WELDING DEFECT PATTERN RECOGNITION IN RADIOGRAPHIC IMAGES OF GAS PIPELINES USING ADAPTIVE FEATURE EXTRACTION METHOD AND NEURAL NETWORK CLASSIFIER

Transcription

1 23 rd Word Gas Conference, Amsterdam 2006 WELDING DEFECT PATTERN RECOGNITION IN RADIOGRAPHIC IMAGES OF GAS PIPELINES USING ADAPTIVE FEATURE EXTRACTION METHOD AND NEURAL NETWORK CLASSIFIER Man author S. MANSOURI ALGHALANDIS * Co-author GH. NOZAD ALAMDARI 2 * Frst Correspondng (Presentng) Author, Ema: yahoo.com - Senor Gas Transmsson Expert, Natona Iranan Gas Company (NIGC), R&D Dept., Dst-8 of Gas Transmsson Operaton, Tabrz , Iran. 2- Wed Interpretng Expert, Natona Iranan Gas Company(NIGC), Mechanca Dept., Dst-8 of Gas Transmsson Operaton, Tabrz , Iran.

2 . ABSTRACT: In ths paper a new method for automatc recognton and separaton between defected radographc mages of weds and correct ones s ntroduced. The method appes oca mage nformaton together wth adaptve feature extracton parameters and neura network cassfer. In practca wed nspecton of gas ppenes, more than sxty percent of radographc mages are not fauty, but need to be separated n vsua check by a certfed expert. Ths s a tme consumng process for nterpreter and reduces the eye senstvty and reabty of nspecton especay for defected parts of radographc mage. The ntroduced pattern cassfer s desgned to sove ths probem. In frst step, preprocessng s done through mage enhancement and nose reducton technques proper for poor quaty and ow contrast radographc mages. Preprocessng step s apped through medan fterng and adaptve hstogram equazaton, together wth preservng grayscae mage nformaton. In next step wed s extracted from background mage and adaptve threshodng s apped usng oca bnarzaton operators wth mantanng crtca mage nformaton. Then weded area s zoned based on the mportance of the mage nformaton and features are extracted for dfferent zonng patterns. A varety of topoogca features and aso grayscae nformaton reated to each zone are used. Some fm defects mstaken nstead of wed defects are aso detected n ths stage. By usng fm defect nformaton n cassfcaton stage, the recognton effcency can be ncreased. The extracted features are fed nto an artfca neura network cassfer. A new structure of neura network cassfer n combnaton wth bnary ogc s ntroduced for cassfcaton stage. The reabty and accuracy of ths new hybrd neuro-ogc structure s compared wth conventona neura networks accordng to adjusted oca threshodng and adaptve zonng parameters. The probem of choosng approprate features are aso dscussed and evauated. The resuts show the recognton performance and fexbty of ths new hybrd neuro-ogc cassfer n comparson wth conventona neura networks structure. Key words: Radographc pattern recognton Wedng defect detecton Adaptve feature extracton Neura network cassfer 2

3 TABLE OF CONTENTS. Abstract 2. Introducton 3. Dgtzng 4. Image enhancement and processng 4.. Medan fterng 4.2. Contrast enhancement 4.3. Wener fterng 5. Image segmentaton 5.. Wed extracton 5.2. Loca bnarzaton 6. Feature extracton 6.. Fm defect feature 6.2. Unavaabty of defected area 6.3. Average sze of defects 6.4. Average dstance of defects from center ne 6.5. Number of defects 6.6. Zonng 7. Cassfcaton 7.. Mut ayer perceptron neura network archtecture 7.2. New hybrd neuro-ogc archtecture 8. Resuts 8.. Data 8.2. Recognton performance 8.3. Dscusson 9. Concuson 0. References. Lst of Tabes 2. Lst of Fgures 3

