AUTHENTICATION OF OTTOMAN ART CALLIGRAPHERS

INTERNATIONAL JOURNAL OF ELECTRONICS; MECHANICAL and MECHATRONICS ENGINEERING Vol.2 Num.2 pp.(2-22) AUTHENTICATION OF OTTOMAN ART CALLIGRAPHERS Osman N. Ucan Mustafa Istanbullu Nyaz Klc2 Ahmet Kala3 Istanbul Aydn Unversty, Engneerng&Archtecture Faculty, Electrcal and Electroncs Eng. Dept. Florya Istanbul, Turkey uosman@aydn.edu.tr mustafastanbullu@aydn.edu.tr 2 Istanbul Unversty Engneerng Faculty, Electrcal and Electroncs Eng. Dept. 332, Avclar, Istanbul, Turkey nyazk@stanbul.edu.tr 3 Istanbul Unversty Economy Faculty, Beyazt, Istanbul, Turkey akala@stanbul.edu.tr Abstract: A content-based retreval system can provde effcent access to the documents from the Ottoman Empre archves, whch contan more than mllon handwrtten fles. Experts want to archve the hstorcal documents to be stored n dgtal mage forms because the documents nclude not only text but also drawngs, portrats, mnatures, sgns, nk smears, etc., whch mght have an assocated hstorcal value. Ths work wll help us to dentfy the Ottoman perod callgraphes; accordngly determne ts value and chronology. Here, we descrbe and apply a computatonal technque on hgh-resoluton scanned form of the orgnal works, specfcally Ottoman art callgraphes for authentcaton. We show prelmnary results from 9 well known Ottoman art callgraphers at varous tmes. Snce pantngs have hgh resoluton, complexty and capacty, t s not possble to process them n one block. In ths paper, Wavelet Transform (WT) and Support Vector Machne (SVM) are employed n cascade form to obtan an objectve approach n authentcaton of callgraphes, and we have succeed up to % percentage for classfy the art works usng these technques. Keywords: Art authentcaton, Ottoman art callgraphers, Wavelet transform, Support vector machne. INTRODUCTION Authentcaton of pantngs, callgraphes can be extremely challengng. Typcally, art experts reach decsons after thorough consderaton of many dfferent types of evdence. Artst s lfetme and documents and tracng the art s hstory of ownershp provde clues. In addton to the relance on the human actor, quanttatve methods can be brought to bear. Avalable technques of examnaton are gven below: Carbon- datng s used to measure the age of an object up to, years old []. Whte lead datng s used to pnpont the age of an object up to, years old [2]. Conventonal X-ray can be used to detect earler work present under the surface of a pantng. Sometmes artsts wll legtmately re-use ther own canvasses, but f the pantng on top s supposed to be from the 7th century, but the one underneath shows people n 9th century propertes such as dress models, the scentst wll assume the top pantng s not authentc. Also x-rays can be used to vew nsde an object to determne f the object has been altered or repared. X-ray dffracton (the object bends X-rays) s used to analyze the

