présenté par: a.djeffal

Download Présenté par:  A.DJEFFAL

If you can't read please download the document

Upload: yetty

Post on 10-Jan-2016

36 views

Category:

Documents


1 download

DESCRIPTION

Departement d’Informatique. Soutenance de Magister en Informatique. Option: I.A & Image. Thème:. Compression des images animées par estimation de mouvement. Présenté par: A.DJEFFAL. Devant le jury: Pr M.C.Baatouche Président Dr M.K.Kholadi Examinateur - PowerPoint PPT Presentation

TRANSCRIPT

  • Prsent par: A.DJEFFALDepartement dInformatique Soutenance de Magister en InformatiqueThme:Devant le jury: Pr M.C.Baatouche Prsident Dr M.K.Kholadi Examinateur Dr A.Boukarem Examinateur Dr A.Khelifa ExaminateurOption: I.A & ImageEncadr par: Dr N.DJEDI

  • PlanIntroductionCompression vidoEstimation de mouvementBlock matchingMthode hybrideConclusion & perspectives

  • IntroductionLes nouvelles technologies utilisent les images animes (squences vido)

  • IntroductionCependant, la manipulation des images ncssite des espaces et des dbits trs importants 180 x 240 x 3 126 KOctets1 minute, 25 images/seconde 180 x 240 x 3 x 25 x 60 Octets 185 MOctets !!! Do la ncessit de la compression Les solutions existantes ne sont pas suffisantes pour les nouvelles applications temps rel tel que les tlphones mobiles, le vido chat, Do la ncessit de nouvelles techniques de compression plus rapides

  • Compression vidoPrincipe: Eliminer la redondance Spatiale Compression des images fixes Eliminer la redondance Temporelle Estimation de mouvement

  • Codage des images fixes (codage IntraFrame)GIF, CIF, JPEG,Codage JPEG (Joint Photo Expert Group)RVBY:LuminanceU,V:ChrominanceU/4, V/4(Nbre de 0s, Premire valeur non nulle qui suit)(0,0): fin du blocCompression vido

  • Compression vidoCodage InterFrameH261,H263,H26L, MPEG1..MPEG21, I(F1)I(F2 - F1)I(F3 F2)I(F4 F3), Codage Prdictif (Images P)Coder une image par rapport une image de rfrence Parfois les infos coder se trouvent dans les images futures et non dans les images passes Codage Bidirectionnel (Images B)

  • Schma dun encodeur vido (MPEG)IBBPBBPIBBPIEntre:Sortie:JPEGJPEG Inv100110100110BBPEstimation de mouvement011011011011Compensation de mouvementIPB111001B000110BCompression vido

  • Caractristique dun encodeur vidoCompression vido1. Taux de CompressionOu encore2. Vitesse de codage : EM +60%3. Dbit4. Qualit

  • Estimation de movementVecteur de mouvement Comment calculer la diffrence entre une image et une image de rfrence ? Calculer le dplacement des objets entre les deux images : Calculer les vecteurs de mouvements Comment reprsenter les objets ?

  • Calculer le vecteur de mouvement de chaque pixelObjet=PixelEquation du flot optiquePixel rcursifEstimation de mouvementMthodes d EM Temps de calcul exhaustif, inadmissible pour les applications temps rel.2. Segmenter limageObjet=Segment Temps de segmentation exhaustif, inadmissible pour les applications temps rel. Problmes de segmentationsChevauchement des objets,Difficult dimplmentation Hard, 3. Considrer les objets comme des blocs non chevauchsObjet=Bloc

    Correspondance de blocs (Block Matching 16x16) Rapide, Efficace, Utilise par la plupart des standards,

  • Block matchingPrincipe:Mesure de Distorsion du Block (MDB):MinMDBMaxMDBIdentiqueRessembleDiffrentMDB:Zone de rechercheBlocs candidats: selon le type de mouvement de la squence (large, rduit ou stationnaire)

