Optimal Selection of Binarization Techniques for the Processing of Ancient Palm Leaf Manuscripts

Rapeeporn Chamchong, Chun Che Fung

School of Information Technology Murdoch University

90, South Street, Murdoch, Western Australia 6150 rapeeporn.c@gmail.com, l.fung@murdoch.edu.au

Abstract-Ancient palm leaf manuscripts in Thailand have been preserved by many organizations for the protection and retrieval of traditional knowledge. With advanced computer technology, digitized media is now commonly used to record these documents. One objective of such work is to develop an efficient

image processing system that could be used to retrieve knowledge and information automatically from these manuscripts. Binarization is an important stage during preprocessing of the manuscripts for subsequent extraction of text and characters. The output is then used for further processes such as character recognition and knowledge extraction. However, there is no single binarization technique that is suitable for all documents. This study aims to improve the binarization process intelligently by

applying machine learning techniques to classify the optimal selection of binarization techniques of palm leaf manuscripts. Experiments results are reported and this technique could be applied for automatic selection of binarization techniques in other image processing problems.

Kqwords-Palm leave manuscripts, Binarization techniques, Image Segmentation.

I. INTRODUCTION The number of multimedia databases and the amount of

information stored and captured are rapidly increasing with advances in computer technology. Although the availability of imaging tools and image processing techniques has made it feasible to process image documents in multimedia formats for future analysis and storage, no single system is capable to retrieve relevant information efficiently and to extract knowledge automatically from documents. It is therefore a key objective of this project to develop an efficient image processing system that could be used to retrieve knowledge and information from historical palm leaf documents in Thailand.

Dried palm leaves have been used as one of the most popular written documents for over five hundred years in Thailand. Such materials have been used for recording Buddhist teaching and doctrines, folklores, knowledge and use of herbal medicines, stories of dynasties, traditional arts and architectures, astrology, astronomy, and techniques of traditional massages. With the passage of time, most of these palm leaves, if left unattended, will deteriorate as they are coming to the end of their lifetime, as they face destructive natural elements such as dampness, fungus, bacteria, insects and bugs. This leads to many projects with objectives to preserve and protect information from ancient manuscripts.

978-1-4244-6588-0/10/$25.00 ©201 0 IEEE

Such projects are initiated and carried out by Thai libraries, universities and institutes including medical departments and religion organizations. Three examples of these projects are the Digitisation Initiative for Traditional Manuscripts of Northern Thailand Project at Chiang Mai University Library [1, 2], the Palm Leaf Manuscript Preservation Project in Northeastern Region of Thailand at Mahasarakham University [3], and the Thailand Herbal Repository Access Initiative (THRAI) at Kasetsart University. In particular, the THRAI project aims at developing an ultimate database for Thai traditional medicinal wisdom for the preservation and propagation of medical knowledge and wisdom from ancient manuscripts.

However, it is recognized that the process of scanning a digital image of the ancient palm leaves also presents some difficulties. Most of the original leaves are aged, leading to deterioration of the writing media, with seepage of ink and smearing along cracks, damage to the leaf due to the holes used for binding the manuscript leaves, dirt and other discoloration. These factors lead to poor contrast, smudges, smear, stains, and ghosting noise due to seeping ink from the other side of the manuscripts between the foreground text and the background. Digital image processing techniques are therefore necessary to improve the readability of the manuscripts.

Prior to the stage of knowledge extraction, characters or text on the images have to be recognized. There are four steps which need to be completed prior to the task of character recognition. First, a manuscript is scanned into a RGB image and then it is converted to a gray-scale image. Next, image enhancement is used to enhance the quality of the image. After this stage, binarization is applied and then text and character separation are carried out before character recognition. Binarization is an essential part of the preprocessing step in image processing, converting gray-scale image to binary image, which is then used for further processing such as document image analysis and optical character recognition (OCR). Consequently, both image enhancement and binarization of historical document are crucial to remove unrelated information, noise and background on the documents. If these steps are ineffective, the original characters from the image may be unrecognizable or more noise may be added. Therefore, these techniques are essential to improve the readability of the documents and the overall performance of the process.