4 2. INTRODUCTION Radography s one of the od and st effectve NDT toos. The most mportant appcaton of ths method n gas ndustres s the nspecton of the weds for transmsson ppenes. Even wth the nventon of onne dgta radography usng senstve fuorescent pates, offne anaog radography usng fms st has ts own benefts and appcatons. Rapd deveopment of gas dstrbuton network reveas the necessty for a reabe and automatc wed nspecton system. Checkng poor quaty and ow contrast fms of weds s a tme consumng process for a certfed expert and reduces hs/her eye senstvty and nspecton accuracy especay for defected parts of radographc mages. The most of the papers pubshed about automatc wed faw recognton systems, a dea wth dentfyng the type of the wed defects. Ths artce deas wth the separatng the correct part of wedng from ts defected part n radographc fms. Snce, more than sxty percent of radographc mages n gas ppenes are not fauty but need to be separated n vsua check, our desgned system can act as an effectve too to reduce the mass work of fm checkng process and put the concentraton of the nspector on defected parts of the fms. In the artces, there s no report about dentfcaton of fm defects, whch reduce the effcency of wed faw recognton system n cassfcaton stage. In ths study a method s ntroduced for separaton of fm defects whch can be mstaken nstead of wed faws, too. Extracton of weds from background mage can be done automatcay by computer or manuay by user seecton. There are varous reported methodooges for automatc wed extracton based on the assumpton that the ntensty of the pxes n the wed area dstrbute more as Gaussan dstrbuton than other areas n the mage. An automatc wed extracton method s used n ths study. Next step s mage segmentaton and feature extracton. In ths artce oca threshodng s used for adaptve segmentaton of weded area. Then some topoogca features are extracted from segmented area and fed nto an MLP neura network cassfer wth back propagaton agorthm. In ths paper, an automatc wed defect separaton system s desgned and tested. At frst, fm dgtzng s descrbed. Then preprocessng and mage enhancement s apped. Next stage oca segmentaton and feature extracton s studed. The extracted features together wth fm defect nformaton are apped to an MLP neura network cassfer. Fnay a comparson between dfferent sets of nput parameters to pattern cassfer s done and the resuts are dscussed. 3. DIGITIZING Fm dgtzer as an entrance gate for mage data nput, s a crtca part of the wed recognton system. Seectng optmzed resouton of scannng and acceptabe quaty of dgtzng pays an mportant roe n whoe system performance. Gamma-ray fm strps were dgtzed usng scannng devce under controed umnaton wth approxmate fm dmensons of 70 mm by 300 mm. A commerca scanner wth ght ntensty of 20,000 cd./sq.m ( cande per square meter) was used. Dfferent resoutons from 300 to 2000 dp ( dot per nch) wth gray scae ranges from 8 bt to 6 bt were tested to fnd an optma seecton. It s known that the human beng s not abe to dffer n gray scae over 28 eves(7 bts) [] and mages of 6 bt depth n gray scae eve occupy sgnfcant memory space and take ong tme to be processed. In fgure, some scanned sampe mages wth dfferent resouton from 300 to 2000 dp are shown. In ths study for compromsng between mage quaty and processng speed, mages wth resouton of 600 dp and gray scae of 256 eves were seected. (.a)300 dp. (.b) 600 dp. (.c) 600 dp. (.d) 2000 dp. Fg.. Dgtzed mages wth dfferent resouton and 8 bts of gray eve. (a)300 dp. (b) 600 dp. (c) 600 dp. (d) 2000 dp. 4