  • Block MatchingAlgorithmesObjectifs: Augmenter le taux de compression Minimiser le temps de calcul Augmenter la qualit des images Comment ? Dimensions de la zone de recherche Choix des blocs candidats (nombre & dispersion) Seuils dacceptabilit

  • Full SearchTemps de calcul exhaustif !Three Step Search (3SS)Block MatchingAlgorithmes1111111119 Points: Central & ses 8 voisinsDistance: w/2Trouver le meilleur ressemblantUtiliser le meilleur comme centreDistance/2, slectionner 8 nouveauxRpter lalgorithme 3 fois22222222333333335 points: central & ses 4 voisinsDistance : w/2Trouver le meilleur ressemblantSi meilleur = centre ou bords alors Distance/25 nouveaux points autour du meilleur prcdent Lorsque distance=1, Utiliser 9 points, arrter.111112223344444444Recherche logarithmique 2D (2DLog)

  • Descent gradient SearchBlock MatchingAlgorithmes9 points: central & ses 8 voisinsDistance : 1Trouver le meilleur ressemblantSi meilleur = centre ou bords alors Arrter8 nouveaux points autour du meilleur prcdent Recommencer11111111122233333Four Step Search (4SS)111111111Point centre et ses 8 voisinsPas=W/4Si meilleur=centre alors Pas = 1 Select 8 nouveaux voisins FinSinon Continuer avec le nouveau centre et le mme pasSi la procdure est rpte troisfois alors, Pas = 1 Select 8 nouveaux voisins Fin222223333344444444Autres BM Hirarchique BM fraction de pixel BM Binaire,

  • AlgorithmesBlock MatchingComparaisonCritres Taux de compression PSNR Temps de calcul: nombre de blocs testsOutils Squences de test Application de testSquences de test1. Squences stationnairesImages fixes dans la plupart (arrire plan) : zone de recherche trs rduite1. Squences lentesImages mouvement des blocs rduits: exploration dune scne1. Squences rapidesImages large mouvement des blocs: squences sportivesClasses selon le vitesse de mouvement des blocsUtilises par les standards de codage vidoApplication de test Codage JPEG fixe, Codage P, Codage B,Lancer Rsultats (Miss America)

    Codec RalisCodec H263Codec MPEG2TC (%) PSNR (dB)TC (%) PSNR (dB)TC (%) PSNR (dB)99.9134.3899.2438.5199.5037.5899.5035.0198.1441.7599.3738.7798.0939.0396.2943.9899.1640.2796.6442.0883.3348.3898.9542.18

  • AlgorithmesRsultats de comparaisonBlock MatchingSquences Rapides 4SSSquences LentesEt statiques GSMthode hybrideMme squence contient des blocs rapides, dautres lents et dautres statiques !

    Graph1

    8.79759.3994.2041065713

    8.79259.3998.3902002461

    8.749.39410.6769165065

    8.7859.3917.6823796939

    8.789.3911.4468864469

    8.7859.4027.6366579948

    8.81759.4039.1082976592

    PSNR

    TC

    Vitesse

    Algorithmes

    Performance

    Squence CarPhone

    Feuil1

    AlgorithmeFS3SS2DLogOSN3SS4SSHBMSGS

    CarphonePSNR35.2935.1935.1734.9635.1435.1235.1435.278.82258.79758.79258.748.7858.788.7858.8175

    Taux94.0593.9993.9993.9493.9193.994.0294.039.4059.3999.3999.3949.3919.399.4029.403

    Blocs Tests1359582535679417878014049019525213104019642116468536.77599557224.20410657138.390200246110.67691650657.682379693911.44688644697.63665799489.1082976592

    Miss AmericaPSNR34.3733.7833.932.7533.1533.0733.934.548.4458.4758.18758.28758.26758.4758.635

    Taux99.699.6499.6699.6899.6299.6399.6699.649.9649.9669.9689.9629.9639.9669.964

    Blocs Tests73904060168177278359925207191047345344878781110811321.78383276693.828488806111.90140873013.29499062586.61597360673.80801994392.7748693037

    NewsPSNR31.6531.6731.6531.6531.6931.731.6531.687.91757.91257.91257.92257.9257.91257.92

