View
212
Download
0
Embed Size (px)
Citation preview
11
Contour Enhancement and Completion via Left-Invariant Second Order
Stochastic Evolution Equations on the 2D-Euclidean Motion Group
Erik Franken, Remco Duits,Markus van Almsick
Eindhoven University of Technology
Department of Biomedical Engineering
EURANDOM workshop
“Image Analysis and Inverse Problems”
December 13th 2006, Eindhoven, NL
22
Outline
• Introduction to Orientation Scores• Invertible Orientation Scores• Operations in Orientation Scores• The Direction Process for Contour completion• Nonlinear diffusion for Contour enhancement• Conclusions
33
1. The retina contains receptive fields of varying sizes multi-scale sampling device
2. Primary visual cortex is multi-orientation
Biological Inspiration
• Cells in the primary visual cortex are orientation-specific
• Strong connectivity between cells that respond to (nearly) the same orientation
Measurement in Primary Visual Cortex
Bosking et al., J. Neuroscience 17:2112-2127, 1997
44
image
Orientation Scores
From 2D image f(x,y) to orientation score Uf(x,y,θ) with position (x,y) and orientation θ
x
y
xorientation score
y
55
Orientation Score is a Function on SE(2)
Properties of SE(2)• Group element
translation rotation• Group product
• Group inverseg¡ 1 = (¡ R ¡ 1
µ b; ¡ µ)
SE (2) = R2o SO(2) = Euclidean Motion GroupWe consider orientation score U 2 L2(SE (2))
66
Approach: Image Processing /analysis via Orientation Scores
“Enhancement” operation
Initial image
“Enhanced” image
Orientation score transformation
Inverse orientation score transformation
Segmented structures
Segment structures of interest
77
Outline
• Introduction to Orientation Scores• Invertible Orientation Scores• Operations in Orientation Scores• The Direction Process for Contour completion• Nonlinear diffusion for Contour enhancement• Conclusions
88
Invertible Orientation Score Transformation
i.e. “fill up the entire Fourier spectrum”.
Image to orientation score
Orientation score to image:
Stable reconstruction requires
Oriented wavelet
99
Invertible Orientation Score TransformationDesign considerations: reconstruction,
directional, spatial localization, quadrature, discrete number of orientations
re(Ã) im(Ã) F [Ã]P Nµ
j =0 F [R j sµ Ã]dµ
re(WÃf )(¢;µ)example image f j(WÃf )(¢;µ)j
1010
Outline
• Introduction to Orientation Scores• Invertible Orientation Scores• Operations in Orientation Scores• The Direction Process for Contour completion• Nonlinear diffusion for Contour enhancement• Conclusions
1111
Represents the “net” operator. It is Euclidean-invariant iff is left-invariant, i.e.
Left Invariant Operators
L g©U = ©L gU; i 2 f »;´;µg
(L gU)(h) = U(g¡ 1h)Where is the left-regular
representation L g
1212
Left-invariant Derivative Operators
areleft-invariant derivatives onEuclidean motion group, i.e.
@»
@́
@µ
x
µ
y
@»@́
@µ@»; @́; and @µ
L g@iU = @iL gU; i 2 f »;´;µg
Not all left-invariant derivatives on SE(2)do commute![@µ;@»] = @µ@» ¡ @»@µ = @́; [@µ; @́] = ¡ @»
=Tangent to line structures
= orthogonal to line structures
1313
Convection-diffusion PDEs on SE(2)
convection diffusion
Time process
Resolvent process
1414
Linear & Left-invariant Operators are G-convolutionsNormal 3D convolution – versus G-convolution on SE(2)
“G-Kernel”
1515
Outline
• Introduction to Orientation Scores• Invertible Orientation Scores• Operations in Orientation Scores• The Direction Process for Contour
completion• Nonlinear diffusion for Contour enhancement• Conclusions
1616
Direction Process on SE(2)Resolvent of linear PDE
Random walker interpretation
1717
Stochastic Completion Fields
Collision probability of 2 random walkers on SE(2):
Forward Backward
The mode line (in red) is the most likely connection curvebetween the two points
1818
An example
Noisy input image
Greens functions:
“Simple” enhancement viaOrientation score
Stochastic completion field
1919
Exact Solution by Duits and Van Almsick
Explicit PDE problem, case (Mumford) :Analytic Solution of Greens function?