5 4. IMAGE ENHANCEMENT AND PROCESSING Preprocessng and mage enhancement s done to remove system nose and aso photographc fm nose n three steps ncudng medan fterng, contrast enhancement and wener fterng. 4.. Medan fterng Frst of a, medan fterng s apped through the mage. For the mask shown n fgure, suppose that MED(,j) s the medan of the vaues n 5Χ5 neghborhood pxes where (, j ) represents pxe coordnates n the mask center of m 3. Ftered mage s obtaned by usng the reaton : MED, j) = Medan( m, m, m,..., m, ) () ( m25 The gray eve of each pxe s repaced by medan of the gray eves n the neghborhood of that pxe. Ths method s partcuary effectve for the nose patterns consstng strong, spke ke components and where the characterstc to be preserved s edge sharpness [ 2]. Orgna mage of wed and ts 5Χ5 medan ftered s shown n fgure 2. (2.a) (2.b) (2.c) Fg. 2. (a) 5Χ5 neghborhood pxes as a mask for medan fterng n Eq.(). (b) Orgna mage (c) 5Χ5 medan ftered mage Contrast enhancement Dfferent technques of hstogram equazaton to enhance contrast of radographc mages have been reported n many researches [],[3],[4],[5]. The enhancement technque of choce s the so caed hstogram equazaton (HE). For dscrete vaues of gray eves, probabtes gven by the reaton: r n k ( rk ) = 0 rk n P, k = 0,,, L - (2) where r k represents the gray eve of the pxes of the mage to be enhanced, wth k back and k r = 0 representng r = representng whte n the gray scae, L s the number of eves, r ) P s the probabty of the k th gray eve, n k s the number of tmes ths eve appears n the mage. A pot of P k ( r k ) versus rk s usuay caed hstogram and the technque used for obtanng a unform hstogram s known as HE technque [2]. In ths technque the cumuatve hstogram H of gray eves G s used as the essenta part of the functon F HE ( G) that maps the orgna gray eves nto the transformed ones: H ( G) F HE N ( G) G mn + G wth G = G max G mn = (3) r ( k 5

6 n whch G max and mn G ndcate the upper and ower mts of the transformed gray vaues, respectvey and N represents the number of pxes over whch the hstogram has been taken [6]. HE s a goba approach based on gray eve dstrbuton over an entre mage and s not sutabe to enhance detas over sma areas. The goba hstogram equazaton s made adaptve by takng the hstogram over a oca regon nstead of the whoe mage: F AHE H ( G) mn N ( G) AHE = G + G (4) Adaptve hstogram equazaton (AHE) has been recognzed as a vad method of contrast enhancement n mage processng. The man advantage of AHE s that t can provde better contrast n oca areas than that achevabe utzng tradtona hstogram equazaton methods. Whereas tradtona methods process the entre mage at once, AHE utzes oca contextua regon. The effect of AHE n contrast enhancement of ths study s shown n fgure 3. AHE (3.a) (3.c) (3.d) (3.b) (3.e) Fg.3. (a) Orgna mage. (b) Hstogram of orgna mage. (c) Adaptve hstogram equazaton(ahe), The contextua regon shown s an m х m mask around a pxe at ocaton (x,y). (d) Image after AHE (e) Hstogram of mage after AHE Wener fterng Appcaton of wener (east-mean-square) fter n mage restoraton has been aready reported as a near fter [2]. In ths study wener fter s apped to the mage adaptvey, taorng tsef to the oca Image varance. Where the varance s arge, t performs tte smoothng, where the varance s sma t 6