    Taux94.5594.5594.5594.5594.5494.5494.5594.549.4559.4559.4559.4549.4549.4559.454

    Blocs Tests135958253751261930801672022076811462052054211505873.99865645147.76880049728.97118455527.222615453510.25956704637.30207719769.9610192115

    FormanPSNR34.5534.5834.6134.5834.5934.6534.634.638.6458.65258.6458.64758.66258.658.6575

    Taux94.0193.9693.9693.9493.9493.8893.9793.969.3969.3969.3949.3949.3889.3979.396

    Blocs Tests135958254006732529862019213016611932392997963086137.487402445411.858363703914.8572956759.944938192215.52481641910.00680462719.7209125993

    TreesPSNR34.0534.0534.0634.0534.0434.0634.0634.068.51258.5158.51258.518.5158.5158.515

    Taux94.294.1794.1794.1894.1494.1494.1994.129.4179.4179.4189.4149.4149.4199.412

    Blocs Tests6889440208079124294109564135300918651458649646610.403956.21475.47826.7654.593257.29324.8233

    RubicPSNR37.2837.2737.2937.1837.1837.2137.2837.339.31759.32259.2959.2959.30259.329.3325

    Taux97.397.3297.3297.3397.3297.3197.3297.329.7329.7329.7339.7329.7319.7329.732

    Blocs Tests68894401745988276254650894065566087742780974.00920972758.457987965512.80878316567.82945216212.57635644997.97793531038.9632124153

    SkiPSNR30.1430.2930.3930.3330.2830.3430.3730.387.57257.59757.58257.577.5857.59257.595

    Taux93.2793.293.1893.293.293.1593.1993.159.329.3189.329.329.3159.3199.315

    Blocs Tests101119203004272084421504102652511534772497652677364.99289344837.19624643799.97274117415.65502109329.77345139666.00564530665.6025338393

    LabPSNR34.9634.9634.9634.9634.9734.9634.9634.968.748.748.748.74258.748.748.74

    Taux94.594.4994.4994.4994.4994.4994.4994.499.4499.4499.4499.4499.4499.4499.449

    Blocs Tests135958253955632087811994602176341634322255181413433.7920634647.18456181367.5203048236.89230543029.17812913026.651353772210.6124816935

    Feuil1

    8.79759.3994.2041065713

    8.79259.3998.3902002461

    8.749.39410.6769165065

    8.7859.3917.6823796939

    8.789.3911.4468864469

    8.7859.4027.6366579948

    8.81759.4039.1082976592

    PSNR

    Taux

    Vitesse

    Algorithmes

    Performance

    CarPhone

    Feuil2

    Feuil3

    Graph2

    8.4459.9641.7838327669

    8.4759.9663.8284888061

    8.18759.96811.9014087301

    8.28759.9623.2949906258

    8.26759.9636.6159736067

    8.4759.9663.8080199439

    8.5559.9614.739635996

    PSNR

    TC

    Vitesse

    Algorithmes

    Performance

    Squence Miss America

    Feuil1

    FS3SS2DLogOSN3SS4SSHBMSGS

    CarphonePSNR35.2935.1935.1734.9635.1435.1235.1435.278.82258.79758.79258.748.7858.788.7858.8175

    Taux94.0593.9993.9993.9493.9193.994.0294.039.4059.3999.3999.3949.3919.399.4029.403

    Blocs Tests1359582535679417878014049019525213104019642116468536.77599557224.20410657138.390200246110.67691650657.682379693911.44688644697.63665799489.1082976592

    Miss AmericaPSNR34.3733.7833.932.7533.1533.0733.934.228.4458.4758.18758.28758.26758.4758.555

    Taux99.799.6499.6699.6899.6299.6399.6699.619.9649.9669.9689.9629.9639.9669.961

    Blocs Tests7390406016817727835992520719104734534487878116329601.78383276693.828488806111.90140873013.29499062586.61597360673.80801994394.739635996