2020
Practical approximations
2121
Exact Green’s Function versus Approximation
The smaller the better approximation
2222
How? •G-convolution with exact/approximate Green’s function•Finite element implementation in Fourier domain (August / Duits)•Explicit numerical schemes (Zweck and Williams)
Application of the Direction process
Non-linear enhancement step
Initial image
Image with completed contours
Orientation score transformation
Inverse orientation score transformation
Compute Stochastic Completion Field
What?
•Orientation-score gray-scale transformation (i.e. taking a power)•Angular/spatial thinning
2323
Automatic Contour completion by SE(2)-convolution
2424
Outline
• Introduction to Orientation Scores• Invertible Orientation Scores• Operations in Orientation Scores• The Direction Process for Contour completion• Nonlinear diffusion for Contour
enhancement• Conclusions
2525
The Diffusion Equation on Imagesf = image
u = scale space of image
D = diffusion tensor
Linear diffusion Perona&Malik Coherence-enhancing diff.
2626
Diffusion equation in orientation scorescurvature
Diffusion orthogonal tooriented structures
Diffusion tangent tooriented structures
Diffusion inorientation
Evolvingorientationscore
Rotating tangent spacecoordinate basis
Left-invariantderivatives
are left-invariant derivatives onEuclidean motion group, i.e.
@»
@́
@µ
x
µ
y
@»@́
@µ
@»; @́; and @µ
L g@iU = @iL gU; i 2 f »;´;µg
2727
Example diffusion kernels
2828
• Oriented regions: D’11 and D33 small, D22 large and κ according to estimate
• Non-oriented regions: D’11 large, D22=D33 large, κ = 0
@»
@́
@µ
x
µ
y
@»@́
@µ
How to Choose Conductivity Coefficients
2929
Measure for Orientation Strength
• Hessian in Orientation Score
Note: non-symmetric due to non-commuting operators!
• Gaussian Derivatives can be used, if one ensures to first take orientational derivatives and then spatial derivatives.
• Measure for orientation strength:
3030
Curvature estimation
• If a vector points tangent to a structure in the orientation score, the curvature in that point is:
Ideally zero
• Estimation of v:1. Determine eigenvectors and eigenvalues of
2. Select the 2 eigenvectors closest to the ξ,θ-plane
3. Take eigenvector corresponding to the largest eigenvalue
3131
@»
@́
@µ
x
µ
y
@»@́
@µ
Chosen Conductivity Coefficients
3232
• Numerical scheme: explicit, left-invariant finite differences.
• Using B-spline interpolation cf. Unser et al.
Implementation
3333
Diffusion in orientation score Coherence enhancing diffusion Results
Size: 128 x 128 x 64
3434
Collagen image
Diffusion in orientation score Coherence enhancing diffusion
Size: 200 x 200 x 64
3535
Results – with/without curvature estimation
3636
Outline
• Introduction to Orientation Scores• Invertible Orientation Scores• Operations in Orientation Scores• The Direction Process for Contour completion• Nonlinear diffusion for Contour enhancement• Conclusions
3737
Conclusions
• We developed a framework for image processing via Orientation scores. Important notion: An orientation score is a function on SE(2) use group theory.
• Useful for noisy medical images with (crossing) elongated structures
• Found Analytic Solution of Greens functions• Stochastic Completion Fields of images using
G-convolutions• Non-linear diffusion on orientation scores to
enhance crossing line structures
3838
Current/Future work
• Improving adaptive/nonlinear evolutions on SE(2)– Numerical methods– Nonlinearities
• Applying in medical applications– 2-photon microscopy images of Collagen fibers– High Angular Resolution Diffusion Imaging
• Apply same mathematics in other groups, e.g. SE(3) and similitude groups.
3939
4040
Line enhancement in 3D via invertible orientation scores
Application: Enhancement Adam-Kiewitz vessel
4141
Enhancement Kidney-Boundaries in Ultra-sound images
Via Orientation Scores:
4242
Medical Application: Cardiac ArrhythmiasHeart rythm disorder by extra conductive path/spot
Catheters in hart provide intracardiogram and can burn focal spots/lines
Navigation by X-ray
Detection cathethers in X-ray navigation
automatic 3D-cardiac mapping from bi-plane.
4343
Efficient calculation of G-convolutions
• Using steerable filters in the orientation score + inspired by Fourier transform on SE(2)
Algorithmic complexity can be reduced from
to