Corn Seed Varieties Classification Based on Mixed Morphological and Color Features Using Artificial Neural Networks

Transcription

1 Research Journa of Apped Scences, Engneerng and Technoogy 6(19): , 013 ISSN: ; e-issn: Maxwe Scentfc Organzaton, 013 Submtted: October 03, 01 Accepted: December 03, 01 Pubshed: October 0, 013 Corn Seed Varetes Cassfcaton Based on Mxed Morphoogca and Coor Features Usng Artfca Neura Networks 1 Areza Pazok, Fardad Farokh and Zohreh Pazok 1 Department of Agronomy and Pant breadng, Shahr-e-Rey Branch, Isamc Azad Unversty, P.O.Box: , Tehran, Iran Department of Eectrca and Eectronc Engneerng, Centra Tehran Branch, Isamc Azad Unversty, Tehran, Iran Abstract: The abty of Mut-Layer Perceptron (MLP) and Neuro-Fuzzy neura networks to cassfy corn seed varetes based on mxed morphoogca and coor Features has been evauated that woud be hepfu for automaton of corn handng. Ths research was done n Isamc Azad Unversty, Shahr-e-Rey Branch, durng 011 on 5 man corn varetes were grown n dfferent envronments of Iran. A tota of 1 coor features, 11 morphoogca features and 4 shape factors were extracted from coor mages of each corn kerne. Two types of neura networks contaned Mutayer Perceptron (MLP) and Neuro-Fuzzy were used to cassfy the corn seed varetes. Average cassfcaton s accuracy of corn seed varetes were obtaned 94% and 96% by MLP and Neuro-Fuzzy cassfers respectvey. After feature seecton by UTA agorthm, more effectve features were seected to decrease the cassfcaton processng tme, wthout any meanngfu decreasng of accuraces. Keywords: Artfca Neura Networks (ANNs), corn, Feature seecton, Mut ayer perceptron (MLP), neurofuzzy, seed INTRODUCTION Corn s one of the major foods n the word. In some crops as corn, because of dfferences between varety s morphoogy and quaty, the seeds dentfcaton s very mportant. Coor and morphoogca features are the man vsua factors n seed nspecton and gradng so cassfcaton of dfferent seed varetes are determned accordng to these features generay. Severa gradng systems usng dfferent morphoogca features for the cassfcaton of dfferent cerea gran varetes have been reported n terature (Barker et a., 199a, b, c, d; Majumdar and Jayas, 000; Zapotoczny et a., 008). Features for varous corn damages were dentfed by red, green and bue pxe vaue nputs to a neura network (Steenhoek et a., 001). Recenty, researchers combned varous externa features (Morphoogca, Coor and Textura) to mprove the cassfcaton accuracy of gran kernes. The cassfcaton of gran kernes cannot be easy usng a unque mathematca functon because of the varaton n morphoogy, coor and texture, so neura networks have the potenta of sovng probems n whch some nputs and correspondng output vaues are known, but the reatonshp between the nputs and outputs s dffcut to transate nto a mathematca functon. Neura network cassfers have been successfuy mpemented for sovng the probems of agrcuture such as gran quaty nspecton and especay gran dentfcaton. Many studes have been reported on appcaton of Artfca Neura Networks (ANNs) n agrcuture (Jang et a., 004; Uno et a., 005; Movagharnejad and Nkzad, 007; Savn et a., 007; Zhang et a., 007; Ehert et a., 008). Pazok and Pazok (011) cassfed 5 ran fed wheat gran cutvars usng artfca neura network. The experment resuts ndcated that the average accuracy was % and after feature seecton appcaton by UTA agorthm ncreased to 87.