3796

Several binarization algorithms have been proposed in the literature [4-13]. However, it is difficult to select the optimal algorithm. The comparison of image qualities from those algorithms is not an easy task as there is no objective evaluation process to compare the results. In contrast, some researchers have proposed a quantitative image measurement of binarization. This performance evaluation of binarization algorithms is recognized to significantly depending on the image content and on the methodologies of binarization. The common approach is to design a set of criteria and scores of criteria. The criteria may be computed by machine or may be decided by visual human. From survey, many researchers have evaluated and compared several algorithms by using different measurements. In addition, most of the measurements have to compare the result image with a ground truth image. It is recognized that there is no single binarization technique that is suitable for all images and there is no automatic selection of the optimal binarization technique.

In this paper, a proposal for the optimal selection of binarization techniques of pahn leaf manuscripts is reported. In

section II, a summary of candidate binarization techniques is provided. Section III describes the framework of the optimal selection of binarization techniques by machine learning. Experimental results are then shown in Section IV and fmally, a conclusion and discussion on future research are given in the last section.

II. THE CANDIDATE BINARIZATION TECHNIQUES Binarization is the technique of converting a gray-scale

image to a binary image by using threshold selection techniques to categorize the pixels of an image into either one of the two classes. Most of the studies [4-7, 13, 14] separated the binarization techniques into two main methods that are global thresholding and local adaptive thresholding techniques

In this paper, the candidates of binarization techniques are Otsu [10], Kittler and Illingworth [15], Kapur [16], Yen and et al [17], Tsai [18], and Bernsen [13]. A summary of these techniques is given in Table 1.

TABLE!. BINARIZATION TECHNIQUES

Binarization Techniques I.Otsu [10]

2.Klittler and Illingworth [15]

3.Kapur [16]

4.Yen and et aI. [19]

5.Tsai [18]

6.Bernsen [13]

1'\ (thr *) = o�(thr*)/o} o�(thr*) = arg max o� (thr) , Os:thr<L-l

Criteria

thr* is optimal threshold, 1'\ is separation criteria, o�(thr) is variance between group of histogram at threshold thr and o} is variance of histogram Top! = arg min{P(T)logof (T) + [1-P(T)Jlog ob(T)

-P(T) log P(T) -[I -P(T)]log[I - p(T)b where Of (T) and Of (T) are foreground and background standard deviations. Topt = argmax[Hf(T)+ Hb(T)] where

Hf (T) = -f p(l) log p(l) and

I=oP(T) P(T)

Hb(T)=- � p(l) log p(l)

I=T +\ P(T) P(T) Topt = arg max[Cb (T)+ Cr<T)] where

Cb (T) = -IOgl M p(l) r} and 1 P(T)

Cf(T)=-IOg{ � [�r} I=T+\ I-P(T)

Topt = arg equal [m\ =b\(T),m2 =b2(T),m3 =b3(T)J wh T k db k k

ere mk = Lp(I)1 an k =Pfmf +Pbm b 1=0

T(x, y) - (Zlow + Zhigh)l2, C(x, y) = (Zhigb - Zlow<) e Window size(r x r)= 15x15 and e= 15

3797

Testing set

Input image

Candidate algorithms

Features from training set

Features from testing set or input image

Selected algorithm

Estimated Model

Selection .........................................................

Binary image

Figure I . Framework of optimal selection ofbinarization techniques

III. A FRAMEWORK FOR THE OPTIMAL SELECTION OF BlNARIZA TlON TECHNIQUES

In most circumstances, a human operator is capable to determine and compare the results from different processing algorithms on images and to select the best result. However, it is not easy for the computer to select the best result unless some quantitative measures are being used. If an automated algorithm selection is implemented, this will assist the process and improve the system performance. This study aims at selecting the most appropriate algorithm by using machine­learning techniques. The proposed framework comprises of three modules: clustering, feature extraction, and selection. The framework is shown in Fig. 1.

A. Clustering Module This module is used to cluster the binary images by using

k-means algorithm [20]. The objective is to group the images of similar characteristics together. When an optimal algorithm is selected, it could be applied to the other images within the same cluster.

B. Feature Extraction Module Color histogram is the most commonly used feature for

image characterization. It is found to be very effective. Beside, color moment, is used in this research in the forms of mean and standard deviation. They are used to convey the color distribution information. This forms a compact representation of the color feature used to characterize a particular color image.

C. Selection Module This module is the main component of the proposed

system. The prediction model is established by learning from the training data set. In this study, Backpropagation [20] is used.