    NewsPSNR31.6831.6731.6531.6531.6931.731.6531.687.91757.91257.91257.92257.9257.91257.92

    Taux94.5594.5594.5594.5594.5494.5494.5594.549.4559.4559.4559.4549.4549.4559.454

    Blocs Tests135958253751261930801672022076811462052054211505873.99865645147.76880049728.97118455527.222615453510.25956704637.30207719769.9610192115

    FormanPSNR34.6134.5834.6134.5834.5934.6534.634.638.6458.65258.6458.64758.66258.658.6575

    Taux94.0193.9693.9693.9493.9493.8893.9793.969.3969.3969.3949.3949.3889.3979.396

    Blocs Tests135958254006732529862019213016611932392997963086137.487402445411.858363703914.8572956759.944938192215.52481641910.00680462719.7209125993

    TreesPSNR34.0634.0534.0634.0534.0434.0634.0634.068.51258.5158.51258.518.5158.5158.515

    Taux94.294.1794.1794.1894.1494.1494.1994.129.4179.4179.4189.4149.4149.4199.412

    Blocs Tests6889440208079124294109564135300918651458649646610.403956.21475.47826.7654.593257.29324.8233

    RubicPSNR37.3437.2737.2937.1837.1837.2137.2837.339.31759.32259.2959.2959.30259.329.3325

    Taux97.3397.3297.3297.3397.3297.3197.3297.329.7329.7329.7339.7329.7319.7329.732

    Blocs Tests68894401745988276254650894065566087742780974.00920972758.457987965512.80878316567.82945216212.57635644997.97793531038.9632124153

    SkiPSNR30.3430.2930.3930.3330.2830.3430.3730.387.57257.59757.58257.577.5857.59257.595

    Taux93.2793.293.1893.293.293.1593.1993.159.329.3189.329.329.3159.3199.315

    Blocs Tests101119203004272084421504102652511534772497652677364.99289344837.19624643799.97274117415.65502109329.77345139666.00564530665.6025338393

    LabPSNR34.9634.9634.9634.9634.9734.9634.9634.968.748.748.748.74258.748.748.74

    Taux94.594.4994.4994.4994.4994.4994.4994.499.4499.4499.4499.4499.4499.4499.449

    Blocs Tests135958253955632087811994602176341634322255181413433.7920634647.18456181367.5203048236.89230543029.17812913026.651353772210.6124816935

    Feuil1

    000

    000

    000

    000

    000

    000

    000

    PSNR

    Taux

    Vitesse

    Algorithmes

    Performance

    Miss America

    Feuil2

    Feuil3

    Graph3

    7.91759.4553.9986564514

    7.91259.4557.7688004972

    7.91259.4558.9711845552

    7.92259.4547.2226154535

    7.9259.45410.2595670463

    7.91259.4557.3020771976

    7.929.4549.9610192115

    PSNR

    TC

    Vitesse

    Algorithmes

    Performance

    Squence News

    Feuil1

    AlgorithmeFS3SS2DLogOSN3SS4SSHBMSGS

    CarphonePSNR35.2935.1935.1734.9635.1435.1235.1435.278.82258.79758.79258.748.7858.788.7858.8175

    Taux94.0593.9993.9993.9493.9193.994.0294.039.4059.3999.3999.3949.3919.399.4029.403

    Blocs Tests1359582535679417878014049019525213104019642116468536.77599557224.20410657138.390200246110.67691650657.682379693911.44688644697.63665799489.1082976592

    Miss AmericaPSNR34.3733.7833.932.7533.1533.0733.934.548.4458.4758.18758.28758.26758.4758.635

    Taux99.699.6499.6699.6899.6299.6399.6699.649.9649.9669.9689.9629.9639.9669.964

    Blocs Tests73904060168177278359925207191047345344878781110811321.78383276693.828488806111.90140873013.29499062586.61597360673.80801994392.7748693037

    NewsPSNR31.6531.6731.6531.6531.6931.731.6531.687.91757.91257.91257.92257.9257.91257.92

