Synthesis of supervised classification algorithm using intelligent and statistical tools

Ali DOUIK et al /Internat ional Jo urnal on C omputer Sci ence and En gineering Vol.1(2), 2009, 89-9 7 89 Synthesis of supervised classification algorithm using intelligent and statistical tools Supervised Classification Ali DOUIK U.R.: Autom atique Trait ement de Si gnal et Im age (ATSI) Ecole Nationale d’I ngénieurs de Monasti r Monastir – TUNISIE Ali.doui k@enim .rnu.tn Mourad MOUSSA JLASSI U.R.: Automatique Traitement de Signal et Image (ATSI) Ecole Nationale d’I ngénieurs de Monasti r Monastir – TUNISIE Mourad.enim@yahoo.fr Abstract — A fundamental task in detecting fore ground objects in both static and dynamic scenes is to take the best choice of color system representation and the effi cient technique for background modeling. We propose in this pa per a non-parametric algorithm dedicated to segment and to detect objec ts in color images issued from a football sports meeting. Indeed segmentation by pixel concern many app lications and reveal ed how the meth od is robust to detect objects, even in presence of strong shadow s and highlights. In the other hand to refine thei r playing strate gy such as in football, handball, volley ball, Rugby..., the coach need to have a maxim um of technic al-tactics inf ormation ab out the on-going of t he game and the players . We propose in this paper a rang e of algorithms allowing the resolution of many proble ms appearing in the automated proce ss of team identification, where each player is affected to his corresponding team relying on visual data. The developed system was t ested on a m atch of the Tunisian national competition. This work is prominent for many next computer vision studies as i t’s detailed in this study. Keywords-component; Soccer Si ngular value de composition; Classification; artificial intell igence; supervised algorithm; Moments Matri I. I NTRODUCTION In the last ten years, motio n detection and analysis have become very im portant for a wi de range of a pplications, especially since complex algorithms can nowadays be processed real-time. Exam ples of appl ications that use mot ion segmentation t echniques are gesture reco gnition, tracki ng applications [1, 2, 3, 4, 5], video surveillance systems [6, 7], industry, robotic s [14], the medical field [15], aeronaut ics [17], Pattern Recognition [13] and r ecently, sports sector [18]. Although a lot of research has done i n this field on object s segmentation, still a lo t of difficulties h ave to be con sidered in this area, especially to pro duce good results i n changing circumstances. The main purpose of this paper assignment is to present an overview of object s segmentation techni ques and classification. A large variety technique has developed and im proved, K. Karman et al. used Kalman filter to model a dynamic background. Similar ly K. Elgamma l e t al. [9] presented a non- parametric backgr ound model to model dynam ic background. Toyama et. al. [10] used Wiener filter to make a linear prediction of the p ixel intensit y values, given th e pixel historic. C. Wren et. al. [11], use a sing le Gaussian model pe r pixel an d the parameters are updated by alpha blending. Unfortunately, these approaches fail in case the distributio n of the backgrou nd colour values d o not fit into a sin gle model. Y ing Ming et al. [12] worked out a statis tical algo rithm inspired from the id ea of Elgammel based on Cauc hy distribut ion; they prove d that ratios of intensity va lues between the background pixels and the current image pixels are adap ted to Cauchy’s distribut ion. In fact it is characterized by a little w ide form covering the tails of the hist ogram; on the other hand Gaussi an distribut ion has an exponential form. Several works was done concerning cl assification, Pal et al. [21] propose d an SVM techni que, their work reports the resul ts of two expe riments in which multi-class S VMs are compared with Maximum Likelih ood (ML) and Artificial Neur al Network (ANN) methods in term s of classification accuracy; SVM achieves a higher level of classification accuracy than either the ML or the ANN classifier. Classification by artificial visi on in soccer sector has been largely mediatized and became a significant research topic. The result of a m atch has serious consequences on t he club life and its external environm ent (medi a, sponsors...). To refi ne their play strategy [20], coach and the l eaders need to have technical-tact ics and relevant i nformation [19] about events of the play as well as of the player s. Indeed the use of the color in computer vision application is yet very recent, musical field [22], metals classification [23] , road scenes analysis and sensing domain [16, 2 9, 8]. In this paper various su pervised classification techniques were applied. They are based