In this study, the image data set is based on ancient palm leaf manuscripts obtained from the Palm Leaf Manuscript Preservation Project in Northeastern Region, Mahasarakham University [3]. There are 380 palm leaf images which have been scanned previously. The resolution of the input images is 200x200 dpi in RGB format. Samples of the original images are shown in Fig. 2. The input images were converted to gray­scale images and then noise is reduced by Gaussians filtering technique. After that, binarization techniques have been applied. Fig. 3 and 4 show some binarized results from the samples.

(a)

(b)

(c)

Figure 2. Samples of original images (a) example I , (b) example 2 and (c) example 3

3798

';�;,�:a-.;��:"j-��6-.>O'c> tJ!)Ck4j);,;;�A-;"'v £'0..:,:;:;5 n : .�"".v'4<>"8"�6wo,1O�,,�' '8:�"'i.3l""oQ)§it,."i5If>'l.v"'V)�(r.;,

d'V1U>�.38.o'7Jo-q��<t>�<><>i:>.o" :>",,,1k>«>ilui>"'O>V £""ff. <fq �J})<n-.?.oib�?�lo . <3o2�".,e:Or.>,�",'oo.;>().38.3.;>"'''-'<!3?V/'l:J I �D<I"'i (f>e1J><1)�"Qfis>,o<<fipQ">Bw"�':>; �;I " "l D<,>:$'l(J)8<J60 I �- .=--- _ ..

(a)

(b) "!t.-� . • -�- ' .. '-" . ,'" 0

-_ .- -_

��-'J/)tf)·:>�4-;n]Qu ·��VV:s:,"iCl"�:;?'.f')6� o iJ;;�(; .. 3C;.nu(J)'It:lz.)Q..?.u.nv �",'.v,,,,,,,,,,.5/>w',1O. \ . \:., '/le'J<1.nn.31""..Q)6il,."i51i> 1.v",'O.J2(f)) ).<:bV:fJ.v� ..... �. j8 .... $:l1J.o1J'1) �_ ·2'<.,1)..38()o;()'�"��J,!:n):r;)4r.\O;}�.oorV[O€)':i<;}�J !

·,',(}><nv.o!:J.)"=..!'«!blo <3ol"i'efJd'>v:O<'>'�'.:J'''o.;>o.3i}.3"" ,<1'-'<!3,.;;l:J

I .. �<1D if>'{J<f><1al<»ilX.D,,'I? r(J>r.><<Rr.> ,,,,8 ei7a� '.0; 1}--17" ",;.s'l <f>8nlbo

1_ ._ . ____ -"=_-"f.:-"- --::.:- ";"";I!"""

(c) :-.'- . -�-, . . ,-, . ., . .- _.-

� 2'7� )tf)', �d'a' 7�lJ liJ:z;z:,J:i,'iCl "�;?.n6..x:; 0 v;{�c,3;);'V(J3 '1t:l t:J £.? ,v.n5 �",?,-, ,,,,,,,,,,.5/>unn. \. \: � f!e,.:>.vo.n.31""..Q)6it,."i51i> I...,,,, 'V.J2(f)., ">.<Sl')v:JtV2..:J8 .... $fJ7..JouJ) :A- 2'<.1>..38()�,,:5·?�J.cn)Jf;;4r.'.D;2i)C'lOvr.oQtf�d!J)

" ''(}?<nv.o 9.),,= ..!'«!blo <361,,<&fJd'>V:O"""�':)' "0.;>0 .30.3.;>, }<1'�<!3 ,.;;l:J " ��10(f);{J(()t'a.1<n�,:.�[(f)tO'<I5{')1J?8ei7.;;� ;of,>:n • .qt:)I:');:Ji2,<r>8IJ€Jo

. . ___ ___ -=--- - -::. "- 7"

(d)

(e)

(f)

Figure 3. The binary result images from example 1, (a) Otsu, (b) Klittler & Illingworth, (c) Kapur, (d) Yen and et at. (e) Tsai and (f) Bernsen

TABLE II. CONFUSION MATRIX OF THE OPTIMAL SELECTION OF BINARIZATION TECHNIQUES

Predicted as � BS KI OT YE TS KP Total

Bernsen (BS) 86 2 I 1 2 0 92

Klittler & 1 47 1 0 0 0 49 Illingworth (KI) Otsu (OT) 2 4 65 1 0 0 72

Yen and et aL 2 0 1 48 0 0 51 (YE)

