multi-dimensional image analysis - petra christian university
TRANSCRIPT
Center for Image Processing
Multi-dimensional Image Analysis
Lucas J. van Vlietwww.ph.tn.tudelft.nl/~lucas
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 2
Image Analysis Paradigm
Texture filtering
analysis
segmentation
Imagerestoration
Imageenhancement
sensorImageformation scene
pre-processing
classification
Measurements:Point: edge location, isophote curvaturesGlobal: size and shape descriptorsLocal: texture attributes anisotropy, orientation, scale
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 3
Multi-dimensional analysis
Goal: Use sampling-error free measurements for analog properties in digitized images
� Point measurements for object identification� Exact boundary location� (Principal) curvatures
� Global measurement for object description� Integrated object intensity (gray-volume)� Size in 2D: area and perimeter� Size in 3D: volume, surface area, length� Shape: Bending energy, Euler numbers
� Local texture analysis� Anisotropy, orientation, scale
Use gray-scale rather than binary operations to obtain high accuracy and precision
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 4
Better geometric measurements
� State-of-the-art cameras offer:� Large images of 1000 x 1000 pixels or more� 12 – 16 bit photometric information
� A binary image is disturbed by aliasing� Thresholding corrupts data irreversible
Faithful representation Binary representation
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 5
Location of curved edges
� Zero-crossing of second derivatives applied to blurred curved edges are biased.� Laplace: Dxx + Dyy = Dcc + Dgg outwards� in grad direction: Dgg inwards� PLUS 2Dgg + Dcc on edge !
� Results on an ellipse: object with slowly varying curvature.
PLUS
SDGDLaplace
PLUS
x
y
g
cg
cg
c g = gradientc = contour
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 6
Isophote curvature in 2D
An isophote is a curve of constant gray-value (level curve)� Curvature is the change of contour direction per unit length� Curvature (κ=1/R) with R the radius of the osculating circle
cc
g
DD
κ
−
=object contour
g =(Dx,Dy)
c =(-Dy,Dx) Usage: corner detection,dominant points, bending energy
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 7
Isophote curvature in 3D
� 2D isophote surface patch in 3D image space� Principal curvatures κ1 and κ2 along c1 and c2: 1 2c c g⊥ ⊥
1 11
c c
g
DD
κ
−
=2 2
2c c
g
DD
κ
−
=
1 2 0κ κ> > 2 1 0κ κ< < 1 20, 0κ κ> =
1 20, 0κ κ> < 1 2 0κ κ= =
Usage for local shape: elliptic (convex, concave), cylinder,saddle, flat
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 8
Sampling-error free measures
� A sampling-error free measurement is a digital measurement performed on a sampled image that is exactly equal to its analog counterpart
The “sum” of all samples is measured without thresholding and does not introduce a sampling error
( ) ( ) ( ) ( )2
12
sum , 2 0,0
iff :x y x y
sampling Nyquist
b b i j B
f f
π= ∆ ∆ ∆ ∆ =
>
∑
Digital track
=
Analog track
reconstruction
operation in analog space
samplingoperation in digital space
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 9
Sum() as measure for …
� Sum() is a sampling-error free measure
� Recipe for object measurements in gray-scale images:1. Transform the input image with the object into an output image
whose sum() is directly proportional to the feature to be measured
2. The transformation must consist of sampling-error free operations
3. Proper sampling is required to avoid aliasing
4. Bias correction terms can be extracted from the mathematical framework! (not empirically)
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 10
2D area & 3D volume
� Clip image a to produce a flat object on a flat background� Use “soft-clipping” to avoid aliasing
output after:
thresholding
hard-clipping
erf-clipping( )
( )
erf
2
3
clip
sumD
D
b a
Ab b
V
=
= =∑
linear input slope
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 11
2D perimeter, 3D surface area
� Transform the “flat”, bandlimited object into a contour
Analytical dilation / erosion by Taylor series around a smooth edge (with height = 1) over a distance δ =½
( ) ( )
( ) ( )
12
212
212
:
:
:
g gg
g gg
g
dilation b r b r b b
erosion b r b r b b
contour bδ
δ δ δ
δ δ δ
=
+ ≈ + +
− ≈ − + −
≈
a
b
|bg |
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 12
Shape: 2D bending energy
� 2D bending energy proportional to the bending energy of a deformed circular rod (Young ’74).
