 621

823