Tsai (TS) 2 0 0 0 81 5 88

Kapur(KP) 0 0 0 0 3 25 28

Total 93 53 68 50 86 30 380

3799

boP

I wi)imv ro7� 'mJ t em".» "1.1W<tf/iUi i>JiJi)i>"iJIl1{)'siP 8(iD v:Vo,{,,,'?3) r(W ?'v"''''� ><,�"9n �l. \ lozil:J rt'-'1<>ifjlv.i5otD'![ri1:J �-" luvll>ll>cm!)�lo0 ga-",;)'$);J,i)(,UlW)!) ;J iJ,(j'f)Bi't"'»Vi)!)-'If),).",(ja'V",, u""""olJm)iJ �'t,7

"U».1)�Ii1�iJv;eJ:>O>S):��'l."wlv�5tm?.8�3')�f1>�":Gi>o").iO,

(a)

(b)

(c)

(d) .::::s.......... _ $-=_

lwi)."..,m�r:mJt"':O('.» "�/,'f)o/IQj i>JiJi)""iJ.zl1{)('siP8(jDv:Vo,{,,,'!i5«W _ '9'v"''''Gi{ >{,�"'9n' �\\ lozil:J m'�if51iY;lDOal)tti6toj utJ ll>ll>cr.o!)lilo0

krr�Mi�-w w- iJ,(j�..,r;�·»:"hl).�:,t}l;."{,."'�"""O U>'D.1) ?'h '-",,:,?��Dj���f!�-t�?oi����"·Gi>08).{o!:,

(e)

(f)

Figure 4. Binary result images from example 2, (a) Otsu, (b) Klittler & Illingworth, (c) Kapur, (d) Yen and et aI. (e) Tsai and (f) Bernsen

TABLE III.

Algorithms

I.Bemsen

2.Klittler & Illingworth

3.0tsu

DETAIL ACCURACY BY CLASS OF OPTIMAL SELECTION OF BINARIZA TION TECHNIQUES.

TPRate FPRate ROC Area

0,935 0,024 0.992

0.959 O.oI8 0.992

0.903 0.01 0.971

4.Y en and et at. 0.941 0.006 0.992

5.Tsai 0.921 0.017 0.995

6.Kapur 0.893 0.014 0.993

IV. EXPERIMENTAL RESULTS AND DISCUSSION

In this experiment, 10 fold cross-validation was applied to the 380 palm leaf manuscript images in order to evaluate the optimal selection of binarization techniques. The selection is calculated by using backpropagation algorithm with a hidden layer with 258 input nodes (256 bins of histogram, mean, and average values), 132 hidden nodes, and 6 output nodes. The confusion matrix and detailed accuracy by class of the selection system is given in Tables 2 and 3 respectively.

From the confusion matrix in Table I, the accuracy rate (correctly classify) of this system is 92.63%. The results of this experiment in each algorithm (class) in Table III have shown that Klittler and Illingworth's algorithm have the highest True Positive (TP) rate at 95.9%, which is a measurement of the positive cases that were correctly classified while Kapur's algorithms has a TP rate at 89.3%. The lowest False Positive (FP) rate is Yen and et al.'s algorithm at 0.6%, which is the negative cases that were incorrectly classified as positive but the highest FP rate is Bernsen's algorithm at 2.4%. The area under the ROC Curve (AUC) is a method of measuring the performance of the ROC curve which is given by the TP Rate and FP Rate. The best prediction from AUC of this experiment is Tsai's algorithm at 0.995 while the weakest prediction is Otsu's algorithm at 0.971. Overall AUC results from each algorithm can be analyzed that even if the prediction is not perfect, it is not random as well. However, all cases are almost 1 so the prediction is almost perfect.

V. CONCLUSTION AND FUTURE WORK

Although many binarization techniques have been applied to document image processing problems, such techniques have difficulties in eliminating background noise on palm leaf manuscripts. Most of the reported work has been carried out on standard images that are not as complex and in poor quality such as the ancient palm leaf manuscripts. From this study, it can be concluded that there is no single binarization technique that is suitable for all images. For this reason, how to choose the best binarization technique for the user is the key issue and currently, there is no automatic selection of the optimal binarization technique. At present, machine learning technique is being investigated in order to determine which algorithm is suitable for binarization techniques. This proposal framework can be applied for automatic selection system but it will be better if the appropriate selections are ranked for user as semi­automatic selection. Users can make a decision from this information by themselves. Subsequent this framework will be improved the ranking of appropriate selection in the future.

