Intégration de connaissance experte

Transcription

1 Intégration de connaissance experte dans des systèmes de fusion d informationsd Florentin BUJOR, Gabriel VASILE, Lionel VALET Emmanuel TROUVÉ, Gilles MAURIS et Philippe BOLON emmanuel.trouve@univ-savoie.fr LISTIC / ESIA Université de Savoie p. 1

2 Travaux du LISTIC Thème central : méthodes/systèmes de fusion d informations Systèmes coopératifs : informations a priori, connaissance experte, Approche basées sur la logique floue; Extraction d informations : Calcul d attributs 2D, 3D, 2D+temps, Images multi-composantes (multi-spectrale, polarisation, interférométrie) Application en télédétection : Images multi-dates (détection de changement), Données PolInSAR : estimation de cohérence, classification Wishart, ORFEO : Optique + Radar + Haute Résolution. p. 2

3 Différents niveaux de fusion Scène «Decision level» : objets, cibles Validation/rejet, tri, simplification Attributs / primitives «Feature level» : attributs, paramètres Classification, détections Pixels «Data level» : albédo, radiométrie filtrage, amélioration de résolution p. 3

4 Fusion «Data level» POLSAR data Attributes Info. Fusion Clasif. Land cover map Info. a priori p. 4

5 POL-InSAR imaging Three different polarization configurations k {HH,VV, XX} ; Estimation of matrix C pol : C pol [ T ] 11 = * [ Ω ] 12 T [ Ω ] 12 [ T ] 22 Coherency matrices: * T [ Tii ] = kiki Polarimetric interferometric covariance: T [ Ω k k * 12 ] = 1 2 HR airborne L-band POL-InSAR images of the Oberpfaffenhofen area, Wessling, Germany; colored composition of diag{t 11 }. p. 5

6 IDAN estimation Principle: In each pixel (seed), a window of variable shape and dimensions is built, containing only connex pixels belonging to the same statistic population as the seed. AN spatial support estimation T 11 T 22 + Ω 12 Goal: reach the number of pixels necessary for reliable complex-correlation estimation; preserve stationarity within estimation window; estimate statistical measures over the largest possible neighborhood. p. 6

7 IDAN estimation original boxcar directional IDAN T 11 intensity image filtering results [ pixels]: p. 7

8 H/α/A /A decomposition EIGENVECTOR DECOMPOSITION OF COHERENCY P i = 3 real eigenvalues λ 1 >λ 2 >λ 3 of T ii k = λ 1 i λ k ENTROPY 3 H = Pi log 3( Pi ) i= 1 α PARAMETER α = P + 1α 1 + P2α 2 P3 α 3 ANISOTROPY A = λ2 λ3 λ + λ 2 3 p. 8

9 H/α/A /A decomposition boxcar directional IDAN H/α/A of T 11 (color compositions) p. 9

10 POLSAR classification results IDAN optical image T 11 Wishart classification results p. 10

11 Fusion «Feature level» SAR data Membership functions Attributes Info. Extraction Info. Fusion Info. a priori Classification expert knowledge p. 11

12 Interactive information fusion Information Extraction A1 A2 A3 Symbolic Fuzzy System Output Experts (Knowledge- Learning) Membership Functions & Rules Information Fusion Quality Assessment p. 12

13 Multi-temporal temporal SAR Data Spatial information x t y Multi-temporal info. Simultaneous temporal & spatial information extraction p. 13

14 Information extraction Spatial edge attribute 1 image ρ n zn = 1- ρ = n µ min µ e iθn µ a b, b µ a ρ N images = arg N N zn zn n= 1 n= 1 = and θ N 1 p. 14

15 Information extraction Temporal change attribute 2 images ν( i,j) N images = µ min µ i j µ, µ j i ν temp = 1- min i j ( ν( i,j) ) p. 15

16 Information extraction 3D-Texture attribute Second order log-cumulants ˆ~ ( ) 2 k2 = logi logi ν ν Spatial-temporal temporal neighborhood 2 p. 16

17 Interactive information fusion Symbolic description of the attributes small large very_large Fuzzy rules IF {A 1 is very_large} AND {A 2 is large} THEN {the output belongs to the class temporal_change}. Aggregation by Zadeh composition rule p. 17

18 Multi-temp temp fusion results Ground truth Fusion result p. 18

19 Feature/Decision level information fusion Information Extraction / Decision D1 D2 D3 Choquet Integral System Output Importance and interaction coefficients Experts (Knowledge- Learning) Reference areas Information Fusion Quality Assessment p. 19

20 Default detection in 3D tomographic data p. 20

21 Pixel Features p. 21

22 2-additive Choquet integral aggregation Cu ( x,..., 1, x2 xn )= I ij φ 0 min( x i, x j ) I ij + I ij π0 max( x i, x j ) I ij + n i= 1 x ( v i i 1 2 j i I ij ) are the Shapley indices, representing the overall importance of each criterion v i n c i ( υ i = 1) i= 1 I ij represents the interactions between pairs of the criteria ( [ 1,1 ]) positive value implies a positive synergy (complementarity) among criteria negative value implies a negative synergy (redundancy) between both criteria null value implies that the criteria are independent I ij ci,cj p. 22

23 Features/Decisions Decision Présence + intensité du défaut Qualité globale de la pièce Action sur le process p. 23

24 Conclusions Fusion Tools available at the different levels Merging Optical, Radar and Prior Information Integrating Expert Knowledge Interactivity ORFEO HR Data analysis Pixel Feature/Object Object Decision Application driven tools p. 24