%. Chen et a. (010) proposed a vson-based approach combned wth pattern recognton technques and neura networks to dentfy corn varetes. Experment showed the average cassfcaton accuracy for fve varetes was up to 90%. Yun (004) presented a detecton agorthm based on Back Propagaton (BP) network for cassfcaton of corn. The average recognton accuracy of the standard corn, broken corn and dfferent kerne s genotype coud reach 95%. Neuro-fuzzy networks are combnaton of artfca neura networks and fuzzy ogc. Neuro-fuzzy technques are apped n many feds as mode dentfcaton and forecastng of near and non-near systems. Rutkowaska and Starczewsk (004) presented Correspondng Author: Areza Pazok, Assocate Professor, Department of Agronomy and Pant breadng, Shahr-e-Rey Branch, Isamc Azad Unversty, P O Box: , Tehran, Iran 3506

2 Res. J. App. Sc. Eng. Techno., 6(19): , 013 an approach to cassfcaton of Irs based on neurofuzzy systems and hybrd earnng agorthms n the fed of mage processng and anayss. In ths study, MLP and Neuro-fuzzy neura networks effcency s presented for corn seed varetes cassfcaton and the accuracy dfferences before and after feature seecton s compared. The specfc goa s to extract the externa features of corn kernes and then generate the optma features set for corn varety dentfcaton usng feature seecton agorthm. MATERIALS AND METHODS (a) (b) Due to dentfcaton of 5 corn (Zea mays L.) seeds varetes whch are grown n dfferent envronments of Iran, ths research was done n Isamc Azad Unversty, Shahr-e-Rey Branch durng 011. The expermented corn varetes were ncuded: KSC60, KSC403, KSC400, KSC600 and KSC704 (Fg. 1). In the presented method, at frst, dfferent types of the features were extracted and fed to Mutayer Perceptron (MLP) and neuro-fuzzy networks for cassfcaton. These features conssted of coor features, morphoogca features and shape factors. The MLP and neuro-fuzzy networks were traned on the randomy seected nstances and tested on the rest of the data for cassfcaton of corn seeds varetes. Fnay, the UTA feature seecton agorthm was performed n order to determne the more effectve features (Utans et a., 1995). The program s wrtten n MATLAB verson 7.8. The proposed method was mpemented usng a Pentum V persona computer wth 4GB RAM and.67 GHz CPU. The system archtecture s shown n the Fg.. (c) (e) Fg. 1: Fve corn seed varetes: (a) KSC60, (b) KSC403, (c) KSC400, (d) KSC600 and (e) KSC704 (d) Image acquston: Dgta mage anayss offers an objectve and quanttatve method for estmaton of morphoogca parameters. Ths process uses dgta mages to measure the sze of ndvdua seeds and mathematcay extract features and shape reated nformaton from the mages. A Panasonc camera (Mode SDR-H90) wth zoom ens mm foca ength used to take the mages of corn seed sampes. Images format was 4 bt coor JPEG wth resouton of pxes. The camera was mounted over the umnaton chamber on a stand whch provded easy vertca movement. The dstance between the camera and each seed sampe was fxed (7 cm) to regret the effect of the dstance on saved mages. In order to reduce the nfuence of surroundng ght, a back umnaton chamber s ocated between the sampes and the ens and equa number of the mages (90 mages) was taken for each varety. The acqured corn seed varetes are shown n Fg Fg. : System archtecture Feature extracton: In ths research, coor, morphoogca features and shape factors used for extractng of ndvdua corn seeds as foows.