ACKNOWLEDGMENT

We thank the Preservation of Palm Leaf Manuscripts Project, Mahasarakham University, Thailand for their support and providing images from their database.

REFERENCES

[ 1] (2008). Infonnation Heritage Available: http://www.ernc.com! leadership/digital-universelinfonnation-heritage-awards-2008.htm

[2] Y. Musiket, "Scanning the Past, " 15 January 20 10, 2009.

[3] Mahasarakham University, "Palm Leaf Manuscript Preservation Project in Northeastern Region, " Mahasarakham University, Thailand 2005.

[4] R. Chamchong and C. C. Fung, "Comparing background elimination approaches for processing of ancient Thai manuscripts on palm leaves, " in 2009 Int. Conf. Machine Learning and Cybernetics, China, 12- 15 July, 2009.

[5] Y. Chen and G. Leedham, "Decompose algorithm for thresholding degraded historical document images, " in lEE Proceeding Visual Image Signal Processing, December, 2005.

[6] J. He, Q. D. M. Do, A. C. Downton, and J. H. Kim, "A comparison of binarization methods for historical archive documents, " in Proc. 8th Int. Conf. Document Analysis and Recognition, 2005, pp. 538-542 Vol. 1.

[7] G. Leedham, Y. Chen, K. Takru, T. Joie Hadi Nata, and M. Li, "Comparison of some thresholding algorithms for textlbackground

segmentation in difficult document images, " in Proc. 7th Int. Conf. Document Analysis and Recognition, 2003, pp. 859-864.

[8] M. Sezgin and B. Sankur, "Survey over image thresholding techniques and quantitative perfonnance eva1uation, " J. of Electronic Imaging, vol. 13, pp. 146- 168, 2004.

[9] W. Niblack, An introduction to digital image processing: Prentice Hall, 1986.

[ 10] N. Otsu, "A threshold selection method from gray-level histogram, " IEEE Trans. Systerns Man Cybernet, vol. 9, pp. 62-66, 1979.

[ 1 1] J. Sauvola and M. Pietikainen, "Adaptive document image binarization, " Pattern Recognition, vol. 33, pp. 225-236, 2000.

[ 12] M. Sezgin and B. Sankur, "Selection of thresholding methods for nondestructive testing applications, " in Proc. 200 1 Int. Conf. Image Processing, 200 1, pp. 764-767 vol.3.

[ 13] O. D. Trier and A. K. Jain, "Goal-directed evaluation of binarization methods, " IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 17, pp. 1 19 1- 120 1, December 12, 1995.

[ 14] R. C. Gonzalez and R. E. Woods, Digital Image Processing. New Jersey: Prentice-Hall, 2002.

[ 15] J. Kittler and J. lllingworth, "Minimum error thresholding, " Pattern Recognition, vol. 19, pp. 4 1-47, 1986.

[ 16] J. N. Kapur, P. K. Sahoo, and A. K. C. Wong, "A new method for gray­level picture thresholding using the entropy of the histogram, " Graph. Models Image Process, vol. 29, pp. 273-285, 1985.

[ 17] Y. Jui-Cheng, C. Fu-Juay, and C. Shyang, "A new criterion for automatic multilevel thresholding, " Image Processing, IEEE Transactions on, vol. 4, pp. 370-378, 1995.

[ 18] W. H. Tsai, "Moment-preserving thresholding: A new approch, " Graph. Models Image Process, vol. 19, pp. 377-379, 1985.

[ 19] L.-K. Huang and M.-J. J. Wang, "Image thresholding by minimizing the measures of fuzziness, " Pattern Recognition, vol. 28, pp. 4 1-5 1, 1995.

[20] F. v. d. Heijden, R. P. W. Duin, D. d Ridder, and D. M. J. Tax, Classification, parameter estimation and state estimation : an engineering approach using MATLAB West Sussex, England: John Wiley & Sons, Ltd, 2004.

3800