on intelligent too ls as fuzzy and ne uronal classification on the one hand, statistic and hybrid classifi cation based respectively on moments difference and determination of three significant color component s on the othe r hand. A comparative study about player recognition rates wa s elaborated enabling us to ISSN : 0975-3397 Ali DOUIK et al /Internat ional Jo urnal on C omputer Sci ence and En gineering Vol.1(2), 2009, 89-9 7 90 conceive an adequate met hod for football play ers classification at the aim to classify each player in his suitable class automatically. II. S EGMENTATION TECHNIQUES E ASE OF U SE A. Detection by hi stogram an alysis In artifici al vision field , colour im ages are taken by a video camera and then di gitized by a computer. Since the soccer video is taken by static camera, the supporter and useless informat ion can be rem oved by delim iting the playing z one with an affine function . Ob jects identification in co lour images is a relevant stage in classifi cation study; therefore we begin by the background sub traction to detect players and th en to classify each of them in its su itable class. The algorithm is based on th e following stag es:  Convert the origin al image to standards rgb lev els removing t he light refl ections.  Detect the high and low th reshol ds of each histogram then carry out the thre sh old on the three chrom atic levels.  Apply a logical operator “AND” on the three levels and the original im ages. B. Detection by statist ical lear ning Detection by histogram analysis cons ists to segment colour images and to rem ove useless inform ation that have no contribution in classification ph ase. This method isn’t a good choice of se gmentati on for m any applicat ions as presented i n figure 2. The major problem of this technique corr esponds at the time when t he background an d foreground have the sam e characteristics, hence after t hresh olding hi stograms m any false detection can be occurred: f or example i t can remove a large part of usef ul inform ation hence we s hould devel op an appropriate technique. Because the parametric b ackground m odel still lacks flexibility when dealing with non-static backgrou nds, a highly flexible non-parametric tech nique is proposed for estimating background probabilities from ma ny recent samples over time using Kernel den sity estimation. In the non-parametric mode l all recently observed pixel values x 1 , x 2 …x N are modeled by probabilit y density fun ctions using a certain kernel estim ator function, which is often chosen to be a Ga ussian. The weighte d sum of all t hese Gaussians results in the final probab ility density function of the pixel value x t :  The kernel estimator fu nction K i s chosen t o be a N ormal function (0 , ) N  with  being the kernel function bandwidt h. For simplicity reasons the color ch annels within  are assumed independen t, but each with their own kernel band width 2 j  . Because of these assumptions the final density estimation can then be written as:  When this probab ility is below a certain threshold, th e pixel is classified as a foreground pixel. The threshold can be adjusted to achiev e a desired proportion of false positiv es. In other kernel density estim ation applications, t he kernel bandwidth has t o be dependent on the num ber of samples. When there are many samples the bandwi dth should be sm aller than if we have few one. However, in th is case tempor al properties are taken into a ccount for determi ning the 2 j  of color channel j. For each color c hannel the median m of the absolute differences of each c onsecutive pairs of samples is calculated. We estimate  by:  This method guaranties t hat the local devi ation is large when there are many large ju mps between consecutive samples and smaller when this is not the case. a b c d e f Figure 1. (a, b) Represent original images, (c, d) Represent delimited images, (e, f) Represent segmented images. Figure 2. Histograms of standards RGB levels (r, g, b) C. Singular Value De composition Approach [24] 1) Introduct ion A non-parame tric backgro und modeli ng technique, has been applied on a soccer vide o images, the main problem that can be appears is th e occurr ing of wrong detection p ixels, indeed shadow pixels are det ected as movin g objects resul t to 1 1 () ( ) N tt i i px K x x n    2 () 2 2 11 Pr( ) 1 2 1 2 xx tj ij k j d xe t j K i j         0, 68 2 m   ISSN : 0975-3397 Ali DOUIK et al /Inter national Journal on C omput er Science and E ngineeri ng Vol.1( 2), 2009, 8 9-97 91 an over segmentation that will be damage, many later work s where this paper it’s reg istered, however this algorithm is extremely important becaus e it’s a part of players classification a nd tracking on a soccer video. Two major issues in the SVD technique: (1) carryin g out a mathematical approach and (2) explain m ain adva ntage of the method pr oposed here a nd sh owing infl uences of t he singul ar values choice on a treated outpu t image, besides we will see the prominent contrib ution usin g SVD theory to restore and eliminate shadow, highlight s and noise from camera displacement and changed circum stances. 