    Taux94.5594.5594.5594.5594.5494.5494.5594.549.4559.4559.4559.4549.4549.4559.454

    Blocs Tests135958253751261930801672022076811462052054211505873.99865645147.76880049728.97118455527.222615453510.25956704637.30207719769.9610192115

    FormanPSNR34.5534.5834.6134.5834.5934.6534.634.638.6458.65258.6458.64758.66258.658.6575

    Taux94.0193.9693.9693.9493.9493.8893.9793.969.3969.3969.3949.3949.3889.3979.396

    Blocs Tests135958254006732529862019213016611932392997963086137.487402445411.858363703914.8572956759.944938192215.52481641910.00680462719.7209125993

    TreesPSNR34.0534.0534.0634.0534.0434.0634.0634.068.51258.5158.51258.518.5158.5158.515

    Taux94.294.1794.1794.1894.1494.1494.1994.129.4179.4179.4189.4149.4149.4199.412

    Blocs Tests6889440208079124294109564135300918651458649646610.403956.21475.47826.7654.593257.29324.8233

    RubicPSNR37.2837.2737.2937.1837.1837.2137.2837.339.31759.32259.2959.2959.30259.329.3325

    Taux97.397.3297.3297.3397.3297.3197.3297.329.7329.7329.7339.7329.7319.7329.732

    Blocs Tests68894401745988276254650894065566087742780974.00920972758.457987965512.80878316567.82945216212.57635644997.97793531038.9632124153

    SkiPSNR30.1430.2930.3930.3330.2830.3430.3730.387.57257.59757.58257.577.5857.59257.595

    Taux93.2793.293.1893.293.293.1593.1993.159.329.3189.329.329.3159.3199.315

    Blocs Tests101119203004272084421504102652511534772497652677364.99289344837.19624643799.97274117415.65502109329.77345139666.00564530665.6025338393

    LabPSNR34.9634.9634.9634.9634.9734.9634.9634.968.748.748.748.74258.748.748.74

    Taux94.594.4994.4994.4994.4994.4994.4994.499.4499.4499.4499.4499.4499.4499.449

    Blocs Tests135958253955632087811994602176341634322255181413433.7920634647.18456181367.5203048236.89230543029.17812913026.651353772210.6124816935

    Feuil1

    000

    000

    000

    000

    000

    000

    000

    PSNR

    Taux

    Vitesse

    Algorithmes

    Performance

    Squence News

    Feuil2

    Feuil3

    Graph6

    9.31759.7324.0092097275

    9.32259.7328.4579879655

    9.2959.73312.8087831656

    9.2959.7327.829452162

    9.30259.73112.5763564499

    9.329.7327.9779353103

    9.33259.7328.9632124153

    PSNR

    TC

    Vitesse

    Algorithmes

    Performance

    Squence Rubic

    Feuil1

    AlgorithmeFS3SS2DLogOSN3SS4SSHBMSGS

    CarphonePSNR35.2935.1935.1734.9635.1435.1235.1435.278.82258.79758.79258.748.7858.788.7858.8175

    Taux94.0593.9993.9993.9493.9193.994.0294.039.4059.3999.3999.3949.3919.399.4029.403

    Blocs Tests1359582535679417878014049019525213104019642116468536.77599557224.20410657138.390200246110.67691650657.682379693911.44688644697.63665799489.1082976592

    Miss AmericaPSNR34.3733.7833.932.7533.1533.0733.934.548.4458.4758.18758.28758.26758.4758.635

    Taux99.699.6499.6699.6899.6299.6399.6699.649.9649.9669.9689.9629.9639.9669.964

    Blocs Tests73904060168177278359925207191047345344878781110811321.78383276693.828488806111.90140873013.29499062586.61597360673.80801994392.7748693037

    NewsPSNR31.6531.6731.6531.6531.6931.731.6531.687.91757.91257.91257.92257.9257.91257.92

    Taux94.5594.5594.5594.5594.5494.5494.5594.549.4559.4559.4559.4549.4549.4559.454

    Blocs Tests135958253751261930801672022076811462052054211505873.99865645147.76880049728.97118455527.222615453510.25956704637.30207719769.9610192115