� The simply-connected, closed contour with the minimum bending energy is the circle (2π/R, not scale invariant).
b |bg| κ κ2
Differential geometry Image analysis
( )2 2be be gE s ds E bκ κ= =∑∫�
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 13
Shape: 3D bending energy
� Elastic rods (SCC space curves)� κ1 ⇒ cross-section of rod� κ2 ⇒ trajectory of rod
� circle has minimal bending energy
� Torque forces are neglected
� Deflected thin plates (SCC surfaces)� principal curvatures κ1 and κ2
� Poisson’s ratio p [0 ,½]. (let p = 0)� sphere has minimal energy (8π). � Dimensionless and therefore scaling
invariant.
( )22rodE s dsκ= ∫� ( )2 2
1 28plateE p dSπ κ κ= + +∫∫�
22rod gE bκ=∑ ( )2 2
1 2plate gE bκ κ= +∑
Differential geometry
Image analysis
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 14
Length in 3D
Resume: An erf-clipped object of unit intensity, bvolume: V = Σ bsurface area: A = Σ |bg|length: L = ??
� image b contains a cylinder (length L,radius R)Σ b = πR2 L = volumeΣ bg = 2πRL = areaΣ bgg = 2πL = length
� Length of spaghetti
� In the plane perpendicular to the string: g,cΣ bgg + bcc = 0 all g • c = 0Σ bgg = –Σ bcc = 2π (see Euler)
13 2D ggL b
π
= ∑
b
bcc
bgg
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 15
Shape: Euler numbers
� Euler numbers characterize the topology2D: number of objects – number of holes3D: number of objects – number of handles
(donut & coffee cup have Euler number = 0)
Differential geometry Image analysis
Hopf:
Gauss-Bonnet:
( ) 12 2
11 2 3 1 24
2
4
D g
D g
s ds N b
dS N b
π
π
κ π κ
κ κ π κ κ
= =
= =
∑∫
∑∫∫
�
�
b
bg
κ
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 16
Curvilinear structures
10 km 10 cm
1 cm
10 cm
∆z = 10 µm
� Anisotropy� Orientation� Scale� Curvature
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 17
Domains vs Scale
� Texture attributes are domain properties
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 18
Orientation and Scale
� Rotational invariant chirp image
Orientation mapby Gradient Structure Tensor
Scale mapby Gaussian Scale Space
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 19
Gradient Structure Tensor
� How to combine vectors?
ϕ
λ 1
λ 2
= ⋅ =
2
2
x x yt
x y y
f f f
f f fG g g
xf
yf
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 20
GST: anisotropy, orientation
� Closed-form solutions for:
12 2
2tan x y
x y
f f
f fϕ
−
= −
1 2
1 2anisotropy: A λ λ
λ λ
−=
+
( ) ( )
( ) ( )
2 22 2 2 211 2
2 22 2 2 212 2
4
4
x y x y x y
x y x y x y
f f f f f f
f f f f f f
λ
λ
= + + − +
= + − − +
ϕ
λ1 λ2
A
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 21
PVC particle roughness
� Contour information (shape features) failed to rank batches according to “quality”
� Lobes show up as lines rather than points� Measure the local anisotropy (ellipses)
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 22
Roughness = Integrated anisotropy
� Three scales: gradient, window, particle
λ
λ= − =
∑∑
2
11 0.44R
λ
λ= − =
∑∑
2
11 0.61R
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 23
Roughness distributions
� (Cumulative) distributions for N=30 particles can be used as roughness measure
� Roughness measure correlates well with product quality
0.39 0.42 0.45 0.48 0.51 0.54 0.57 0.