2) SVD approxi mation of an im age The main obje ctive of backgr ound segment ation techniq ue is to use singular value deco mposition of a given image A represented by a matrix A p = [a ij ], when it ca n be de composed into a product of three ma trix T kk k US V as shown in figure 3. Wher e a ij is the appearance freq uency of chromat icity and intensity of backg round pi xels (p= r ed, gree n, blue). Figure 3. Singular values decomposition of the Matrix A. SVD technique consists to redu ce the size for each initial chromati c level fr om r to k ra nk by suppression of r -k colu mn. Matrix S represents diagonally the singular values classified by decreasing or ders. Low val ues have no infl uence on the total energy of A. U k and V k are orthogonal matrices issued from matri ces U and V. T he determi ned singular values for each plan were presented in frequency space. There representation proves that for each one corresponds a discrete frequency. The noise that can occur in t he signal (in fre quency space the am plitude of noise i s constant) c orresponds to a l ow amplit ude of singular value whe reas high am plitudes of these represents glob al signal ene rgy. 3) Confidence intervals research In this section, we describe th e basic background m odel and the background s ubtraction process wi th singular value decompositi on. It’s both used in t he restoration or t he reconstruction of an im age, to increase the compactness distribution o f different classes and also t o provide useful image information. To evaluate mathe matical contribu tion of SVD , a quantification of global signal energy distribu tion according to the weight of each singular value S kk was done. The figure 4 illustrates the en ergy distribution E defined by:   The relative energy contained by each singular value K, noted p k is defined by:  Where the energy of the K singular value is e qual to 2 . kk S Figure 4. Weight distribution showing 7 dom inant representing 99 % of the signal energy among 108 weights As it’s shown in figures 5a, 5b and 5c, the size of treated image will be deduce d from th e curves representing standard déviation of each colour levels according to the singular value decompositi on. In fact a good c hoice of the size leads to reduce both compact ness in different di stributions and i n computi ng time. According t o figures 5a, 5b an d 5c, we can de note two zones, the first one defined in the interval   0, kkl i S     where   kkl i S is the singular value lim its corresponding to the l inear part of the cur ve (i = red, bl ue , green), in t his zone the curve presents a slope, beyond   kkl i S a second zone appears where the standard dev iation varied slig htly therefore th e optimal singular value   ˆ kkl i S must necessaril y belong to the first zone of each curve. a b c Figure 5. Evaluation of standard deviati on according to singular values respectively the red, green and blue channel. Table1 illustrates in itial and improved stan dard deviations for three channels (RGB). TABLE I. E VALUATION OF IMPROVEMENT PARAMETERS R 4.2044 119.61 1156.9 G 4.7227 152.08 1073.3 B 4.313 88.988 1141.7 R_SVD 3.9274 119.07 1229.8 G_SVD 4.6204 152.05 1104.2 B_SVD 3.9954 88.407 1205.5 2 1 k k i EA    2 kk k S p E  ISSN : 0975-3397 Ali DOUIK et al /Internati onal Jour nal on Com puter Science and Engineer ing Vol.1( 2), 2009, 8 9-97 92 The choice of singular values will be kept depending on two issues: the first one is the energy c urve evaluat ed by figure 4 and the second one is the standard deviation curves of each chromatic level shown i n figures 5a, 5b and 5c. In fact we specify for each component the si ngular value lim it previewed. The table 2 shows con fidence interv als as well as the limit and the optimal si ngular values that vary from a level to another. TABLE II. S PECIFICATION OF CONFIDENCE INTERVALS   kkl S  kk ˆ S confidence intervals Red plan 29 19 [0, 29] Green plan 28 13 [0, 28] Blue plan 19 13 [0, 19] 4) Foreground segmen tation and shado w suppression Using th e probability Pr( ) x t calculated in equat ion 2, the pixel is considered as a for eground pixel if Pr( ) x th t  . The threshold th is a global threshold over al l the image that can be adjusted to achieve a d esired percentage of false positives. The shadows detection as foregroun d regions is a source of confusion for subseq uent phases of analy sis. Color inform ation [18] is usef ul for shadows suppression by separati ng color fr om lightness inform ation. For a gi ven three color vari ables, R, G and B, the chromaticity coordinat