3 Res. J. App. Sc. Eng. Techno., 6(19): , 013 Coor feature extracton: Coor s an mportant feature that human perceve when vewng an mage. Human vson system s more senstve to coor nformaton than gray eves so coor s the frst canddate used as the feature. There are severa coor spaces. In order to study the effect of coor features on the dentfcaton performance of corn varetes, three transformatons of RGB (red, green and bue) coor space were evauated,.e., HSV, YCbCr and I 1 I I 3. RGB: RGB coor space s the most common used one for mage representaton on computers. An RGB mage, sometmes referred as a true coor mage, s stored as an m-by-n-by-3 data array that defnes red, green and bue coor components for each ndvdua pxe. HSV: MATLAB and the Image Processng Toobox m-fes do not support the HSI coor space (Hue Saturaton Intensty). Therefore, we used the HSV coor space that s very smar to HSI. From the Red (R), Green (G) and Bue (B) coor bands of an mage, Hue (H), Saturaton (S) and Vaue (V) were cacuated usng the foowng equatons (Image Processng Toobox, 007): Max Max (R, G, B) (1) Mn Mn (R, G, B) () V Max (3) Max Mn S (4) Max 1 G B V R 6 Max Mn 1 B R 1 H V G 6 Max Mn 3 1 R G V B 6 Max Mn 3 f H 0 H H 1. (5) YCbCr: The Y eement represents the umnance component and the Cb; Cr eements represent two chromnance components. Equaton (6) represents the YCbCr transformaton of RGB coor space (Umbaugh, 005). I 1 I I 3 : The transformaton of RGB coor space nto I 1 I I 3 coor space can be acheved by the Eq. (7) (Ohta, 1985). I ( R G B) / 3 1 I ( R - B) / I (-R G - B) / 4 3 (7) Furthermore, mean (m) of these coor components were cacuated. In tota, 1 coor features were extracted for dentfcaton. Morphoogca feature extracton: The foowng morphoogca features were extracted from abeed mages of ndvdua corn seeds varetes. Geometry reated features ncudng area, permeter and major and mnor axs engths were measured from the bnary mages (Pawa et a., 001; Zhao- Yan et a., 005). Area (A): The area of a regon s defned as the number of pxes contaned wthn ts boundary. Permeter (P): The permeter s the contour ength of the boundary. Major axs ength (L): The ength of the major axs s the ongest ne that can be drawn through the object. Mnor axs ength (): The ength of the mnor axs s the ongest ne that can be drawn through the object perpendcuar to the major axs. Aspect rato: Major axs ength K = (8) Mnor axs ength Equvaent dameter (Eq): It was the dameter of a crce wth the same area as the corn seed regon. 4 Area Equada (9) Convex area (C): It was the number of pxes n the smaest convex poygon that can contan the corn seeds regon. Sodty (S): The proporton of the pxes n the seeds regon that are aso n the convex hu. Extent (Ex): The proporton of the pxes n the boundng box whch are aso n the seeds regon. Y 0.99R 0.587G 0.114B Cb R G 0.500B 18 Cr 0.500R G B 18 (6) 3508 Roundness (R): Ths s gven by R = 4 Area Permeter (10)

4 Fg. 3: Mutayer perceptron neura network Compactness (CO): The compactness provdes a measure of the object's roundness: CO= 4 Area π L Res. J. App. Sc. Eng. Techno., 6(19): , 013 (11) Shape features: From the vaues of axs ength and Area, shape factors were derved (Symons and Fucher, 1988a) as foow: Major axs ength Shape factor1( SF1) : (1) Area Area Shape factor( SF) : (13) Major axs ength 3 Shape factor3( SF3) : Area (Major axs ength/)(major axs ength/) π Shape factor4( SF4) : Area (Major axs ength/)(mnor axs ength/) π (14) (15) The feature vector was made from above features and fed two artfca networks for cassfcaton. Artfca neura networks: Artfca Neura Networks (ANN) s a mathematca too, whch tres to represent ow-eve ntegence n natura organsms and t s a fexbe structure, capabe of makng a non-near mappng between nput and output spaces (Rumehart et a., 1986). In ths study, Mut Layer Perceptron network (MLP) and Neuro-fuzzy network were used to cassfy corn varetes. Mut Layer Perceptron (MLP) network: An artfca neura network s composed of many artfca neurons that are nked together accordng to specfc network archtecture. The objectve of the neura network s to transform the nputs nto meanngfu outputs The MLP network conssts of an nput ayer, one or more hdden ayers and an output ayer. Each ayer conssts of mutpe neurons. An artfca neuron s the smaest unt that consttutes the artfca neura network (Kantardzc, 003). The network needs to be traned usng a tranng agorthm such as back propagaton. The goa of every tranng agorthm s to reduce the goba error by adjustng the weghts and bases. We apped a MLP neura network wth hdden ayers. The nput ayer had 7 neurons because the data sets contaned 7 parameters and 5 neurons (KSC60, KSC403, KSC400, KSC600 and KSC704) n the output ayer. The apped tranng structure for corn seeds varetes cassfcaton was Typca Mutayer perceptron neura network archtecture s shown n Fg. 3. Neuro-fuzzy cassfcaton network: Many dfferent systems have been apped n cassfcaton probems. In the area of computatona ntegence, neura networks, fuzzy systems and neuro-fuzzy systems are wdey empoyed as cassfers. In the fed of artfca ntegence, neuro-fuzzy refers to combnatons of artfca neura networks and fuzzy ogc. In ths study, we propose an approach to desgn fuzzy system where the membershp functons are chosen n such a way that certan crteron s optmzed. The structure of the fuzzy system s specfed frst and some parameters n the structure are free to change, then these free parameters are determned accordng to the nput-output pars (Wang, 1997). Frst, we specfed the structure of the fuzzy system. The fuzzy system was chosen wth product nference engne, sngeton fuzzfer, center average defuzzfer and Gaussan membershp functon. We apped a neuro-fuzzy cassfer wth the structure as MLP neura network that contaned 60 neurons (rues). The fuzzy system was derved as foow (Wang, 1997): f (x) = M = 1 M = 1 y 1 [ n = 1 [ n = 1 x _ x exp(-( ) )] σ x _ x exp(-( ) )] σ (16) where, M s the number of rues consdered and y -, x - and σ (q) are free parameters and woud determne n earnng phase. Desgnng a fuzzy system means determnng these three parameters. To determne these parameters n some optma fashon, t s hepfu to represent the fuzzy system f (x) of Eq. (16) as a feed forward network.

5 Res. J. App. Sc. Eng. Techno., 6(19): , 013 Fg. 4: Network representaton of the fuzzy system (Wang, 1997) Specfcay, the mappng from the nput x є U R n to the output f(x) є V R can be mpemented accordng to the foowng operatons (Wang, 1997): The nput x s passed through a product Gaussan operator: x _ x z = n exp(- ( ) = 1 ) (17) σ The z are passed through a summaton operator and a weghted summaton operator to obtan b and a: b = M z = 1 (18) a = M y z = 1 (19) Fnay, the output of the fuzzy system s computed: F = b a (0) Neuro-fuzzy system for dentfcaton of corn varetes s shown n Fg. 4. Feature seecton: Feature seecton s the probem of choosng a subset of features deay necessary to perform the cassfcaton task from a arger set of canddate features. There are severa ways to determne the best subset of features. UTA s a smpe method whch s based on traned artfca neura network. In the bass of