components that make up the pant an artst used, and to detect pentment. X-ray fluorescence (bathng the object wth radaton causes t to emt X-rays) can reveal f the metals n a metal sculpture or f the composton of pgments s too pure, or newer than ther supposed age [3]. Quanttatve technques could potentally be helpful n dentfyng mages, copes or forgeres thus nvestgators may rely on other methods such as computerzed authentcaton. In lterature, regardng to ths case, there are some works based on mage processng. [, ] In vew of all these thngs, there s not so much workng about Ottoman art callgraphes for authentcaton. In ths case we thnk that ths work wll be an mportant pont for authentcaton of Ottoman art callgraphes. Wavelet Transform s a mathematcal tool, whch analyses and decomposes one and two dmensonal sgnals. Wavelet decomposton dvdes sgnals nto sub-bands. These subbands are evaluated to determne textures, assgnng a frequency to each sub-band and related coeffcents of sub-bands can be utlzed as feature extracton block and followed by a classfer. Then, these data are used as nput nto Support Vector Machne for testng []. All ths chosen ones are well known n ther lfe perod and have a lot of dstnctve callgraphes. We have chosen ther art callgraphes for ths work. These are some examples of the chosen ones: Fgure Some callgraphy examples of Ottoman Empre Fgure Computaton Scheme Here some mportant ottoman art callgraphers and ther works have been chosen. Abbaskaml, Hafz Rashbral, Hasan Rıza, İsmet el-mevlev, Mahmud Celaleddn, Sabr, Seyyd Dervş Hall, Seyyd Muhammed Bahr and Velyyuddn are all mlestones for art callgraphes of Ottoman Empre. These art callgraphes and ther dgtal photographs are separated by two consecutve approaches; wavelet and Support Vector Machne (SVM) n evaluaton of hgh capacty 2D pantng mages. In hgh resoluton mages, one step drect mage processng s almost mpractcal, because of tme consumng long teratons. The frst approach s based on compresson wth mnmum loss va feature extracton. The second s classfcaton of orgnal art callgraphes and ts fakes by SVM, where extracted features are taken as nput. SVM developed by Vapnk [7, 8] have been used n a range of problems ncludng pattern recognton, bonformatcs and text categorzaton [9-]. SVM provdes a novel approach to the two and mult-class classfcaton problem. 2

2. METHODS In ths work, callgraphes some are gven n Fgure 2, segmented properly and samples augmented to peces. Then these peces appled wavelet transform for feature extracton. In ths step, wavelet coeffcents of mages are calculated. At the frst stage, 8 wavelet transform outputs; approxmaton, vertcal, horzontal, dagonal of Level- DAUB and HAAR type wavelet decompostons and maxmum, mnmum, mean, varance, standard devaton values of raw mage are obtaned. In the second stage of cascade structure, these 3 parameters are chosen as nput of SVM for classfcaton. As a result, authentcaton accuracy of ths cascade block, has reached up to %. 2.. Wavelet Based Feature Extractıon Wavelets decompose data nto dfferent frequency sub-bands components and then study each component wth a resoluton matched to ts scale. Wavelets have come out as powerful new mathematcal tools for analyss of complex datasets. In classcal approach, Fourer transform provdes representaton of an mage based only on ts frequency contents. Hence ths representaton s not spatally localzed whle wavelet functons are localzed n space. Whle Fourer transform groups a sgnal nto a spectrum of frequences whereas the wavelet analyss decomposes nto a herarchy of scales rangng from the coarset scale. Hence wavelet transform whch provdes representaton of an mage at varous resolutons s a better tool for feature extracton from mages [, 2]. The wavelet transform s a useful mathematcal tool that currently has receved a great attenton n dfferent applcatons lke compresson and feature extracton. Feature extracton s defned that extracton some mportant features from the mage and obtanng feature vector [3]. Here, wavelet transform s used as a feature extractor. In numercal analyss and functonal analyss, a dscrete wavelet transform (DWT) s any wavelet transform for whch the wavelets are dscretely sampled. Therefore, snce our data s dscrete, we use Dscrete Wavelet Transform. DWT employs a dscrete set of the wavelet scales and translaton obeyng some defned rules as an mplementaton of the wavelet transform. As wth other wavelet transforms, a key advantage t captures both frequency and key advantage t captures both frequency and locaton nformaton. Here, both of scale and translaton parameters are dscrete. Thus, DWT can be represented n Eq.(), W [m, n] f [ x] m, n [ x] x where, dscretzed scale and () translaton parameters are gven by, a 2 ve b k 2 j j ( k, j Z ). Then, wavelet bass functon s wrten n Eq.(2), j,k [ x] 2 j / 2 (2 j x k ) (2) One dmensonal transforms are easly extended to two dmensonal functons lke mages. In ths case, the DWT s appled to each dmenson separately. Ths yeld a mult resoluton decomposton of the mage nto four sub-bands called the approxmaton (low frequency component) and detals (hgh frequency component). The approxmaton (A) ndcates a low resoluton of the orgnal mage. The detal coeffcents are horzontal (H), vertcal (V), and dagonal (D). Fgure2 DWT Transform 2.2. Support Vector Machne Based Classfcaton Support vector machnes are a set of relates supervsed learnng methods that analyss data and recognze patterns, used for classfcaton. The orgnal SVM algorthm was nvented by Vladmr Vapnk and current standard ncarnaton (soft margn) was purposed by Cornna Cortes and Vladmr Vapnk. The standard SVM takes a set of nput data and predcts, for each gven nput, whch of two possble classes the nput s a member of, whch makes the SVM a nonprobablstc bnary lnear classfer. It s formulzed under the concept of structural rsk mnmzaton rule unlke neural network based classfer []. At the begnnng, SVM was actually desgned for bnary classfcaton n order to construct an optmal hyperplane so that the margn of separaton between the 27