es r, g and b(r=R/(R+G+B), g=G/(R+G+B), S=(R+G +B)/3). Consider the case where the background is co mpletely static , and let the expected value for a pixel be  ,, rg s ii i . Assume that this pixel is covere d by shadow and let  ,, rg s tt t be the observed val ue for this pixel at this frame. Then it is expected that 12 / th s s th t i  . To prove robustness of this algo rithm, the figure 6 illustrates different player window s and shows how we can overcome segment ation (rem oving shadow pixels and pixels background hav ing the same characteristics that th ose foreground pixels). Figure 6. Effect of the singular value decomposition level. Where (a and b) are Original images, (c and d) Dete ct ion using chromaticity coordinates r, g and the lightness variable s. (e and f) Detection using chr omaticity coordinates r, g and s with SVD. III. C LASSIFICATION ALGORITH MS A large variety of s upervised classification a lgorithm s was developed, ranged from statistica l to intellig ent tools and they operate on only color regions. Each one of them is expressed in adapted colo r system representation th at will contribute to an optimal cl assification. A. Hybrid classification The advantage of this algorith m is that the color w ill be represented in a syst em of the three most discriminati ng levels, in order to be able to separate t he colour nuance di stributions corresponding to pixel players from each team. 1) Hybrid col our system After extraction of useful inform ation which represe nts the pixel players, we separate the two classes using col orimetric analysis. Nevertheless traditional RGB space cannot be the most discriminating representati on space. Indeed, other c olour systems, deduced from the RGB components, can be more suitable according to the consid ered case. For this reason, treatment of pi xel players in various c olour systems l eads to a hybrid space represented by the three best colour com ponents. 2) Method descript ion N.Vandenbro ucke [29] considered a mu ltidimensional space composed of the chromatic levels currently used as the following: E = {R , V , B , r , v , b, X, Y, Z , I1 , I 2, I3 , y , i , q , u , v , l , t , s }. In each level α ( α  E) the algorithm of discrimi nation is based on various phases:  Select three training player windows J 1A , J 2A and J 1B in the RGB system.  Convert in the α level player windows.  Calculate the average of pixel coordinates (x, y) representing each player: α (x, y): pixel value (x, y) in the plan α . R 1A and S (R 1A ) are respectively area and surface of the player J 1A . R 2A and S (R 2A ) are respectively area and surface of the player J 2A . R 1B and S (R 1B ) are respectively area and surface of the player J 1B . Therefore we can evaluate for each level α : α -average re gion for J 1A : 1 1 1 (, ) () A R A A x y m SR     α -average re gion for J 2A : 2 2 2 (, ) () A R A A x y m SR     α -average re gion for J 1B : 1 1 1 (, ) () B R B B x y m SR      Calculate the distance bet ween J 1A and J 2A as well as the distance between J 1A and J 1B a b c d e f ISSN : 0975-3397 Ali DOUIK et al /Internati onal Jour nal on Com puter Science and Engineer ing Vol.1( 2), 2009, 8 9-97 93 1, 2 1 2 AA A A Dm m     1, 1 1 1 AB A B Dm m    (7) Where these expression 1, 1 AB D  and 1, 2 AA D  are respectively the distance between J 1A ,J 2A and J 1A ,J 1B in α level. The discriminati ng criterion adopt ed is determ ined as the difference D  between these two distances: 1, 1 1, 2 AB A A DD D    (8) We proceed with the same w ay for all α level from E set. The criterion value Classification in a descending or der lead to the determina tion of the most d iscriminating chromatic components (tabl e 3). TABLE III. E VALUATION OF DISCRIMINATE L EVELS α β γ Hybrid color space Saturation Green Blue 3) Region model ling After the conversion of RGB im age for the players A and B in the hybrid s ystem, the m odelling phase consist s in assigning to each P(x, y) of a level th e average value of the pixels intensities for t he corresponding l evel [30]. 4) Parameters and metho dology The attributes used in this al gorithm are the average values of each colorimetric component wh ereas the criteria of decision making are the Euclidian distance between A (s, v, b) a nd JA where A the coordinat es for model A and JA current player t o be classified (table 4) . Similarly we calculate the distances separati ng the model B from JB. TABLE IV. E UCLIDIAN DISTANCE BETWEEN T WO OBJECTS AND M ODELS d2 (B, J A ) 0,54616 d2 (A, J A ) 0,063875 d2 (B, J B ) 0,022593 d2 (A, J B ) 0,63185 The membership’s decision of each player belonging to a team is specified by evaluati ng the distance that separates player window and the t wo models. Ind eed, a very weak distance corresponds to pla yers of the same clust er; the opposite case corresponds t o two players of d ifferent