ths method, average of one feature n a nstances s cacuated. Then the seected feature n a nput vectors has been repaced by the cacuated mean vaue. Then the traned network was tested wth the new features Utans et a. (1995). The comparson error was defned n our strategy as foow: E = (FP (new) + FN (new)) - (FP (od) + FN (od)) (1) where, FP (od) FN (od) = Fase postve = Fase negatve usng the whoe orgna features FP (new) and FN (new) s those vaues when one of the feature repaced by the mean vaue. Three dfferent states coud happen: One nput s consdered reevant f E s postve and the hgher the E s, ndcates the features mportance among other features. One nput s neffectve f E s zero. One nput s not ony neffectve but aso nosy and shoud be removed from the nput vector f E s negatve. RESULTS AND DISCUSSION Identfcaton of corn seed varetes on mage of each corn kerne that contaned sampes of 5 varetes tested. There were 90 mages for each varety. Images 3510

6 Res. J. App. Sc. Eng. Techno., 6(19): , 013 Tabe 1: Average accuracy before UTA agorthm Varetes accuracy (%) Neura networks KSC60 KSC403 KSC400 KSC600 KSC704 Average accuracy (%) MLP Neuro-fuzzy Tabe : Comparson error of morphoogca features n UTA agorthm (MLP) Feature's Error (E) Varetes A P L R C S EX Eq K CO SF1 SF SF3 SF4 KSC KSC KSC KSC KSC Tota (T) Tabe 3: Comparson error of coor features n UTA agorthm (MLP) Feature's error (E) Varetes Rm Gm Bm Hm Sm Vm Ym Cbm Crm I 1 m I m I 3 m KSC KSC KSC KSC KSC Tota (T) Tabe 4: Comparson error of morphoogca features n UTA agorthm (Neuro- fuzzy) Feature's error (E) Varetes A P L R C S EX Eq K CO SF1 SF SF3 SF4 KSC KSC KSC KSC KSC Tota (T) format was 4 bt coor JPEG and pxes accuracy n MLP and neuro-fuzzy neura networks consdered for mages sze. were 94% and 96% respectvey. As t was shown the There were 60 tranng data and 30 test data for performance of neuro-fuzzy neura network was better each corn seed varetes (300 tranng data and 150 test n overa cassfcaton. In ths case, maxmum data for 5 expermented corn seed varetes). Tweve accuraces beonged to KSC60 n MLP (98%) and coor features (Rm, Gm, Bm, Hm, Sm, Vm, Ym, Cbm, neuro-fuzzy (99%). Crm, I 1 m, I m and I 3 m), 11 morphoogca features Due to determne the more effectve features and (Area, Permeter, Major axs ength, Mnor axs ength, dscard the rreevant features, UTA agorthm apped Aspect rato, Equvaent dameter, Convex area, and tota feature's error (T) evauated. In the MLP Sodty, Extent, Roundness and Compactness) structure, 8 effectve features I m (50), SF (48), I 3 m extracted from seed varetes mages and features such (34), Mnor (8), Aspect rato (4), Sm (1), Permeter as area, permeter, major and mnor axs ength (8) and Hm (6) seected (Tabe and 3) because of ther computed on the bnary mages and four shape factors hgher feature's error (Utans et a., 1995). (SF1, SF, SF3 and SF4) were derved from these man After dong UTA agorthm n neuro-fuzzy geometrc features. The program wrtten and tested structure the feature error cacuated and 9 effectve usng MATLAB 7.8 software. features Sm (68), Hm (0), I 3 m (0), Convex Area (14), Many features were hghy correated wth each I 3 m (14), I 1 m (10), SF (8), SF4 (8) and Roundness (6) others and f one of the features was seected, the rest of seected (Tabe 4 and 5). So nneteen ess effectve the features w not contrbute sgnfcanty n features for MLP and 18 features n neuro-fuzzy cassfcaton. removed from the nput vector. The MLP and neuro-fuzzy neura networks, As seen n Tabe 6, the average accuraces after accuraces were evauated (Tabe 1). The average dong UTA agorthm n MLP neura network and 3511