negatve and postve data set wll be maxmzed. Let x, y,, 2,...l, y, and x IR be the tranng samples where the n tranng vector s x and y s ts correspondng labeled class. As a result SVM can be expressed as follows; l f ( x) sgn y K ( x, x ) b (3) for u sgn(u ) for u where l s the number of learnng patterns, y s the target value of learnng pattern x, b s a bas, and K ( x, x ) s a kernel functon that hgh-dmensonal feature space: K ( x, x ) ( x). ( x ) () The polynomal kernel whch s shown n equaton () and the Gaussan radal bass functon (RBF) kernel whch s shown n equaton () are frequently used kernel functons. K(x, x ) exp[ x - x ] p () () In polynomal kernel, p s the degree of polynomal kernel; f p s equal to, kernel s called lnear and f t s equal to 2, then kernel s called quadratc kernel. 2.2.. Tranng of Support Vector Machne SVM fnds the hyperplane that causes maxmzes the separatng margn between two postve and negatve classes. Mathematcally, ths hyperplane can be found by mnmzng the followng cost functon, P ( w) w 2 to the am s to maxmze the separatng margn subject to constrants. Ths problem s transformed to dual form for solvng whch follows: LD j y y j x x j 2 j (8) y. f (x ),,2,..., k Subject to Where, K ( x, x ) x.x Subject 2 (7) C and y. Where, C s regularzaton parameter that controls the tolerance to classfcaton errors n tranng. The tranng vector x whose correspondng s nonzero s called support vector. 3. AUTHENTICATION OF ART CALLIGRAPHIES Wavelet coeffcents of mages are calculated. After decomposton wth Dscrete Wavelet Transform, we have four coeffcents for each mage; approxmaton, vertcal, horzontal and dagonal. Level- DAUB and HAAR type wavelet decompostons are preferred. The frst level decomposton vector sze s too large to be gven as an nput to a classfer. Snce hgh dmensonal of feature vectors ncreased computatonal complexty and hence, n order to reduce to dmensonalty of the extracted feature vectors, statstcs over the wavelet coeffcents are used. The statstcal features were also chosen as; maxmum, mnmum, mean, varance, standard devaton values of {approxmaton, vertcal, horzontal and dagonal} sub bands. The computed statstcal features of 3 dscrete feature coeffcents are used as the nputs of the network of SVM. In SVM tranng, lnear, quadratc and RBF kernel s used as n Tables -. General revew s shown n table. The best error level s acheved when RBF kernel parameters.. CONCLUSION In ths study, we have classfed callgraphes of 9 Ottoman callgraphers and ts mages wth the help of wavelet transform and SVM, whch are used for feature extracton and supervsed machne learnng approaches. We have 28