cluster. The results of classification by this method are illustrated in tables 5 with a very encouragi ng classification rate of around 93 %. TABLE V. C LASSIFICATION RATE FOR THE H YBRID TECHNIQUE Classe Total number of players Number of classified players Percentage of classification Team A 210 201 95 % Team B 210 195 92 % B. Difference momen ts algorithm (dm om) The colour moments are m easurements that can be employed t o differentia te images based on their colors characteristics. They provide a measurement of the similarity between color images. The parame ter values of each model can be compared with the images constituting database. The colour moments [26] can be c onsidered as a di scriminati ng criterion between texture and c olor obj ects. In fact the colour distributi on for these regi ons follows a certain densi ty of probability (Gaussian ). 1) Used paramete rs In this m ethod Sticker a nd Orengo [27] em ployed three central colour moments expressed as follows:  The average: 1 1 N E P ii j N j    Where E i is the average value of the player pixels to be classified  Standard deviat ion: 2 )) 1 (( 1 i E N P ii j N j      i  represents the dispersion degr ee between pixels player and their average.  Skewness: 3 3 () ) 1 ( 1 i E N SP ii j N j     Si represents the asymmetry de gree between pixels player and their average. 2) Methodology This method i s carried out i n HSV sy stem (Hue, Sat uration and Value). The moments of candi date image that are fixed to nine are evaluated for each leve l. The discriminating criterion between two images (reference a nd candidate image) is defined as the sum of differences bet ween the moments distri butions with a weighting fact or expressed by the f ollowing equation: (, ) 1 1 23 N A B dmom JA M B W E E ii i j A BA B WW S S ii i i ii         dmom: represents the dist ance between two players. ISSN : 0975-3397 Ali DOUIK et al /Inter national Journal on C omputer Sci ence and En gineeri ng Vol.1( 2), 2009, 89 -97 94 M A : Model for Player J A . w i : are the weights to be specified which are related to the specific case, they can be grante d so that various prefe rences are given to various attribu tes of an im age. They can be modified to increase or d ecrease the im portance of a colorimetri c component wh ich appears to be i nteresting. Classification by this method is carried out through se veral stages as follows: Stage 1: Convert image RGB to HSV. Stage 2: Calculate the moments matrices of the two models (M A , M B ). Stage 3: Calculate the moments matrix of to classify each player. Stage 4: Calculate the dmom for the foll owing m atrix of weights: 1 2 1 1 2 1 1 2 1 W       We calculate dmo m between two m odels and a gi ven object (player). It is well noticed that these d istances to an attribu te can contrib ute to m ake d ecision on classification. TABLE VI. T HE D MOM BETWEEN TWO OBJECT S AND MODELS dmom (B, JA) 7,0757 dmom (B, JB) 1,7987 dmom (A, JA) 1,8814 dmom (A, JB) 5,3967 3) Experimental results The classification by moments difference technique for two clusters are elaborated , the results are illustr ated in table 7, the global classification rate for two classes reaches 78 %. TABLE VII. C LASSIFICATION RATE FOR THE DMOM TECHNIQUE Class The total number of players Number of classified players Percentage of classification Team A 210 129 62 % Team B 210 200 95 % C. Fuzzy classification [31] Fuzzy logic is frequentl y used in comput er vision, it m ay affect various applicat ions. The most common among t hese are regulation, cont rol and classificati on. Many fuzzy system s can be used in this cont ext: Sugeno m odel, Tsukam oto model and Mamdani model (u sed in this algorith m). 1) Statisti cal study With an ex tensible colour images database ta ken during th e warming up tim e (at the beginning of the m atch), a statistical study [16] was elaborat ed allowing to establi sh correlation between parameters during cl assification. Fi gure 7a, 7b and 7c represents respectively th e distributio n curves of the intensities average values of the green (gr een_m oy), the blue (blue_moy ) and the red (re d_moy). Figure 7. The intensities distribution of th e three components (RGB) for two clusters The distributi on curves highlight the importa nce of each parameter during cl assification. Only the parameter red_m o y presents an overlapping (Fig.7 c), the two other parameters green_moy, blue_m oy (Fig.7a and Fig.7 b) haven't it , thus a statistical classificat ion [25] can be used in thi s case. 2) Methodology The fuzzy classification is done rely ing on the f ollowing stages:  Fuzzification: this st age is used to quantify the i nput and output variables (Fig.8). Th is stage consists to define the membership funct ions of the lingui stic variables. Figure 8. Membership Functions of the input and output param eters 0, 41 0, 36 0, 35 0, 22 0, 16 0, 19 1. 