Comparaison d`estimateurs de tangente sur les courbes
Transcription
Comparaison d`estimateurs de tangente sur les courbes
Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Comparaison d’estimateurs de tangente sur les courbes discrètes + Modèles déformable discrets + Onco-SP-IM : Déconvolution François de Vieilleville Irit, Université Paul Sabatier Réunion d’équipe TCI F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Outline 1 Tangent Estimators + Evaluation + Enhancement Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement 2 Digital Deformable Model Introduction and Problem Statement Proposed Solution Experimental Evaluation 3 Onco-SP-IM : Image Restoration F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Caractéristiques Géométriques en Géométrie Discrète Question Quel estimateur/algorithme choisir pour un descripteur géométrique donné ? Selon quels critères objectifs ? Comment les définir ? F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Caractéristiques Géométriques en Géométrie Discrète 111111111111111111111 000000000000000000000 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 000000000000000000000 111111111111111111111 1111111111111111111111 0000000000000000000000 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 0000000000000000000000 1111111111111111111111 Aire discrète Aire Global Global Question Quel estimateur/algorithme choisir pour un descripteur géométrique donné ? Selon quels critères objectifs ? Comment les définir ? F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Caractéristiques Géométriques en Géométrie Discrète Normales Local Normales discrètes Local Question Quel estimateur/algorithme choisir pour un descripteur géométrique donné ? Selon quels critères objectifs ? Comment les définir ? F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Digital Objects - Digital Curves Digital plane ≡ Z2 . Digital object ≡ Finite set of points of Z2 . From euclidean shapes to digital shapes : E = X ∩ hZ × hZ. The digital boundary can be extracted using inter-pixel contours or cellular decomposition. Elements of the digital curve are ordered and indexed. → Computation of local geometric characteristics use windows containing finite number of points. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Digital Objects - Digital Curves h = 0.02 Circle Ellipse F. de Vieilleville Réunion Equipe TCI Flower Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Problems When Analyzing Digital Curves (1) There exists infinitely many continuous Euclidean shapes for a digitized shape. (2) How to determine the size of computation window w.r.t. local geometry ? (3) Time to spend on computations may be limited. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Solution 1 : Continuous Estimators and Fixed Sized Window (1) Add hypotheses for the reference shape : smoothness, compactness, convexity, minimal perimeter or maximal area. → Brings bias in estimation : special underlying curve. (2 & 3 ) Fix the size of the computation window. → explicit trade-off between time computation and precision, loss of local geometry consistency. → For tangent estimation we chose median filtering, Gaussian derivative and low order polynomial fitting. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Solution 2 : Use Maximal Digital Straight Segment Based Algorithms (1) No hypotheses on the reference shape : study of the digital linear part of the digital object. (2 & 3 ) Size of the computation window computed automatically. → At very low resolution, consistency with local geometry ? → For tangent estimation we choose the λ-MST tangent estimator and the Global Minimum Curvature estimator. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Polynomial Fitting (Ci = (xi , yi ))1≤i≤2q+1 set of 2q + 1 consecutive samples on digital curve. Minimize the functional : 2 P2q+1 Pj=N E (a0 , . . . , aN ) = i=1 yi − j=0 aj xij . ELR (a, b) = E (a, b, 0, . . .) Linear Regression, EIPF (a, b) = E (0, a, b, 0, . . .) Implicit Parabola Fitting, EEPF (a, b, c) = E (a, b, c, 0, . . .) Explicit Parabola Fitting. Estimation on whole curve in O ((2q + 1) ∗ number of points) F. de Vieilleville Réunion Equipe TCI Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Polynomial Fitting 1 Th EPF q=12 3.25 IPF q=12 RL q=12 3.2 3.15 3.1 3.05 polar angle 1.5 1.53 1.56 1.59 1.62 1.65 1.68 F. de Vieilleville AAE of tangent orientation tangent orientation 3.3 IPF q=8 EPF q=8 LR q=8 IPF q=16 EPF q=16 LR q=16 IPF q=32 EPF q=32 LR q=32 0.1 0.01 0.001 10 Réunion Equipe TCI 100 inv. of grid step 1000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Polynomial Fitting Done separately on each coordinate (X and Y). → Parametrisation problem induced : in fact a length estimation is required. Iterative refining of the length estimation : tangent → tangent → length .... → In practice, no real improvement even on constant curvature shapes (disk) F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Polynomial Fitting abs. error of tan. orientation AAE of tan. orientation 0.05 0.045 ICIPF q=8 ITICIPF q=8 0.04 0.035 0.03 0.025 0.02 ICIPF q = 8 ITSIPF q = 8 ICIPF q = 16 ITICIPF q = 16 ICIPF q = 32 ITICIPF q = 32 1 0.1 0.015 0.01 0.005 0.01 polar angle inv.of grid step 0 1.2 1.25 1.3 1.35 1.4 1.45 1.5 F. de Vieilleville 10 Réunion Equipe TCI 100 1000 10000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Gaussian Smoothing Gσ0 (t), first derivative of Gσ (t) = σq = √1 σ 2π Pq Gσ0 q (−i)Ci 2 , exp −t 2 2σ estimated derivative at C0 obtained as : 2q+1 3 Estimation on whole curve in O ((2q + 1) ∗ number of points). F. de Vieilleville Réunion Equipe TCI i=−q Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Median Filtering Choosing the median orientation among the following 2q vectors centered on : Ci : ((Ci−q Ci , , ) . . . , (Ci−1 Ci , , ) (Ci Ci+1 , , ) . . . , (Ci Ci+q , )) F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Digital Straight Segment Digital Straight Segment : piece of the digitization of a ray D(a, b, µ, ω) = {M ∈ Z2 |µ ≤ aMx − bMy < µ + ω} (a, b, µ, ω) ∈ Z4 , ba : slope of the line, ω : thickness ω = |a| + |b| 4-connected line. → Exists various classes of DSS, one of interest in the class of the Maximal Segments. → Recognition in linear time, incremental algorithms. → Strong link with continued fraction. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Maximal Segment → Recognition of the whole set of MS in linear time on the whole digital curve. → Digital convexity ≡ monotony of slopes of MS[Doerksen et.al.’04]. → Average length of maximal segments follows Θ(h−1/3 ), average number of maximal segments follows Θ(h−2/3 ), on digitization of shapes with C 3 border. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Digital Estimator : λ-MST Weighted average of orientation of MS, weights depend on point location and λ function. Exists simple λ function making no false concavities (triangle). Point to point multigrid convergence proven (th. Θ(h1/3 ), xp. Θ(h2/3 )). F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Global Minimum Curvature Choose a shape that minimizes its squared curvature along its border among the shapes whose digitization is almost the same as the input digital shape. Global optimisation scheme, in the space of tangent directions parametrized by the curvilinear abscissa. Each MS on the digital curve defines bound in the research space. Iterative relaxation scheme. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement 3.3 tangent orientation tangent orientation Defects of Continuous Methods :Digitized Circle (h=0.01) Th. Tangent GD q=3 GD q=7 GD q=11 3.25 3.2 3.15 3.1 3.05 3 3.2 3.15 3.1 3 polar angle polar angle 2.95 1.4 1.45 1.5 1.55 1.6 1.65 1.7 tangent orientation tangent orientation Th. Tangent ICIPF q=3 ICIPF q=7 ICIPF q=11 3.05 2.95 3.3 3.3 3.25 Th. Tangent Matas q=3 Matas q=7 Matas q=11 3.25 3.2 3.15 3.1 3.05 3.3 1.4 1.45 1.5 1.55 1.6 1.65 1.7 Theoretical Tangent GMC L-MST 3.25 3.2 3.15 3.1 3.05 3 3 polar angle 2.95 polar angle 2.95 1.4 1.45 1.5 1.55 1.6 1.65 1.7 F. de Vieilleville 1.4 1.45 1.5 1.55 Réunion Equipe TCI 1.6 1.65 1.7 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Angle