6
0.63 0.66 0.69 0.72
smoothmiddle
rough
0
2
4
6
8
10
12
coun
t
roughness
particle roughness histograms
smoothmiddlerough
0.39 0.42 0.45 0.48 0.51 0.54 0.57 0.60 0.63 0.66 0.69 0.72
smoothmiddle
rough0%
25%
50%
75%
100%
roughness
cumulative particle roughness histograms
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 24
Orientation-driven analysis
� Dominant orientation of Gradient Structure Tensor� Strongest peak in orientation space
Dominant orientationof GST yields the gradient2-weightedorientation map.
Strongest orientationin ϕ-space allows veryabrupt changes inorientation map.
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 25
Orientation space
� Decompose the image into narrow orientation bands
� Apply a nonlinear operator in orientation space:� Labeling yields segmentation� Peak selection for detection
ΦΦΦΦ ΦΦΦΦ−−−−1111
ΦΦΦΦ
selectivity
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 26
Circle in image-space …
� A circle in 2D image space becomes a double-helix in orientation space.
� Note that orientation-axis is periodic with π
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 27
Overlapping circles
� Two overlapping circles in 2D image space compose a single object.
� Since the circles cross at different orientations, they become separated in orientation space.
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 28
Orientation selectivity
� Green = first peak, Red = second peak, Blue = remainder
N=8 N=16 N=32
orientation
N=8
N=16
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 29
Multi-Scale
Series of images filtered of decreasing scale: Scale-space� Sample the scales logarithmically using filters of size = basescale
yields n scales per octave
Input imagescale 0
scale 1
scale 2
scale 3
scale 4
scale 5
var (scale 1)
var (scale 2)
var (scale 3)
var (scale 4)
var (scale 5)
Local variance between scales n and n-1.
{ }11 13212 ,2 ,2 ,...,2 nbase ∈
Scale difference ≈ scale derivative
Scal
e sp
ace
Scal
e sp
ace
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 30
Chirp exampleSc
ale-
spac
e
Scale derivative Spatial variance
Scal
e-sp
ace
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 31
Gaussian scale space
� Chirp of varying contrast� Low-pass filters of increasing scale:
color code (fine to course)� Normalization: sum per pixel is constant
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 32
Scale analysis
� Scale information reveals geological structures in seismic data
Absolute Normalized
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 33
Morphological scale space
� Measure the hole-size distributionThe image acts like a sieve
� Subtract closings of increasing scaleThe hole-size is the scale that closes the gap
2 4 8 16 32 64 128 256
Scale256 x 256
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 34
Labeling of holes by size
2 4 8 16 32 64 128 256 2 4 8 16 32 64 128 256
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 35
Pore-size distribution
2 4 8 16 32 64 128 256
Milk + substrateMilk + substrate + enzime C
Milk
Milk (blue) + substate (red) + enzyme (green)
Average pore-size distributions imagesN=32
Center for Image ProcessingCenter for Image Processing
CBP course: Multi-Dimensional Image Analysis 36
Literature
pdf files available at: www.ph.tn.tudelft.nl/~lucas
� L.J. van Vliet and P.W. Verbeek, Better geometric measurements based on photometric information, Proc. IEEE Instrumentation and Measurement Technology Conf. IMTC94 (Hamamatsu, Japan, May 10-12), 1994, 1357-1360.
� G.M.P. van Kempen et al., The application of a local dimensionality estimator to the analysis of 3D microscopic network structures, in: SCIA'99, Proc. 11th Scandinavian Conference on Image Analysis (Kangerlussuaq, Greenland, June 7-11), 1999, 447-455.
� M. van Ginkel et al., Improved Orientation Selectivity for Orientation Estimation, in: SCIA'97, Proc. 10th Scandinavian Conference on Image Analysis (Lappeenranta, Finland, June 9-11), 1997, 533-537.