7 Res. J. App. Sc. Eng. Techno., 6(19): , 013 Tabe 5: Comparson error of coor features n UTA agorthm (Neuro- fuzzy) Feature's error (E) Varetes Rm Gm Bm Hm Sm Vm Ym Cbm Crm I 1 m I m I 3 m KSC KSC KSC KSC KSC Tota (T) Tabe 6: Average accuracy after UTA agorthm Varetes accuracy (%) Neura networks KSC60 KSC403 KSC400 KSC600 KSC704 Average accuracy (%) MLP Neuro-Fuzzy Tabe 7: Dfference of accuraces before and after UTA Varetes accuracy (%) Neura networks KSC60 KSC403 KSC400 KSC600 KSC704 MLP Neuro-Fuzzy neuro-fuzzy were 96% and 95% respectvey. Comparson of varety's accuraces showed that the hghest accuracy n MLP observed n KSC60 varety (100%) and the owest one beonged to KSC400 varety (93%) and n neuro-fuzzy, KSC60 varety (99%) had the hghest accuracy and KSC600 varety (90%) had the owest accuracy. The dfferences between accuraces before and after performng UTA agorthm for MLP and neurofuzzy networks was shown n Tabe 7. In MLP structure, feature seecton reduced accuracy ony n KSC704 varety (-1%) and accuracy of the other varetes ncreased. So feature seecton had postve effect for corn varetes cassfcaton usng MLP neura network. In the neuro-fuzzy case, feature seecton was neffectve n KSC60 varety (0%) and reduced accuraces for the rest varetes. Therefore, the resuts showed that feature seecton for neuro-fuzzy cassfer was not ncreased the average accuraces of corn varetes. As the mater of fact, the average accuracy before and after evauatng the UTA agorthm was near to each other and we thnk that the overa features seecton s acceptabe enough to get the best corn seed cassfcaton wth owest tme and cost by usng the mnmum number of features. CONCLUSION Neura networks can present nternay the knowedge necessary to sove a gven probem. After earnng the neura network's knowedge about sovng probems, t spread n agrcutura scence, so the MLP and neuro-fuzzy neura networks presented for cassfyng 5 corn seed varetes. A 450 sampes seed varetes nvestgated and 7 features extracted from seeds by MATLAB verson 7.8. After feature seecton usng UTA agorthm, the optmum sets of features for two neura networks created ndvduay. It found that, after feature seecton dong n MLP method 8 features (I m, SF, I 3 m, Mnor and Aspect rato, Sm, Permeter and Hm) and n neuro-fuzzy structure, 9 effectve features (Sm, Hm, I 3 m, Convex Area, I 3 m, I 1 m, SF, SF4 and Roundness) extracted among 7 orgna nputs. The hghest accuracy for seeds dentfcaton n MLP (100%) and neuro-fuzzy (99%) beonged to KSC60 varety, So we conducted that feature seecton n MLP ncreased and n neuro-fuzzy decreased average accuraces, however neuro-fuzzy before feature seecton ganed to hghest accuracy average (96%) among a expermented cases. ACKNOWLEDGMENT The authors acknowedge the great hep and assstance provded by Dr Amr Hossen Shran Rad Assocate professor of Seed and Pant Improvement Research Insttute, Dr. Davood Habb Assstant professor of Isamc Azad Unversty, Karaj Branch and cerea Research Department (SPII) for ther support and usefu suggestons. REFERENCES Barker, DA., T.A. Vour, M.R. Hegedus and D.G. Myers, 199a. The use of ray parameters for the dscrmnaton of Austraan wheat varetes. Pant Varet. Seeds, 5(1):