7 performs more smoothng. Ths approach s more seectve than comparabe near fter, preservng edges and other hgh-frequency parts of the mage. It works best when nose s constant-power addtve nose, such as Gaussan nose. Snce the ntenstes of the pxes n the wed area dstrbute more as a Gaussan dstrbuton than other areas n the mage of wed [7], t was seected for fterng radographc mages of weds. 5. IMAGE SEGMENTATION The segmentaton methodoogy ncudes the foowng steps: 5.. Wed extracton Before dong any further oca processng on the radographc mage, t s preferred to extract wed regon by separatng between weded area and ts background mage. For ths reason, proposed methodoogy by Lao et a. [8], s apped n ths research Loca bnarzaton Bnarzaton of scanned gray scae mages s an mportant step n most mage anayss systems. For poor quaty mages such as radographc fms, goba threshodng doesn t have good performance for areas wth varabe background ntensty, ow contrast and stochastc nose. Therefor t s essenta to fnd bnarzaton methods whch w correcty abe a the nformaton present. Dfferent oca bnarzaton methods are reported by many researchers[9]. Nback s method s known to gve the best performance for the context of dgt recognton [0]. In ths study, Nback s method s used by tunng ts parameters for radographc mages of weds. The dea of ths method s to vary the threshod over the mage, based on the oca mean and oca standard devaton. The threshod at pxe ( x, y) s cacuated as T ( x, y) m ( x, y) + k. s ( x, y) = (5) Where m( x, y) and s( x, y) are the sampe mean and standard devaton vaues, respectvey, n a oca neghborhood of ( x, y). The sze of the neghborhood shoud be sma enough to preserve oca detas, but at the same tme, arge enough to suppress nose. The vaue of k s used to adjust how much of the tota mage object boundary s taken as a part of the gven object. In ths research 5х5 neghborhood and k= gave we separated defect objects n the wed area. 6. FEATURE EXTRACTION In ths artce features descrbng number, sze, ocaton and avaabty of defected area s used to make dstncton between true and fauty areas n the wed regon. 6.. Fm defect feature Durng the deveopment process of the fm, sometmes wed area n the mage s damaged by na effect of the operator. Ths fm defect s very smar to faw patterns and may be mstaken nstead of wed defect. As shown n fgure 4, the gray eve of fm defect s hgher than gray eve of rea wed defect. So, n ths feature the gray eve of defected area s detected and f t s more than a defned threshod, the feature equas one. Otherwse the feature equas zero. (4.a) (4.b) Fg.4. (a) Image wth fm defect(na effect). (b) Image wth wed defect, 7

8 6.2. Unavaabty of defected area If there s no separated area known as defect, ths feature equas to one, otherwse t equas to zero Average sze of defects: Average sze of defects( of defects( n ) A n ): the rato between sum of defected areas ( A ) to the number 6.4. Average dstance of defects from center ne: Average dstance of defects from center ne ( d n ) : the rato between sum of the dstances of each defect from the center of the wed bead ( d ) to the number of defects( n ) Number of defects: Number of defects ( n ): number of separated areas n the wed bead known as defected area Zonng: Zonng s an mportant method for dervng ow resouton structura features [3],[4]. In ths study, zonng s adopted wth the structure of the weded area. Weded regon s extracted to a rectanguar frame array. In norma zones heght(h) and wdth(w) are seected to be the same, where W s aways equa to the wed wdth. In proposed zoned bock, heght of the weded area s seected to be one thrd of W. Features mentoned n subsectons 6. to 6.5 are extracted for these two types of zoned bocks. The frst set s for norma zone(h=w) and the second one s for proposed zone(h=w/3). 7. CLASSIFICATION The task of pattern cassfcaton s to assgn an nput pattern represented by a feature vector, to one of the output specfed casses. In compex systems wth nonnear reatons between nputs and outputs, conventona approaches proposed for sovng the probem of nonnear pattern cassfcaton can be found n certan we-constraned envronments, non s fexbe to perform we outsde ts doman. Artfca neura networks (ANNs ) cassfer can provde fexbe aternatves and many appcatons coud beneft from usng them. ANNs are composed smpe eements operatng n parae. These massvey parae systems wth arge number of nterconnectons may sove a varety of chaengng cassfcaton probems. One of the more update nes of research s the cassfcaton of wed defects usng ANNs and aso there are many reports on ths subject [],[4],[]. In ths secton two types of neura networks structure to construct the reatonshp between system nputs and outputs w be expaned. The frst ANNs s a feed forward mut ayer perceptron wth back-propagaton earnng agorthm. The second one s a new hybrd neuro-ogc cassfer structure proposed n ths research. 7.. Mut ayer perceptron neura network archtecture Mut ayer perceptron (MLP) s the most common type of feed-forward ANNs. The abty of earnng s a fundamenta property n neura network archtecture. Durng earnng process connectons of the weghts are updated. Error-correcton earnng rue and back-propagaton earnng agorthm s used for updatng purpose of ths MLP structure. Proposed MLP s composed of many nterconnected neurons that often caed nput, output and hdden ayers as shown n fgure 5. The neurons of nput ayer are used to receve nput vector X and neurons of output ayer are used to produce the correspondng output vector Y. Pattern cassfer n ths secton s made up of three ayers, wth fve neurons n nput ayer, and two neurons n output ayer. The actvaton functon f s commony chosen to be sgmod n order to resembe the two state output of boogca neurons whch orgnay nspred the networks. Output vector s defned equa to ( Y =, Y 2 =0) for no defect status and equa to 8