25 1. 19 1. 19 A M        0. 76 0. 70 0. 74 0. 10 0. 08 0. 10 0. 01 0. 01 0. 01 B M        Moments Matrix of Player (J A ) Moments Matrix of Player (J B ) 0. 71 0. 66 0. 70 0. 08 0. 07 0. 88 0. 31 0. 28 0. 28 MA         0. 40 0. 33 0. 32 0. 25 0. 17 0. 20 0. 84 0. 74 0. 74 MB        a b c d ISSN : 0975-3397 Ali DOUIK et al /Inter national Journal on C omputer Sci ence and En gineeri ng Vol.1( 2), 2009, 89 -97 95  Inference rules: is a set of rules between the fuzzy subsets at the aim to dr aw deductions.  Defuzzification: specifies the s uitable out put m odel according the inference rules, in fact the decision was made by calculating the gravity centre of the resulting function as illustrated in the fo llowing equation:  3) Experimental results of classification The classification by f uzzy technique gives the resul ts illustrated in tab le 8. The global classificati on rate for the two classes reaches 97 %. TABLE VIII. C LASSIFICATION RATE FOR THE F UZZY TECHNIQUE Class The total number of players Number of classified players Percentage of classification Team A 210 206 98 % Team B 210 203 96 % D. Neural Networks classification algorithm (NN) [32] 1) Introduct ion Classification by artifici al vision [2 8] in sport s field became a significant research topic; it tr eated several problems that can result in an overlapping between vari ous classes. The neural network is an inform ation processing sy stem inspired f rom cells organization of the human brai n. In a neural network as represented in figure 9 , we distingui sh three types of neu rons:  Input neurons: they have the property to gather data whose source is apart from the network.  Output neuron s: define the output l ayer of the networ k. It contains as ma ny neurons as the number of cl asses to be discriminat ed.  Hidden neuron s: they don’t h ave any relation wi th the external world, it is used to coordinate between th e input and outp ut neurons. Figure 9. Neural Networks architecture Similarly to human b rain, the artificial neural n etworks can be learned by experim ent: Indeed, the ob jective of supe rvised retropropagati on algorithm is to minimize a cost functi on 'E' representing a quadrat ic error for an in put-output. Where d K the desired out put for the K th neuron and S K the obtained output by the network [2 0]. 2 () kk k Ed s     The weight is obtained acco rding equation (12) where  α is a positive real that sp ecifies the step of weig hts modification. 1 .. tt ij ij i j WW C S        To carry out cl assification, w e highlight relati ons between objects on the one ha nd, objects and thei r parameters on the other hand. In this secti on we present a classificati on by Neural Network (NN) l earned with ret ropropagat ion algorit hm. This technique allows discrim inating clusters (players) present in color images issued from sports meeting. It can be summarized in 4 stages shown in fi gure10. Figure 10. Regions classification Chain 2) Training ph ase The used parame ters to classify play er windows by t his technique are the sam e used in dmom technique. For the training phase 9 param eters are used, th e initial base was constituted by 560 regions in va rious positions am ong this base 300 represent two clust ers C1 and C2, 260 represent respectively the C3 and C4 cl usters. Indeed, NN allowed weights adjustm ent by retro-pr opagation algori thm until reaching a null error defined in la st section. The training phase was done as follows: we select ed several player wi ndows from the training datab ase, and extracted p arameters will b e used to classify these regions. The player windows con tains respectively four classes C1, C2 , C3 and C4 then the desired output are indexed res pectively by 1, 2, 3 and 4. To appreciate the classification algorithm, th e curves representing training error according epochs number on one hand in figure 11 and error according neurons of the hidden layer on the other hand has been evaluated i n Figure 12. () () s zz d z Zs s zd z      Numerical ima g e Parameters extraction Neural networks Decision module: Classification ISSN : 0975-3397 Ali DOUIK et al /Inter national Journal on C omputer Sci ence and E ngineeri ng Vol.1( 2), 2009, 89 -97 96 The choice of the neur ons num ber in the hidden l ayer depends on the t raining phase. Indeed, we must learn netwo rks some models in many circum stance and various positions, each experience contains a differ ent neurons number from hidden units, the choice taken is the first one leading to a w eak training error. Figure 12 shows variation of training error according to neurons number i n hidden laye r. Figure 11. Evaluation of training error ac cording to epochs number Figure 12. Evalution of training error accord ing to neurons num ber in hidden layer 3) Validation phase In this stage we worked with a large amount o f samples from