    FormanPSNR34.5534.5834.6134.5834.5934.6534.634.638.6458.65258.6458.64758.66258.658.6575

    Taux94.0193.9693.9693.9493.9493.8893.9793.969.3969.3969.3949.3949.3889.3979.396

    Blocs Tests135958254006732529862019213016611932392997963086134.99160163037.90557580269.90486378346.629958794810.34987761276.67120308486.4806083995

    TreesPSNR34.0534.0534.0634.0534.0434.0634.0634.068.51258.5158.51258.518.5158.5158.515

    Taux94.294.1794.1794.1894.1494.1494.1994.129.4179.4179.4189.4149.4149.4199.412

    Blocs Tests6889440208079124294109564135300918651458649646610.403956.21475.47826.7654.593257.29324.8233

    RubicPSNR37.2837.2737.2937.1837.1837.2137.2837.339.31759.32259.2959.2959.30259.329.3325

    Taux97.397.3297.3297.3397.3297.3197.3297.329.7329.7329.7339.7329.7319.7329.732

    Blocs Tests68894401745988276254650894065566087742780974.00920972758.457987965512.80878316567.82945216212.57635644997.97793531038.9632124153

    SkiPSNR30.1430.2930.3930.3330.2830.3430.3730.387.57257.59757.58257.577.5857.59257.595

    Taux93.2793.293.1893.293.293.1593.1993.159.329.3189.329.329.3159.3199.315

    Blocs Tests101119203004272084421504102652511534772497652677364.99289344837.19624643799.97274117415.65502109329.77345139666.00564530665.6025338393

    LabPSNR34.9634.9634.9634.9634.9734.9634.9634.968.748.748.748.74258.748.748.74

    Taux94.594.4994.4994.4994.4994.4994.4994.499.4499.4499.4499.4499.4499.4499.449

    Blocs Tests135958253955632087811994602176341634322255181413433.7920634647.18456181367.5203048236.89230543029.17812913026.651353772210.6124816935

    Feuil1

    000

    000

    000

    000

    000

    000

    000

    PSNR

    Taux

    Vitesse

    Algorithmes

    Performance

    Squence Rubic

    Feuil2

    Feuil3

  • Dtecter le type de mouvement du bloc Choisir lalgorithme de recherche qui convient Choisir la zone de recherche optimale Choisir le point de dpart appropriMthode hybride

  • Calculer un vecteur de mouvement moyen partir des estimations prcdentes Dans la mme imageMmes vecteurs de mouvement ou vecteurs trs proches Bloc courant Blocs prcdentsMthode hybrideDtection du type du bloc Dans les images successivesMmes vecteurs de mouvement ou vecteurs trs proches En Prenant en compte la distance et la direction entre limage code et limage rfrence !!!

  • Vecteur de mouvement moyen estim VM(0,0)Bloc Stationnaire](0,0),(3,3)]Bloc Lent> (3,3)Bloc rapideCalculer BDMRetourner VMCalculer BDMCalculer VMUtiliser GSZDR= 3X3Commencer de VMRetourner VMCalculer VMUtiliser 4SSZDR= 16X16Commencer de VMRetourner VMImages PImages B

  • RsultatsMthode hybride Adapte tous type de squences voire blocs Amlioration jusqu 81% en temps de calcul (plus -> recherche par interruption)AvantagesInconvnientsParamtres importantsNombre dimages BParamtres MinBDM & MaxBDM Ajout des instructions de dcision (ngligeables) Perte en TC et PSNR (non significative)

  • Conclusion Etudier les principes, les normes et les techniques de codage vido Etudier les mthodes dEM. Raliser un Codec vido bas sur lEM. Etudier et comparer les algos de BM . Proposer une mthode hybride . Raffiner le codage Intra Audio + Synchro Paralllisation Implmentation HARD Tech. Apprentissage Avances& Perspectives

  • Merci

    Magister Informatique - A.DjeffalDynamically para.. Concentrating on Me and DCTMagister Informatique - A.DjeffalIntroduction to our work.Variation in video content and its processingDynamic parameterization ME & DCTOn-going work and conclusionsMagister Informatique - A.DjeffalThis work leverages on the 2 fundamental properties:Video content is highly non-uniform in space and timeVideo processing can gracefully degrade in power constrained environments by exploiting perceptual tolerance.Mpeg-4 used due to its object based coding techniques.A lot of recent work focuses on low power devices and circuits. However it is accepted that the largest gains in power efficiency comes from high level algorithm changes and this is what our work focuses on.