8 Res. J. App. Sc. Eng. Techno., 6(19): , 013 Barker, D.A., T.A. Vour and D.G. Myers, 199b. The use of sce and aspect rato parameters for the dscrmnaton of Austraan wheat varetes. Pant Varet. Seeds, 5(1): Barker, D.A., T.A. Vour and D.G. Myers, 199c. The use of Fourer descrptors for the dscrmnaton of Austraan wheat varetes. Pant Varet. Seeds, 5(1): Barker, D.A, T.A. Vour, M.R. Hegedus and D.G. Myers, 199d. The use of Chebychev coefcents for the dscrmnaton of Austraan wheat varetes. Pant Varet. Seeds, 5(1): Chen, X., Y. Xun, W. L and J. Zhang, 010. Combnng dscrmnant anayss and neura networks for corn varety dentfcaton. Comp. Eectron. Agrc., 71S: Ehert, D.L., B.D. H, D.A. Raworth and B. Estergaard, 008. Artfca neura network modeng to predct cutce crackng n greenhouse peppers and tomatoes. Comp. Eectron. Agrc., 61: Image Processng Toobox for O-Matrx, 007. Reference Manua, Verson 1.0, Anona Labs Ltd., Retreved from: (Accessed on: September 5, 010). Jang, S.D., X. Yang, N. Cnton and N. Wang, 004. An artfca neura network mode for estmatng crop yeds usng remotey sensed nformaton. Int. J. Remote Sens., 5(9): Kantardzc, M., 003. Data Mnng Concepts, Modes, Methods and Agorthms. IEEE, Pscataway, NJ, USA. Majumdar, S. and D.S. Jayas, 000. Cassfcaton of cerea grans usng machne vson. IV. Combned morphoogy, coor and texture modes. Trans. ASAE, 43(6): Movagharnejad, K. and M. Nkzad, 007. Modeng of tomato dryng usng artfca neura network. Comp. Eectron. Agrc., 59: Ohta, Y., Knowedge-Based Interpretaton of Outdoor Natura Coor Scenes. Ptman Pubshng Inc., Marshfed, MA. Pawa, J., N.S. Vsen and D.S. Jayas, 001. Evauaton of neura network archtectures for cerea gran casscaton usng morphoogca features. J. Agrc. Engng. Res., 79(4): Pazok, A.R. and Z. Pazok, 011. Cassfcaton system for ran fed wheat gran cutvars usng artfca neura network. Afrcan J. Botechno., 10(41): Rumehart, D.E., J.L. McCeand and R.J. Wams, Parae Recognton n Modern Computers, n Processng: Exporatons n the Mcrostructure of Cognton, MIT Press, Foundatons, Cambrdge, MA. Rutkowaska, D. and A. Starczewsk, 004. A Mut-NF Approach wth a Hybrd Learnng Agorthm for Cassfcaton. In: Sncak, P., J. Vascak and K. Hrota (Eds.), Machne Integence: Quo Vads, Word Scentfc, Sngapore, pp: 458, ISBN: X. Savn, I.Y., D. Stathaks, T. Negre and V.A. Isaev, 007. Predcton of crop yeds wth the use of neura networks. Russan Agrc. Sc., 33(9): Steenhoek, L.W., M.K. Msra, W.D. Batecheor and J.L. Davdson, 001. Probabstc neura networks for segmentaton of features n corn kerne mages. App. Eng. Agrc., 17(): Symons, S.J. and R.G. Fucher, 1988a. Determnaton of wheat kerne morphoogca varaton by dgta mage anayss, I Varaton n eastern Canadan mng quaty wheats. J. Cerea. Sc., 8: Umbaugh, S.E., 005. Computer Imagng: Dgta Image Anayss and Processng. Tayor and Francs, New York. Uno, Y., S.O. Prasher, R. Laerox, P.K. Goe, Y. Karm, A. Vau and R.M. Pate, 005. Artfca neura networks to predct corn yed from compact arborne spectrographc mager data. Comp. Eectron. Agrc., 47: Utans, J., E. Moody and S. Rehfuss, Input varabe seecton for neura networks: Appcaton to predctng the U.S. Busness cyce. Proceedng of IEEE/IAFE Computatona Integence for Fnanca Engneerng, pp: Wang, L.X., Desgn of Fuzzy Systems Usng Gradent Descent Tranng, In: A Course n Fuzzy Systems and Contro. Prentce Ha PTR, Upper Sadde Rver, N.J., pp: 44, ISBN: Yun, L., 004. Study on Gran Appearance Quaty Inspecton usng Machne Vson. Chna Agrcuture Unversty, Bejng, (In Chnese). Zapotoczny, P., M. Zenskaa and M. Ntab, 008. Appcaton of mage anayss for the vareta cassfcaton of barey: Morphoogca features. J. Cerea Sc., 48: Zhang, W., X.C. Ba and G. Lu, 007. Neura network modeng of ecosystems: A case study on cabbage growth system. Eco. Mode., 01: Zhao-Yan, L., C. Fang, Y. Y-Bn and R. Xu-Qn, 005. Identfcaton of rce seed varetes usng neura network. J. Zhejang Unv. Sc., 6(11):