9 ( Y =0, Y 2 =) for defected status. Neurons n the hdden ayers, sum up vaues from nput nodes after weghtng them wth approprate weghts W j and compute the output Yo as a functon of summaton. In tranng process, the actua output vector Y o generated by network may not equa to the desred output vector Y. The back-propagaton (BP) agorthm s the most commony adopted MLP tranng d agorthm and t s the most wdey apped neura network archtecture. BP computaton agorthm s as foows [2] (5.a) (5.b) Fg.5. (a) Confguraton of artfca neura networks(anns) wth one nput, one hdden and output ayers usng back-propagaton earnng agorthm. (b) Traned MLP neura network structure wth three ayers and fve nput neurons.. Intaze the weghts( W j ) to sma random vaues. (µ ) X 2. Randomy choose an nput pattern 3. Propagate the sgna forward through the network. 4. Compute δ n the output ayer( where δ = f ( h )[ Y µ d Y o ] O = ) Y o h represents the net nput to the th unt n the th ayer, and f s the dervatve of the actvaton functon f. 5. Compute the detas for the precedng ayers by propagatng the errors backwards: δ = f ( h ) 6. Update weghts usng W j W = η δ Y + j j δ + j for = ( L ), L,. 7. Go to step 2 and repeat for the next pattern unt the error n the output ayer s beow a specfed threshod or a maxmum number of teratons s reached. Therefore the BP agorthm starts wth the output ayer and teratvey computes the δ vaues for the neurons n a ayers. It s common to add a momentum factor nto a BP agorthm to ncrease ts earnng speed. The momentum factor determnes how much the prevous weghts change nfuences the new weght change. The new equaton for W s shown as foows j W + j = η δ j X + η W where η s the momentum coeffcent wth vaue 0 µ. The resuts are presented n secton 8. j 9

10 7.2. New hybrd neuro-ogc archtecture In secton 7. a fve extracted features are fed drecty nto an ANNs wth BP earnng archtecture (Fg.5a) then traned network s tested and evauated. In ths secton a new hybrd neuro-ogc structure s proposed. It s a combnaton of neura networks and bnary ogc structures. Ony three features n subsectons 6.3, 6.4 and 6.5 s fed nto three nput neurons wth BP earnng agorthm. Other two features n subsectons 6. and 6.2 havng bnary vaues are drecty fed nto a ogca structure. (6.a) (6.b) Fg.6. (a) Traned MLP neura network structure usng back-propagaton earnng agorthm wth three ayers and three nput neurons. (b) Proposed hybrd neuro-ogc structure. 0