deviation Angle deviation Isotropy Behaviour :Digitized Circle (h=0.01) 0.1 0.01 0.001 0.001 0.1 0.01 0.001 0 moy MATAS q=3 moy MATAS q=9 max MATAS q=3 max MATAS q=9 0.8 polar angle 0.01 1.6 0 Angle deviation Angle deviation 0 moy GD q=3 moy GD q=7 max GD q=3 max GD q=7 polar angle 0.1 moy ICIPF q=3 moy ICIPF q=9 max ICIPF q=3 max ICIPF q=9 0.8 polar angle moy L-MST max L-MST moy GMC max GMC 0.1 0.01 0.001 1.6 F. de Vieilleville 1.6 0 0.8 Réunion Equipe TCI polar angle 1.6 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Perimeter estimation :Digitized Circle(r=50, 100 random center shifts) θTAN (Ci ) is the estimated normal at point Ci , the estimated length, denoted LθTAN (Ci ), equals : | cos(θTAN (Ci ))| if the associated linel is horizontal, | sin(θTAN (Ci ))| otherwise. Win.size q=1 q=2 q=4 q=8 q=16 q=32 q=64 GD 30.6724 26.9072 6.46475 0.79856 0.12082 0.05776 0.13503 ICIPF 30.6724 2.35272 0.40573 0.11547 0.06791 0.08221 0.24110 MATAS 27.14933 11.32574 3.622088 1.495153 1.336123 1.336123 1.336123 F. de Vieilleville ITICIPF 29.67998 4.032775 0.897381 0.181935 0.070328 0.145093 0.462371 λ-MST GMC 0.50378 0.02553 Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Single Criterion : Multi-Grid Convergence Euclidean Object (X ) Geometric Descriptor (G) Evaluation −→ G (X ) −→ Ĝ (DigGh (X )) ↓ DigGh (·) ↓ , ,· · · Digital Object (DigGh (X )) Estimation of Geometric Descriptor (Ĝ) F. de Vieilleville Réunion Equipe TCI Estimation Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Single Criterion : Multi-Grid Convergence Euclidean Object (X ) Geometric Descriptor (G) −→ ↓ DigGh (·) ↓ G (X ) ↑ limh→0 ? ↑ −→ , Evaluation Ĝ (DigGh (X )) ,· · · Digital Object (DigGh (X )) Estimation of Geometric Descriptor (Ĝ) F. de Vieilleville Réunion Equipe TCI Estimation Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Multi-Grid Convergence On Digitized Disks : Maximal Absolute Error 1 ME of tangent orientation ME of tangent orientation 1 0.1 0.01 0.001 GD q=1 GD q=4 GD q=16 GD q=64 GD q=128 10 inv. of grid step 100 1000 0.001 ICIPF q=1 ICIPF q=4 ICIPF q=16 ICIPF q=64 ICIPF q=128 10 inv. of grid step 100 1000 100 1000 10000 1 ME of tangent orientation ME of tangent orientation 0.01 10000 1 0.1 0.01 0.001 0.1 MATAS q=1 MATAS q=4 MATAS q=16 MATAS q=64 MATAS q=128 10 inv. of grid step 100 1000 0.1 0.01 0.001 10000 F. de Vieilleville L-MST GMC x^(-1.1/2) x^(-1) 10 inv. of grid step Réunion Equipe TCI 10000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement AAE of tangent orientation AAE of tangent orientation Multi-Grid Convergence On Digitized Ellipses : Average Absolute Error 0.1 0.01 0.001 GD q=1 GD q=4 GD q=16 GD q=64 GD q=128 ICIPF q=1 ICIPF q=4 ICIPF q=16 ICIPF q=64 ICIPF q=128 inv. of grid step inv. of grid step 0.001 100 1000 10000 0.1 0.01 0.001 0.01 10 AAE of tangent orientation AAE of tangent orientation 10 0.1 100 1000 100 1000 10000 0.1 0.01 MATAS q=1 MATAS q=4 MATAS q=16 MATAS q=64 MATAS q=128 10 inv. of grid step 100 1000 10000 F. de Vieilleville 0.001 L-MST GMC x^(-2/3) x^(-1.3/3) 10 inv. of grid step Réunion Equipe TCI 10000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Multi-Grid Convergence On Digitized Flowers : Average Absolute Error 1 AAE of tangent orientation AAE of tangent orientation 1 0.1 0.01 0.001 0.01 GD q=1 GD q=4 GD q=16 GD q=64 GD q=128 10 inv. of grid step 100 1000 0.001 10000 inv. of grid step 100 1000 100 1000 10000 1 AAE of tangent orientation AAE of tangent orientation ICIPF q=1 ICIPF q=4 ICIPF q=16 ICIPF q=64 ICIPF q=128 10 1 0.1 0.01 0.001 0.1 0.1 0.01 MATAS q=1 MATAS q=4 MATAS q=16 MATAS q=64 MATAS q=128 10 inv. of grid step 100 1000 10000 F. de Vieilleville 0.001 L-MST GMC x^(-2.5/3) x^(-1/2) 10 inv. of grid step Réunion Equipe TCI 10000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Mixed Criterion : Average Abs. Error * Time Precision is not the only criterion to be taken into account. Product of AAE by the time required for the computation. → Penalises huge errors on average even if small time. → Penalises multi-grid convergence estimators that require too many computations. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Asymptotic Behaviour : Average Absolute Error By Time on Circle AAEBT AAEBT 100 100 10 10 x^(1/3) L-MST GD q=8 GD q=16 GD q=32 GD q=64 GD q=128 1 0.1 10 100 1000 x^(1/3) L-MST ICIPF q=8 ICIPF q=16 ICIPF q=32 ICIPF q=64 ICIPF q=128 1 0.1 10 inv. of grid step AAEBT AAEBT 100 100 10 10 x^(1/3) L-MST MATAS q=8 MATAS q=16 MATAS q=32 MATAS q=64 MATAS q=128 1 0.1 10 100 1000 inv.of grid step F. de Vieilleville 100 1000 inv.of grid step 1 0.24 x^(1/3) 17 x^(0.34131) L-MST GMC 0.1 10 100 Réunion Equipe TCI 1000 inv. of grid step Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Size of Computation Window Required to Reach Best Accuracy for Average Size of Computation Window AAE of tangent orientation Absolute Error on Circle 0.01 GD q=12 GD q=16 GD q=24 0.001 GD q=32 GD q=48 GD q = 64 GD q=96 GD q=128 x^(2.5/3) 0.0001 10 100 1000 inv.of grid step F. de Vieilleville x^(1/2) GD 100 inv.of grid step 10 10 100 Réunion Equipe TCI 1000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Computation Window Size : Based on MS ? Should be in O((1/h)1/2 ). However, no know digital primitives with such growing speed... q1−0 q1−1 q1−2 q2 q3 q4 q5 max(||B(Cj )) − Cj ||1 , ||F (Cj )) − Cj ||1 ), min(||B(Cj )) − Cj ||1 , ||F (Cj )) − Cj ||1 ), )/2, (q1−0 + q1−1 1/6 qP , 1−0 (1/h) b( ||MS || )/nb(MS i 1 i )c , i j P k 3/2 = (( i ||MSi ||1 )/nb(MSi )) , = b(||F (Cj )) − B(F (Cj ))||1 − 1)/2c . = = = = = F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Adaptivity of the proposed computation windows Window Adaptivity q1−0 Local q1−1 Local q1−2 Local q2 Local q3 Global q4 Global q5 Local F. de Vieilleville Average Expected Size 1 13 O (h) 1 O ( h1 ) 3 1 O ( h1 ) 3 1 O ( h1 ) 2 1 O ( h1 ) 3 1 O ( h1 ) 2 1 O ( h1 ) 3 Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Improvements of The Gaussian Derivative Estimator (a) 3.3 3.2 Th. Tangent H2 GD H3 GD H4 GD H5 GD 3.25 tangent orientation 3.25 tangent orientation (b) 3.3 Th. Tangent H1-0 GD H1-1 GD H1-2 GD 3.15 3.1 3.05 3 3.2 3.15 3.1 3.05 3 polar angle 2.95 polar angle 2.95 1.4 1.45 1.5 1.55 1.6 1.65 1.7 1.4 moy H1-0 GD moy H1-1 GD max H1-0 GD max H1-1 GD 0.1 0.01 0.001 0 0.8 1.5 1.55 1.6 1.65 1.7 (d) Angle deviation Angle deviation (c) 1.45 polar angle moy H2 GD moy H3 GD max H2 GD max H3 GD 0.1 0.01 0.001 1.6 F. de Vieilleville 0 0.8 Réunion Equipe TCI polar angle 1.6 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Improvements of The Gaussian Derivative Estimator (f) (e) ME of tangent orientation Angle deviation 1 0.1 0.01 moy H4 GD moy H5 GD max H4 GD max H5 GD 0.001 0 0.8 polar angle 1.6 0.1 H1-0 GD H1-1 GD H1-2 GD 0.01 H2 GD H3 GD H4 GD H5 GD x^(2.5/3) x^(1/2) 0.001 10 inv. of grid step 100 1 H1-0 GD H1-1 GD H1-2 GD H2 GD H3 GD H4 GD H5 GD x^(-2.7/3) x^(-1/2) 0.1 10000 (h) AAE of tangent orientation AAE of tangent orientation (g) 1000 0.01 0.1 H1-0 GD H1-1 GD H1-2 GD H2 GD H3 GD H4 GD H5 GD x^(-2.7/3) x^(-1.23/2) 0.001 10 0.01 inv. of grid step 0.001 10 100 1000 10000 F. de Vieilleville inv. of grid step 100 Réunion Equipe TCI 1000 10000 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Digital Objects - Digital Curves Classic estimators + Digital Estimators Evaluation Enhancement Improvements of The Gaussian Derivative Estimator Experimental results confirm that “classic” estimators can benefit from digital primitives. Theoretical results show that q5 is to be multi-grid convergent. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Active Contours [Kass’88] Let Ω be a parametrization of the curve embodying the active contour geometry : Ω = [0, 1] → R2 x(s) v (s) = y (s) It energy being defined as : Z 1 0 00 EDM (v ) = Eint (v (s), v (s), v (s)) + Edata (v (s)) ds 0 v = argminv ∈F EDM (v ) F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Active Contours [Kass’88] Eint (·) monitored by : α elasticity, β rigidity. 