    Magister Informatique - A.DjeffalAll for low power video processing. Magister Informatique - A.DjeffalMagister Informatique - A.DjeffalOur architecture is on the lines of a Partially Predefined Configuration Architecture i.e. a reconfigurable implementation in custom silicon. aSOC shown below : a number of tiles with different cores, local SRAM and a statically scheduled interconnection switch. Tiles can be reconfigured internally as well as from an external source.Magister Informatique - A.DjeffalMagister Informatique - A.DjeffalShows horz. Component of the MVs for the 2 sequences. The MV distribution varies considerably but is also seen to be correlated in space and time. The variability is due to the changing characteristics of the video signal. The table tennis seq. shows that even in the same sequence, the MV distribution differs due to the changing picture content.Magister Informatique - A.DjeffalMagister Informatique - A.DjeffalHere we see the non-uniform content variation in space, where the background shows very low content variation, while the ball and the paddle are the in the high content variation region.Magister Informatique - A.DjeffalMagister Informatique - A.DjeffalA Video processing system is shown with I/p signals, the output being measured in terms of prec., qual, .. With constraints such as area, power and latency. A predictor block takes inputs as 1, 2, 3 and gives an output in terms of archi. and functional parameters to the VPB. The VPB then processes the input according to the set parameters.Magister Informatique - A.DjeffalFunctional parameter adjustment. The 2 slides above are the same frames but encoded using different algorithms: Full Search and Logarithmic. There will be no quality degradation. The difference is in the latency values to determine the MVs, the compression ration (or transmission bit-rate) and power consumption. Full Search also involves higher computations mapping to higher power consumption. Magister Informatique - A.DjeffalLike mentioned before a larger search space increases power consumption as shown in the graph. Power consumption from computations and I/O (memory accesses). Different algorithms reduce memory accesses and computations. Magister Informatique - A.DjeffalMagister Informatique - A.DjeffalThe parameters in ME are the algorithms used for the determination of the MV, Pel sub-sampling and Half-Pel ME.Magister Informatique - A.DjeffalSearch Space is the area in which the candidate blocks for the comparison are present. Larger search space improves the chances to get a good match, which improves compression. The plot shows how, for a specific sequence, increase in search space reduces the bits per pixel value. This increase is effective only up to a point. Also, a larger search space increases computations, thus increasing power consumption. Magister Informatique - A.DjeffalMagister Informatique - A.DjeffalAs an audience of video coding experts you all must be aware about Motion Estimation We look at the different algorithms, starting with Full search, where a all the blocks in a search window are compared against to form a MV.Magister Informatique - A.DjeffalNext, we look at spiral search which is data dependent during run time. The search starts at a predictor location and continues around it in a spiral fashion. The search stops whenever a value for the matching criteria is obtained which is below the threshold set by the user. Reduced calculations but lower compression.Magister Informatique - A.DjeffalCompare against sampled blocks around a predictor location. The block fulfilling the matching criteria best is chosen as the next predictor location and the step size is reduced for block comparison around it. The previous step is repeated and this continues till the step size is 1 or a boundary is reached, thus converging on to the MV location. Once again, lower computations means lower power but also lower compression.Magister Informatique - A.DjeffalInstead of comparing each and every pixel we compare alternate pixels in 2:1 sub sampling or every 4th pixel in 4:1 sub sampling. Reducing computations.Magister Informatique - A.DjeffalBasically comparing at sub pixel level. Small a,b,c represent the half pels given by the formula.