11 Tranng s done accordng to the structure shown n fgure (6.a) but neuro-ogc structure of fgure (6- b) combned of neura networks and ogca gates s used for testng and cassfcaton. Bocks named AND, OR and NOT are common ogca gates. Bocks caed T are bnarzer. If the nput to ths bock s more than a predefned threshod, the output s one, otherwse the output s zero. Seected threshod s 0.5 for these bocks. If one of the features 6. or 6.2 becomes actve, means there s no defect, regardess of the resuts n the ANNs cassfer, the output w be defect free (e. Y =0, Y 2 =) otherwse the resuts of the neura network cassfer w be domnant. 8. RESULTS 8.. Data Data bank was seected from mages of γ -ray radography of weds n 6 gas ppenes. Dgtzng method was descrbed n secton 3. Nnety mages of weds were seected from the data bank and used to tran MLP neura network structure. Two thrd of the seected data were defect free and the remanng one thrd were defected. Error toerance for tranng data was set at the vaue of 0.0 wth the defaut earnng rate of By varyng the number of neurons n the hdden ayer and foowng the tranng errors, the best performance was reached at forty neurons n the ntermedate ayer whch s optmum number of neurons for the data used. Tabe. Recognton resuts of two types of cassfer for two sets of features ncudng norma zone(h=w) and proposed zone(h=w/3). No fm defect data used for tranng. Correct Data wth no fm defects Rejecton (%) Error (%) Recognton (%) Features from norma zone (H=W) Features from proposed zone (H=W/3) BP structure (Fg.5) Neuro-ogc structure (Fg.6) BP structure (Fg.5) Neuro-ogc structure (Fg.6) Tabe2. Recognton resuts of two types of cassfer for two sets of features ncudng norma zone(h=w) and proposed zone(h=w/3). Fm defect data used for tranng. Correct Data ncudng fm defects Rejecton (%) Error (%) Recognton (%) Features from norma zone (H=W) Features from proposed zone (H=W/3) 8.2. Recognton Performance BP structure (Fg.5) Neuro-ogc structure (Fg.6) BP structure (Fg.5) Neuro-ogc structure (Fg.6) Recognton performance s shown n two tabes. Both of the tabes have the resuts of two types of cassfer for two sets of features. The frst cassfer s ANNs wth BP structure and the second one s the new hybrd neuro-ogc structure proposed n ths paper. One of the feature sets s extracted from