our dat a base with the sam e color classifi cation parameters. We made the simulatio n of the network that is already created in the training stage. In the v alidation netw ork stage we evaluated by testing 546 pl ayers of the vari ous classes and for the t wo sports m eeting dividing it as f ollows: 161, 167, 114 and 104 play er windows respectivel y for C1, C2, C3 and C4 clusters. Figure 13 represent s many colour regions detect ed by the segmentation alg orithm. This test allows ap preciating the performances of t he neuronal syst em. If the perform ances are not satisfactory, it w ill be necessary eith er to modify the network architect ure, or to m odify the t raining base. The obtained classification results are illustrated in table 9. Another test w as carried out by widening the traini ng data base by increasing i t from 130 t o 180 player wi ndows for e ach class of second meeti ng and from 150 t o 200 for the two other classes, this test lead s to improve th e recognition rate in Table 10, indeed it reaches 100 % for the m ost classes (C1, C3, C4). Figure 13. Player windows for each cluster for many positions TABLE IX. C LASSIFICATION RATE FOR THE (NN) TECHNIQUE IN THE 1 ST TEST Classe The total number of players Number of classified players Percentage of classification C 1 161 140 87 % C 2 167 135 81 % C 3 114 100 88 % C 4 104 89 86 % TABLE X. C LASSIFICATION RATE FOR THE NN TECHNIQUE IN THE 2 ND TEST Class The total number of players Number of classified players Percentage of classification C 1 111 111 100 % C 2 117 114 97 % C 3 64 64 100 % C 4 54 54 100 % IV. C ONCLUSION In this paper, many algorithms were elaborat ed to detect objects in colour images sequen ces issued from sports meeti ng. Several automatic classif ication sy stems were made to classify different color regio n representing foot ball players. Al l classification techniques de veloped are supervised, we can discriminate on t he one hand intelli gent techniques such as neuronal and fuzzy algorithms and on the other hand hybri d algorithm based o n the determ ination of a col or representati on system contai ning the three most signi ficant com ponents and using a metric distance in this base as a decision m aking approach taking in to account suitable established models. In fact, we can verify in this p aper that RGB system is n ot always the suitable system for solving several problems raised by the researchers in computer v ision field ( e.g. overlappi ng, screening…). To overcom e these problems, we determined three discriminat ing components (v, B, S) whic h are not from the same represent ation syst em and leads to a p romising classification rates for various color regions. Indeed, we obtained with t he intelli gent tools good rates, 97% an d 100% respectively for fuzzy and ne uronal algo rithms, while for hybrid technique it reaches a rate of 98%. V. R EFERENCE [1] A. Senior. "Tracking with Probabilistic Appearance Models" , ECCV workshop on Performance Evaluati on of Tracking and Surveillance Systems, pp. 48-55, 2002. [2] C. Stauffer and W.E.L. Grims on. " Adaptive Background Mixture Models for Real-time Tracking", Pr oc. IEEE Int'l Conf. on Com puter Vision and Pattern Recogniti on: pp. 246- 252, 1999. [3] S.J. McKenna, S. Jabri, Z. Duric, A. Rosenfeld and H. Wechsler, "Tracking Groups of People", Com puter Vision and Image Understanding 80, pp. 42- 56, 2000. [4] R. Cucchiara, C. Grana, M . Picardi and A. Prati, "Detecting Moving Objects, Ghosts, and Shadows in Vi deo Streams", IEEE Trans. Pattern Analysis and Machine Intelligen ce. 25(10): pp. 133 7- 1342, 2003. [5] A. Mittal and L.S. Davis, "M2 T racker: A Multi-View Approach to Segmenting and Tracking People in a Clutter Scene", International Journal of Computer Vision. 51( 3): pp. 189-203, 2003. ISSN : 0975-3397 Ali DOUIK et al /Inter national Journal on C omputer Sci ence and E ngineeri ng Vol.1( 2), 2009, 89 -97 97 [6] D. Gutchess, M. Trajkovic, E. Cohen-So lal, D. Ly ons and A. K. Jain, "A Background Model Initialization Algor ithm for Video Surveillance", IEEE Int'l Conference on Com puter Vision, 2001. [7] I. Haritaoglu, D. Harwood and L. S. Davis, "W4: Real- Time Surveillance of People and Their Activities", IEEE Trans. Pattern Analysis and Machine Intellig ence. 22(8): pp. 809-830, 2000. [8] C.A. Shah, P. K. Varshney, and M. K. Arora, "ICA mixtur e model algorithm for unsupervised classifica tion of remote sensing im agery", International Journal of Remote Se nsing, 28(8), pp. 1711 -1731, 2007. [9] A. Elgammal, R. Duraiswami, D. Harwood, and L. S. Davis, "Background and Foregr ound Modeling Using Nonparametr ic Kernel Density Estimation for Visual Surv eillance”, Proc . IEEE, 90(7), pp. 1151–1163, 2002. [10] K. Toyama, J. Krumm , B. Brumitt, and B. Meye rs. "Wallflower: principles and practice of bac kground m aintenance", ICCV, pp. 255– 261, 1999. [11] C. Wren, A. Azar bayejani , T. Darr ell, and A. Pentland, " Real-time tracking of the human body", IEEE Trans. Pattern Anal . Machine Intell., 19, pp. 780-785, 1997. [12] Y. Ming, J. Jiang, J. Ming, "Background Modeling and Subtraction Using a Local-Linear-Dependence-Ba sed Cauchy Statistical Model", Proc. VIIth Digital Image Computing: Techniques and Applications, Sun C., Talbot H., Ourselin S. and Ad riaansen T. (Eds.), pp. 469-47 8, 2003. [13] Z. Zhao, A. Vashist, A. Elgamm al, I. Muchnik, C. Kulikowski, "Combinatorial and statistical met hods for part selection for object recognition", International Journal of Computer Mathem atics, 84(9), Computer Vision and Pattern Rec ognition, pp. 1285- 1297. 