preferred wavelet type s Level- Daub and Haar after expermental study. For compresson and unfcaton, only some statstcs of frst level wavelet coeffcents are used as an nput of SVM. These are; {maxmum, mnmum, mean, varance and standard devaton values of the {approxmaton, vertcal, horzontal and dagonal} sub bands. Our expermentaton showed that classfcaton accuracy has reached up to % by Daub type Wavelet and cascaded by RBF Kernel Type SVM. Ths statstc suggests that our combned approach may be used to facltate separaton from orgnal pantng to ts fake.. REFERENCES [] http://amath.colorado.edu/courses/2 /2Spr/Lab/Forgeres.pdf, Detectng Art Forgeres [2] http://www.mystudos.com/artstsmnd/volume--22.html [3] LYU S., D. ROCKMORE, H. FARID (2) A dgtal technque for art authentcaton, Proceedng of the Natonal Academy of Scences, (9):7-7. cancer dagnoss usng least square support vector machne, Dgtal Sgnal Processng, 7:9-7. [9] BURGES C (998) A tutoral on support vector machnes for pattern recognton, Data Mnng and Knowledge Dscovery, 2:-3. [] DAUBECHIES I (992) Ten lectures on wavelets, CBMS Conference Lecture Notes. SIAM. Phladelpha. [] GONZALES, R.C., R.E WOODS (993) Dgtal Image Processng, Addson Wesley, Readng, MA. [2] DEVIJVER, P.A., J., KITTLER (982) Pattern Recognton: A Statstcal Approach, Prentce-Hall, London. [3] ŞENGÜR A., TÜRKOĞLU I. (2) Değşmez Momentlerle Türkçe Karakter Tanıma, Doğu Anadolu Bölges Araştırmaları. [] ŞAYKOL E., SINOP A. K., GÜDÜKBAY U., ULUSOY Ö., ÇETIN E., Content-Based Retreval of Hstorcal Ottoman Documents Stored as Textual Images, IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 3, NO. 3, MARCH 2, p: 3-32 [] GUNSEL B., S. SARIEL, O. ICOGLU (2) Content-based access to art pantngs, Proc. of IEEE ICIP, Genova, Italy, September 2. [] KIRAN K. SIMHADRI, S. S. IYENGAR, Ronald J. HOLYER, Matthew LYBANON, John M. ZACHARY, (998) Wavelet-Based Feature Extracton from Oceanographc Images, Ieee Transactons on Geoscence and Remote Sensng, Vol. 3, No. 3, May 998, p: 77-778 [7] VAPNIK V (999) An overvew of statstcal learnng theory, IEEE Transacton on Neural Network, :989-999. [8] POLAT K., S.GUNEŞ (27) Breast 29

. LIST OF TABLES Model :In 9 categores, obtanng results usng only Appr coeffcents and db wavelet for data: (SVM, fold, c=, RBF gama:,) Table a) General performance b) Performance of classfcaton due to callgraphers c) Confuson matrx Correctly Classfed Instances (88%) Incorrectly Classfed Instances (2%) Kappa Statstc.8 Mean Absolute Error.77 Root Mean Squared Error.287 Relatve Absolute Error 88.9% Root Relatve Squared Error 9.% Total Number of Instances a) Class TP Rate FP Rate Precson Recall F-Measure ROC Area.7..7.7.7.82 Abbaskaml Hafızrashbra Hasanrıza.8.8.889.99 Mahmudcelaleddn Sabr.22.833.99.989 Seyyddervşhall.87.2.87.87.87.98 Seyydmuhammedbahr.833.833.99.87 Velyyuddn.8..7.8.727.9 Weghted Avg..88..89.88.882.9 İsmetelmevlev b) CLASSIFIED AS a b c d e f g h a =Abbaskaml b = Hafızrashbral c = Hasanrıza d = İsmetelmevlev e = Mahmudcelaleddn f = Sabr g = Seyyddervşhall 7 h = Seyydmuhammedbahr = Velyyuddn c) 22

