Presentation is loading. Please wait.

Presentation is loading. Please wait.

Air Systems Division Definition of anisotropic denoising operators via sectional curvature Stanley Durrleman September 19, 2006.

Similar presentations


Presentation on theme: "Air Systems Division Definition of anisotropic denoising operators via sectional curvature Stanley Durrleman September 19, 2006."— Presentation transcript:

1 Air Systems Division Definition of anisotropic denoising operators via sectional curvature Stanley Durrleman September 19, 2006

2 1 MIA’06 - September 2006 Air Systems Division The problem coast crest Purpose : Denoising homogeneous areas… …without smoothing the signal at the interfaces

3 2 MIA’06 - September 2006 Air Systems Division The problem Autoregressive model :

4 3 MIA’06 - September 2006 Air Systems Division The problem Autoregressive model :

5 4 MIA’06 - September 2006 Air Systems Division The problem Autoregressive model :

6 5 MIA’06 - September 2006 Air Systems Division The problem Autoregressive model : Burg algorithm enables : -better estimation in case of short sample signals -fewer interference peaks -recursive computation : real time algorithm -estimation of the spectral density function :

7 6 MIA’06 - September 2006 Air Systems Division The problem Example : record of turbulent atmospheric clutter Images du CR 1 magnitude angle

8 7 MIA’06 - September 2006 Air Systems Division What’s in the image proceesing toolbox ? -Statistical models of noise Bayesian models, Markov fields… : - good model of noise - how to take the geometry into account ? -Geometrical models : Linear filters (Gaussian,…) : do not preserve the discontinuities Non-linear filters : - Curvature motion & morphologic filters (AMSS, mean curvature motion, median filter) : - noise = level set of small areas - specific for gray-level images - Geometric filters : (Kimmel, Sochen, Barbaresco) : - model data as a sub-manifold - depend on the way data are parametrized (mean curvature flow) - model of noise ?  Our goal : define anisotropic operators that can denoise data…  of any dimension (gray-level images, radar signal…)  independently of the data parametrization  and restore piecewise constant data

9 8 MIA’06 - September 2006 Air Systems Division Outline  Noise characterization via sectional curvature  De-noising algorithms  Results

10 Air Systems Division I. Noise characterization through sectional curvature MIA – September 19, 2006

11 10 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 1. Question : what is noise ?? statistics : Bayesian filters, maximum likelihood… geometry : which tool ? Gradient ? Curvature !

12 11 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 2- Basic idea : the surface Gaussian Curvature

13 12 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 2- Basic idea : the surface Gaussian Curvature Examples :

14 13 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature Noise and curvature Axiom : pixel of noise = pixel of big curvature

15 14 MIA’06 - September 2006 Air Systems Division How to denoise ? By minimizing the following energy : I. Noise & Sectional Curvature

16 15 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 3 – Modeling A generic ‘image’ :

17 16 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 3 – Modeling

18 17 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 3 – Modeling Curvature of a metric :

19 18 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature 3 – Modelisation Curvature of a metric : That is the surface Gaussian curvature !

20 19 MIA’06 - September 2006 Air Systems Division I. Noise & Sectional Curvature Summary : 1/ One defines : h metric on the data space e metric on the acquisition space => a ‘mixed’ metric : g 2/ One computes the sectional curvature: K 3/ One defines the energy : E

21 Air Systems Division II. De-noising algorithms MIA’06 - September 19, 2006

22 21 MIA’06 - September 2006 Air Systems Division II. De-noising algorithms Purpose : Minimizing : 2 methods : - Partial Differential Equation - Stochastic algorithm

23 22 MIA’06 - September 2006 Air Systems Division 1.Descent gradient scheme : 1/ initialise with the given noisy image 2/ Evolve towards a minimum of : using the gradient : Hence, the evolution equation : implemented with a finite difference scheme. II. De-noising algorithms

24 23 MIA’06 - September 2006 Air Systems Division II. De-noising algorithms Case of gray-level images :

25 24 MIA’06 - September 2006 Air Systems Division II. De-noising algorithms 2. Stochastic method : - One picks randomly a pixel in the (noisy) image. - One adds a small random Gaussian variable to the pixel’s value. - If the energy decreases : one keeps the change Else : the change is rejected.

26 Air Systems Division III. Results MIA’06 - September 19, 2006

27 26 MIA’06 - September 2006 Air Systems Division III. Results 1 – Gray-level images (1) Metric : Curvature : Flow :

28 27 MIA’06 - September 2006 Air Systems Division III. Results t = 0t = 1t = 10t = 100 Flow equation

29 28 MIA’06 - September 2006 Air Systems Division III. Results Transversal view :

30 29 MIA’06 - September 2006 Air Systems Division III. Results Stochastic Stoch. + PDE Stochastic algorithm :

31 30 MIA’06 - September 2006 Air Systems Division III. Results 2 – Gray level images (2) Adaptative metric : ‘Dilate’ geodesics in D far from the minimum

32 31 MIA’06 - September 2006 Air Systems Division III. Results originalPDE Stoch. Algo.Adaptative metric

33 32 MIA’06 - September 2006 Air Systems Division III. Results median

34 33 MIA’06 - September 2006 Air Systems Division III. Results RSO Image original amss PDE Stoch

35 34 MIA’06 - September 2006 Air Systems Division III. Results 2. Radar signal Reminder : complex auto-regressive analysis Parametrization thanks to complex auto-regressive analysis 8 complex magnitudes 7 reflection coefficients Doppler spectrum Burg Algo.

36 35 MIA’06 - September 2006 Air Systems Division III. Results Data : Reflection coefficients

37 36 MIA’06 - September 2006 Air Systems Division III. Results Simulated data : Image of CR 1 : magnitude angle azimut 16

38 37 MIA’06 - September 2006 Air Systems Division III. Results Simulated data : Image of CR 1 : magnitude angle azimut 16

39 38 MIA’06 - September 2006 Air Systems Division III. Results After de-noising:

40 39 MIA’06 - September 2006 Air Systems Division III. Results After de-noising :

41 40 MIA’06 - September 2006 Air Systems Division III. Results Real data : Images du CR 1 magnitude angle azimut 19

42 41 MIA’06 - September 2006 Air Systems Division III. Results After de-noising :

43 Air Systems Division Thank you for your kind attention MIA’06 - September 19, 2006

44 43 MIA’06 - September 2006 Air Systems Division

45 44 MIA’06 - September 2006 Air Systems Division

46 45 MIA’06 - September 2006 Air Systems Division

47 46 MIA’06 - September 2006 Air Systems Division


Download ppt "Air Systems Division Definition of anisotropic denoising operators via sectional curvature Stanley Durrleman September 19, 2006."

Similar presentations


Ads by Google