2 Eint (v (s), v 0 (s), v 00 (s)) = α(s) v 0 (s) + | {z } lenght penalisation 2 β(s) v 00 (s) | {z } /2 curvature penalisation Image term is decreasing with the norm of the image gradient. Direct computation not tractable, problem rewritten as a variational problem, solved with finite differences, sometimes with greedy heuristics on energy. → hard to find global minimum, no topological changes. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation And so on... Active Contours [Kass’88] −(αC 0 )0 + (βC 00 )00 + ∇P(C) = 0 Level Sets [Osher, Sethian’89] → handle topological changes but does not necessarily stops. 2 ∂C(s,t) ∂ C = P(C) ∂s 2 + w~n ∂t Geodesic Active Contours [Caselles ’95] → handle topological changes and stops on strong gradient 2 ∂C(s,t) = P(C) ∂∂sC2 − < ∇P, ~n > ~n ∂t Minimal Pathes [Cohen, Kimmel ’96] → ReachR a global minimum for curve from A to B. E (C) = Ω (w + P(C(s))) ds F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation ...on Open Digital 4-Connected Path ? Γ ∈ C(A, B) : Set of simple oriented open 4-connected path from point A to point B. Given boundary conditions, finite number of possibilities ! Classical mathematics resolution techniques (DPE stuff) not available... → Greedy heuristics based on energy. → Problem is how to define energies with good properties. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Requirements for Internal Energy D Eint (Γ) = αL̂(Γ) Internal energy relying only on length estimation and requires : convex functional, low variability of global minimum, has to be a good length estimator. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy : Length Estimation P Using DSS-Based Methods : L̂(Γ) = s∈surfel(Γ) < s, τ̂ (s) > with θ(τ̂ (s)) equals to atan(slope) of rightmost max dss of surfel s. n + 2 points Γ 2 +2 L̂(Γ) = √(n−1) (n−1)2 +1 Γ’ 2 +1 L̂(Γ0 ) = √(n−2) (n−2)2 +1 For n = 10, L̂(Γ) ≈ 9, 15 and L̂(Γ0 ) ≈ 9, 47. → Does not bring a convex functional. F. de Vieilleville Réunion Equipe TCI + √ 2 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy : Length Estimation Naive length estimation using elementary surfel length : B A → Classic Euclidean distance but with horizontal and vertical steps, → brings a global minimum, → but too many of them ! ! F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy : Length Estimation D Eint (Γ) = αL2 (CMLP(Γ)) → Known as a good estimator, multi-grid convergent estimator [Klette’99], → convex functional, → global minimum is very close to the digitization of the euclidean segment joining A et B. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Data Energy Regarding the image energy term we define it as : X D EData (Γ) = max(||∇I ||) − ||∇I (c)||, c∈Γ in order for the energy to be positive everywhere and decreasing with the norm of the image gradient. Our digital combinatorial minimization problem becomes : Γ∗ = arg min Γ∈C(A,B) D D D D EDM (Γ), where EDM (Γ) = EInt (Γ) + EData (Γ). F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Elementary Deformations Outside corners, − + − − − 0 − 0 − 0 + + 0 + 0 Flip + 0 − + into 0 − + + . 0 + 0 + Inside corners, − + − − − 0 − 0 + + 0 − 0 0 + Flip − − − + into 0 + 0 + . Inside or outside bumps, − − Flat to (). Flat outside parts, + + − + + + 0 0 0 Bump + outside to Flat inside parts, − − − − − 0 0 0 + + + − − + − − 0 0 0 Bump − to F. de Vieilleville 0 0 0 + + Réunion Equipe TCI + . − + − + . Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Examples − 0 + − 0 + 0 0 −0 − − + F. de Vieilleville Réunion Equipe TCI + Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Constrain Topology Maintain a simple 4-connected path from A to B : check if deformation is valid, does not collide with the contour, filtered when listing the possible deformations of Γ. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation CMLP and Path Convex vertices correspond to outside corners or outside bumps, convex deformation are Flip + or Flat. Concave vertices correspond to the inside corners or inside bumps, concave deformation are Flip − or Flat. − − + F. de Vieilleville + Réunion Equipe TCI + Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy Descent Theorem Let Γ ∈ C(A, B) being finite, choosing at each step any one of the valid convex or concave deformation leads to CMLP(Γ) = [AB] after a finite number of steps. [Requires] Proposition Let Γ ∈ C(A, B) being finite , if CMLP(Γ) 6= [AB] there exists at least one valid convex deformation on a convex vertex or one valid concave deformation on a concave vertex of CMLP(Γ). F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy Descent Proposition Let Γ ∈ C(A, B) being finite , if CMLP(Γ) 6= [AB] there exists at least one valid convex deformation on a convex vertex or one valid concave deformation on a concave vertex of CMLP(Γ). → Flat ok, remains Flip if Γ made of 2 Freeman code, Flip ok else contour would be infinite since : → Flip not valid → contour has 2 consecutive outside corners associated to 2 convex vertices of the CMLP (+0m +) and m ≥ 1 → Flip not valid → contour has 2 consecutive outside corners associated to 2 convex vertices of the CMLP (+0m +) and m ≥ 1 ... F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy Descent Theorem Let Γ ∈ C(A, B) being finite, choosing at each step any one of the valid convex or concave deformation leads to CMLP(Γ) = [AB] after a finite number of steps. − − + F. de Vieilleville + Réunion Equipe TCI + Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Greedy1 Algorithm : Lowest Energy Among all Deformations Input: Γ : a Digital Deformable Model Output: Γ0 : when returning true, elementary deformation of Γ with D D EDM (Γ0 ) < EDM (Γ), otherwise Γ0 = Γ and it is a local minimum. Data: Q : Queue of (Deformation, double) ; D E0 ← EDM (Γ); foreach valid Deformation d on Γ do Γ.applyDeformation(d); D Q.push back(d, EDM (Γ) ); Γ.revertLastDeformation(); end (d, E1 ) = SelectDeformationWithLowestEnergy (Q); Γ0 ← Γ; if E1 < E0 then Γ0 .applyDeformation(d); return E1 < E0 ; F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Greedy2 Algorithm : First Valid Decreasing the Energy Input: Γ : a Digital Deformable Model Output: Γ0 : when returning true, elementary deformation of Γ with D D (Γ), otherwise Γ0 = Γ and it is a local minimum. EDM (Γ0 ) < EDM D E0 ← EDM (Γ); foreach valid Deformation d on Γ do Γ.applyDeformation(d); D if EDM (Γ) < E0 then Γ0 ← Γ; return true; Γ.revertLastDeformation(); end Γ0 ← Γ; return false; F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Synthetic Images Initial α=0 α = 200 α = 300 F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Synthetic Images Inital α=0 α = 800, 4000 α = 1600, 8000 F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Real Images : Top-Down Approach Binary Mask Image Initialisation Minimisation 32x 32 Initialisation From Scaling Minimisation Initialisation From Scaling Minimisation 64x64 128x128 F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Real Images : Top-Down Approach Initial Minimisation α = 400, internal energy term is based on the CMLP. The high value of α penalise the length of the contours, over-smoothing the contours, only top-left contour seems to be correctly delineated. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Real Images : Top-Down Approach Initial Minimisation α = 150. The internal energy term is based on the CMLP. Top-left and bottom-left contours are correctly delineated. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Real Images : Top-Down Approach Initial Minimisation α = 100. The internal energy term is based on the CMLP. Global delineation of contours is correct. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Real Images : Top-Down Approach Initial Minimisation α = 100. The internal energy term is based on the length of the freeman code of the path, other positive values for the balance term bring very similar contours. Although the global delineation seems correct, the obtained contours are not smoothed at all. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation On Real Images : Top-Down Approach Initial Minimisation Results of the top-down segmentation process using the deformable model described in [William’92]. Parameters are such that (α, β) equal (0.1, 0.45), the neighborhood is chosen as a 3 × 3 square, and minimisation is done until no smaller energy can be found. Top-right contour and middle contours are well delineated, top left and bottom left are partially correctly delineated. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Internal Energy Descent SpiraleFIRST.avi → Keep the topology of Γ. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction and Problem Statement Proposed Solution Experimental Evaluation Conclusion Proposed method is as good as classic snake formulation but inherits its drawback : tuning of parameter α. Probabilistic deformation scheme ? Dynamic CMLP algorithm ? More comparison with “classic” active contours ? F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction-Motivation Images obtenues par un microscope sur des objets marqués par des fluorophores : Laser gaussien + lentille cylindrique+ objectif à l’emission, chambre d’immersion (4 degrés de liberté)+ agar + spécimen, objectif d’imagerie (grossissement) + filtre + caméra CCD. Illumination sélective de l’objet par feuilles de lumière : Moins phototoxique, possibilité d’imager en temps réel, d’imager plus en profondeur dans le spécimen. Pas de photo-blanchiment des fluorophores. Coupes 2D d’un objet 3D. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Introduction-Motivation Restaurer les images (3D) ayant subi des dégradations lors du processus de formation : déconvolution. Connaissance a priori sur la fonction de dégradation de l’objet et le processus de formation de l’image. → type de bruit de la caméra : Poissonien + thermique → optique linéaire, psf théorique connue F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution : Richardson-Lucy Bruit Poissonien : f (λ, k) = λk exp−λ k! Modèle th. de formation : i(x) = P(o ⊗ h)(X ) Approche Bayésienne : p(o|i) = p(i|o)p(o) p(i) MaximiserQla vraisemblance selon la loi de Poisson : i(x) −(h⊗o)(x) p(i|o) = x [(h⊗o)(x)] i(x)!exp Minimiser la log-vraisemblance P −log (p(i|o) = x [(h ⊗ o)(x) − i(x) log(h ⊗ o)(x) F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution : Richardson-Lucy MinimiserR la fonctionnelle : J1 (o) = x [(h ⊗ o)(x) − i(x) log(h ⊗ o)(x)]dx CN : ∇J1 (o) = 0 ⇔ h(−x) ⊗ i(x) (h⊗p)(x) =1 Algorithme itératif h multiplicatif : i i(x) ok+1 (x) = ok (x) h(−x) ⊗ (h⊗o k )(x) F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution : Richardson-Lucy Algorithme multiplicatif non-stable : Applification du bruit après quelques itérations Problèmes numériques Algorithme additif (descente en gradient) : i h i(x) ok+1 (x) = ok (x) + δt 1 − h(−x) ⊗ (h⊗o k )(x) Nécessite une régularisation ! F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution : Richardson-Lucy avec Total Variation Minimiser la fonctionnelle : R J2 (o) = J1 (o) + λTV x |∇(o)(x)|dx CN : ∇J2 (o) = 0 ⇔ 1 − h(−x) ⊗ i(x) (h⊗p)(x) − λTV div ok+1 (x) = h ∇o ok (x) + δt −1 + λTV div |∇o| + h(−x) ⊗ F. de Vieilleville Réunion Equipe TCI ∇o |∇o| i(x) (h⊗ok )(x) i =0 Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution : Richardson-Lucy avec Total Variation Descente en gradient avec deux paramètres Balance de régularisation Paramètre d’intégration Performance de la méthode liée au schéma numérique employé Calcul des différences finies Schéma ROF [Rudin et al.’92] on retombe dans des problèmes d’estimation ! ! F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution : Richardson-Lucy avec Total Variation F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution d’images SPIM : Problèmes a posteriori Les images présente des raies sombres et claires : Absorption dues à des éléments non imagés ! ! Focalisation Dirigées selon la direction de la feuille de lumière. F. de Vieilleville Réunion Equipe TCI Tangent Estimators + Evaluation + Enhancement Digital Deformable Model Onco-SP-IM : Image Restoration Déconvolution d’images SPIM Création d’images synthétiques pour obtenir une “vérité terrain” afin de mieux comparer les méthodes. Tests d’autres schémas numériques pour comparaison et évaluation des performances. Séparer les problèmes : débruitage, suppression des raies, déconvolution. F. de Vieilleville Réunion Equipe TCI