Model 2: In 9 categores, obtanng results usng only Appr coeffcents and Haar wavelet for data (SVM fold, RBF gama:,) Table 2 a) General performance b) Performance of classfcaton due to callgraphers c) Confuson matrx Correctly Classfed Instances 3 (8%) Incorrectly Classfed Instances 7 (%) Kappa Statstc.88 Mean Absolute Error.79 Root Mean Squared Error.28 Relatve Absolute Error 88.72% Root Relatve Squared Error 9.73% Total Number of Instances a) Class TP Rate FP Rate Precson Recall F-Measure ROC Area.8.8.889.989.7..7.7.7.82 Abbaskaml Hafızrashbral Hasanrıza.8.8.889.993 Mahmudcelaleddn Sabr.22.833.99.989 Seyyddervşhall.87.2.87.87.87.98 Seyydmuhammedbahr.833.23.833.833.833.8 Velyyuddn.8..7.8.727.9 Weghted Avg..8.9.87.8.8.92 İsmetelmevlev b) CLASSIFIED AS a b c d e f g h a =Abbaskaml b = Hafızrashbral c = Hasanrıza d = İsmetelmevlev e = Mahmudcelaleddn f = Sabr g = Seyyddervşhall 7 h = Seyydmuhammedbahr = Velyyuddn c) 22

Model 3: In 9 categores, obtanng results, takng statstcs of appr and detal coeffcents and usng db wavelet for data (SVM, RBF, gama:,) Table 3 a) General performance b) Performance of classfcaton due to callgraphers c) Confuson matrx Correctly Classfed Instances (%) Incorrectly Classfed Instances (%) Kappa Statstc Mean Absolute Error.728 Root Mean Squared Error.28 Relatve Absolute Error 87. % Root Relatve Squared Error 89. % Total Number of Instances a) TP Rate FP Rate Precson Recall F-Measure ROC Area Abbaskaml Class Hafızrashbral Hasanrıza İsmetelmevlev Mahmudcelaleddn Sabr Seyyddervşhall Seyydmuhammedbahr Velyyuddn Weghted Avg. b) CLASSIFIED AS a b c d e f g h a =Abbaskaml b = Hafızrashbral c = Hasanrıza d = İsmetelmevlev e = Mahmudcelaleddn f = Sabr g = Seyyddervşhall 8 h =Seyydmuhammedbahr = Velyyuddn c) 222

Model : In 9 categores, obtanng results, takng statstcs of appr and detal coeffcents and usng Haar wavelet for data: Table a) General performance b) Performance of classfcaton due to callgraphers c) Confuson matrx Correctly Classfed Instances (%) Incorrectly Classfed Instances (%) Kappa Statstc Mean Absolute Error.728 Root Mean Squared Error.28 Relatve Absolute Error 87. % Root Relatve Squared Error 89. % Total Number of Instances a) TP Rate FP Rate Precson Recall F-Measure ROC Area Abbaskaml Class Hafızrashbral Hasanrıza İsmetelmevlev Mahmudcelaleddn Sabr Seyyddervşhall Velyyuddn Weghted Avg. Seyydmuhamme ahr b) CLASSIFIED AS a b c d e f g h a =Abbaskaml b = Hafızrashbral c = Hasanrıza d = İsmetelmevlev f = Sabr g = Seyyddervşhall 8 h =Seyydmuhammedbahr = Velyyuddn e= ahmudcelaleddn c) 223

Model-, only wavelet appr coeffcents and db wavelet; * Model-2, only wavelet appr coeffcents and Haar wavelet; * Model-3, some statstcs (mn, max, mean, std, moment3) of wavelet coeffcents (appr and detal) and db wavelet; * Model-, some statstcs (mn, max, mean, std, moment3) of wavelet coeffcents (appr and detal) and Haar wavelet Table General revew of all models Classfcatory Model Accurate Of Classfcaton (%) SVM Model- * 88 9 category Model-2 * 8 data Model-3 * Model- * 22