2007. [14] J. B. Hayet, "Contribution to the navigation of a mobi le robot on visual land- marks textured in a struct ured environment", PHD thesis, university Paul Sabatier, Toulouse, Janvier 2003. [15] M. Antonie, O.R. Zaiane, and A. Coman, "Application of data mining techniques for medical image classi fication", Proc. 2nd Intl. Wor kshop on Multimedia Data mining (MDM/KDD’2001) in conjunction with Seventh ACM SIGKDD, pp. 94-101, August 2001. [16] L. Durieux, J. Kropácek, G. D. Grandi , and F. Achard, " Object-oriented and textural image classification of the Siberia GBFM radar mosaic combined with MERIS imagery for continental scale land cover mapping", International Journal of Rem ote Sensing, 28(18), pp. 4175- 4182. [17] J. Ferryman, M. Borg, D. Thir de, F. Fusier, V. Valentin, F. Brem ond, M. Thonnat, J. Aguilera, and M. Kamp el, "Automated Scene Understanding for Airport Apro ns", In Austra lia n Joint Conference on Artificial Intelligence, pp. 593–503, Sydney, Australia, 2005. [18] [18] T. Misu, S. Gohshi, Y. Izumi, Y. Fujita, and M. Naemura, "Robust tracking of athletes using multiple feat ures of multiple views", Journal of WSCG, 12 (1–3), pp. 285–292, 2004 . [19] [19] J.Assfalg, M. Bertini, C. Colom bo, A. Del bimbo, W. Nunziati, "Semantic annotation of soccer videos: Automatic highlights identification", Computer Vision and Image understanding, pp. 285-305, Elsevier 2003. [20] F. Ramos, A. Junco and E. Espinosa, "S occer strategies that live in the B2B world of negotiation and decisi on-ma king", International Journal on Decision Support System s, 35, pp. 287– 310(200 3). [21] M.Pal, and P. M. Mather, "Support vect or m achines for classification in remote sensing", International Journa l of Remote Sensing, 26(5), pp. 1007-1011, 2005. [22] A. Meng, and J. Larsen, "Decisi on time horizon for music genr e classification using short time features", EUSIPCO, pp. 1293–1 296, Vienna, Austria, 2004. [23] K.Y. Choi, S.S. Kim, "Morphological an alysis and classification of types of surface corrosion damage by digital image processing", Corrosion Science, 45, pp. 1-15, 2005. [24] M. MOUSSA JLASSI, A. Douik, H. MESSAOUD, "Non-parametric Motion segmentation In Soccer Video Images", 2nd International Conference on Electrical Engine ering Design and Technologies, Hammamet, Tunisia, October 8- 10, 2008,. [25] [25] J. Li and James Z. Wang, "Autom atic linguistic indexing of pictures by a statistical modeling appro ach" IEEE Transaction on Pattern Analysis and Machine Inte lligence, 25(9): 1075–1088, 2003. [26] [26] H. Zhang and J. Feng, "Color textur e moment for content based image retrieval", Proc. IEEE Intl Conf. on Im age Processing, September, 2002. [27] M. Stricker, M. Orengo, "Sim ilarity of Color Images", Proc. Storage and Retrieval for Image and Video Databases III, February 1995. [28] C. Vasquez, A. Hernandez, F. Mora , G. Carrault, an d G. Passariello, "Atrial activity enhancement by wiener filtering using an artificial neural network", IEEE Trans. Biomed. Eng., 48, pp. 940–9 44, Aug. 2001. [29] S. L.Benfield, H. M. Guzman, , J. M. Mair, and J. A. T . Young, " Mapping the distribution of coral reefs and associated sublittoral habitats in Pacific Panama: a comparison of optical satellite sensors and classification methodologies", Interna tional Jour nal of Remote Sensing, 28(22), pp. 5047- 5070, 2007. [30] Z. Jing, J. Tianzi, L . Bing, J. Xingpeng and Z. Huizhi. " Systematic benchmarking of m icroarray data feat ure extraction and classification". International Journal of Computer Mathematics, 85( 5) , pp. 803-811, May 2008. [31] M. Moussa Jlassi, A. Douik and M. Annabi, "Synthesis of Classification Supervised Algorithms for Players Identification Duri ng a Sports Meeting", 14th IEEE International Conference on El ectronics, Circuits and Systems, ICECS’07, Mar rakech – Morocco, December 11- 14, 2007. [32] M. Moussa Jlassi, A. Douik, H. Messaoud, "Classification Supervisée par Approche Neuronale Applicati on: reconnaissanc e des joueurs de football", 5èm e Conférence Intern ationale d’Electrotechnique et d’Automatique, JTEA’08, Hammamet – Tunisie, Mai 02-04, 2008. ISSN : 0975-3397