12 the norma zones(h=w) and the other one s obtaned from proposed zone(h=w/3). Proposed neuro-ogc structure has ess nput neurons and takes shorter tme for tranng n comparson wth BP structure wth fve nput neurons. It has aso better recognton performance accordng to the resuts shown n tabes and 2. Comparng the resuts from the zonng pont of vew ndcates that n smar cassfers, wed separaton effcency ncreases for the features extracted from the proposed zone(h=w/3). Tabe shows data wth no fm defect nformaton used for tranng. In tabe 2, seven fm defect data are repaced n tran set for tranng the cassfers. It s observed that drecty usng the fm defect nformaton n tranng data, reduces recognton effcency Dscusson Test resuts ndcates that proposed neuro-ogc structure gves better recognton performance and aso speeds up the tranng process by separaton between bnary(two eve) and numerc features, then tranng ony numerc features and combnng the whoe resuts together by a ogca structure. Athough drect tranng of neura network structure wth fm defect nformaton, reduces recognton effcency and creates more confuson n wed separaton stage, but ths new neuro-ogc structure have the abty to reject fm defects by ts ogca structure wth no need for fm defect tranng. Aso more meanngfu resuts can be obtaned from features extracted from proposed zone(h=w/3). Overa recognton rate and reabty can be ncreased by deveopng data sampes and usng a bg tranng set. 9. CONCLUSION Ths paper has descrbed a method for automatc recognton and separaton between defected radographc mages of weds and correct ones. A new structure of neura network cassfer n combnaton wth bnary ogc s ntroduced. Expermenta resuts show that ths new hybrd neuro-ogc structure gves better recognton performance, especay when ANNs are confused by fm defect nformaton durng tranng stage. Neuro-ogc structure has the abty to reject fm defects by ts ogca part wth no need for tranng the fm defects. It s mportant because drect tranng of the fm defect nformaton to neura networks, reduces the recognton effcency. For better recognton performance adaptve preprocessng and oca bnarzaton technques are used for mage enhancement and segmentaton then features are extracted from proposed zoned bocks. 0. REFERENCES. da Sva, R.R., Caoba, L.P., Squera, M.H.S. and Rebeo, J.M.A.(2004). Pattern Recognton of Wed Defects Detected by Radographc Test. NDT & E Internatona, 37: Gonzaez, R.C. and Wntz, P.(987). Dgta Image Processng. Addson-Wesey, 2nd Edton. 3. Wanga, X. and Wong, B.S. (2004). Image Enhancement for Radography Inspecton. The Thrd nternatona Conferenceon Expermenta Mechancs, SPIE Proceedngs, Wang, G. and Lao, T.W.(2002). Automatc Identfcaton of Dfferent Types of Wedng Defects n Radographc Images. NDT & E Internatona,35: Shafeek, H.I., Gademawa, E.S., Abde-Shafy, A.A. and Eewa, I.M. (2004). Assessment of Wedng Defects n Ppene Radographs Usng Computer Vson. NDT & E Internatona,37: Casteman, K.R.(979). Dgta Image Processng. Engewood Cffs N.J., Prentce Ha, Wang, G. and Lao, T.W. (2002). Automatc Identfcaton of Dfferent Types of Wedng Defects n Radographc Images. NDT & E Internatona, 35: Lao, T.W. and N,J. (999). An Automated Radographc NDT System for Wed Inspecton: Part I- Wed Extracton. NDT & E Internatona, 29: Trer, Ф.D. and Taxt, T. (995). Evauaton of Bnarzaton Methods for Document Images. IEEE Transactons on Pattern Anayss and Machne Integence,7: Trer, Ф.D. and Jan, A.K. (995). Goa Drected Evauaton of Bnarzaton Methods. IEEE Transactons on Pattern Anayss and Machne Integence,7: Juang, S.C., Tarng, Y.S. and L, H.R. (998). A comparson between Back-propagaton and Counter-propagaton Networks n the Modeng of TIG Wedng Process. J of Materas Processng Technoogy, 75:

13 2. Jan, A.K., and Mao, J. (996) Artfca Neura Networks : A Tutora. IEEE Computer Magazne, March 996: Trer, Ф.D. and Jan, A.K. (996) Feature Extracton Methods for Character Recognton A Survey. Pattern Recognton, 29: Srkantan, G. et a. (996) Gradent Based Contour Encodng for Character Recognton. Pattern Recognton, 29: LIST OF TABLES Tabe. Recognton resuts of two types of cassfer for two sets of features ncudng norma zone(h=w) and proposed zone(h=w/3). No fm defect data used for tranng. Tabe2. Recognton resuts of two types of cassfer for two sets of features ncudng norma zone(h=w) and proposed zone(h=w/3). Fm defect data used for tranng. 2. LIST OF FIGURES Fg.. Dgtzed mages wth dfferent resouton and 8 bts of gray eve. (a)300 dp. (b) 600 dp. (c) 600 dp. (d) 2000 dp. Fg.2. (a) 5Χ5 neghborhood pxes as a mask for medan fterng n Eq.(). (b) Orgna mage (c) 5Χ5 medan ftered mage. Fg.3. (a) Orgna mage. (b) Hstogram of orgna mage. (c) Adaptve hstogram equazaton(ahe), The contextua regon shown s an m х m mask around a pxe at ocaton (x,y). (d) Image after AHE (e) Hstogram of mage after AHE. Fg.4. (a) Image wth fm defect(na effect). (b) Image wth wed defect, Fg.5. (a) Confguraton of artfca neura networks (ANNs) wth one nput, one hdden and output ayers usng back-propagaton earnng agorthm. (b) Traned MLP neura network structure wth three ayers and fve nput neurons. Fg.6. (a) Traned MLP neura network structure usng back-propagaton earnng agorthm wth three ayers and three nput neurons. (b) Proposed hybrd neuro-ogc structure. 3