content based image retrievalromip.ru/russir2008/slides/cbir/lecture-3.pdf · content based image...
TRANSCRIPT
Content Based Image Content Based Image RetrievalRetrievalNatalia [email protected]
HP Labs Russia
© 2008 Hewlett-Packard Company
Tutorial outlineTutorial outline• Lecture 1
− Introduction− Introduction
− Applications
• Lecture 2− Performance measurement− Performance measurement
− Visual perception
− Color features
• Lecture 3• Lecture 3− Texture features
− Shape features
− Fusion methods− Fusion methods
• Lecture 4− Segmentation
Local descriptors− Local descriptors
• Lecture 5− Multidimensional indexing
2/51
− Survey of existing systems
Lecture 3Lecture 3
Texture featuresTexture features
Shape featuresShape features
Fusion methods
Lecture 3: Outline
• Texture features
Lecture 3: Outline
• Texture features• Texture features
− Statistical
− Spectral
• Texture features
− Statistical
− Spectral− Spectral
− Comparison
− Spectral
− Comparison
• Shape features
− Boundary based
• Shape features
− Boundary based− Boundary based
− Region based
− Comparison
− Boundary based
− Region based
− Comparison− Comparison
• Fusion methods
− Comparison
• Fusion methods
4/51
Texture featuresTexture features
• What is texture?• What is texture?
5/51
Smooth Rough Regular
Texture featuresTexture features
Texture features
General statistics parameters
Haralick’s co-occurrence matrices
Statistical
PWT
Spectral
Model basedTamura features
Geometrical
TWT
DCT, DST, DHT
Complex wavelets
Gabor filters
ICA filters
Markov random fields
Fractals
Model-based
Voronoi tesselation features
Structural methods
GeometricalICA filters
6/51
Texture featuresTexture features
• General statistics• General statistics
Based on intensity histogram of the whole image or its regions:
– histogram of intensity, L – number of intensity levels.– histogram of intensity, L – number of intensity levels.
– central moment of order n.
– average intensity.
– variance, is a measure of contrast.
, R=0 where intensity is equal., R=0 where intensity is equal.
– a measure of histogram assimetry.
7/51
– a measure of histogram assimetry.
Texture featuresTexture features
• General statistics (2)• General statistics (2)
– a measure of contrast of homogeneity
(max for homogeneous areas ).(max for homogeneous areas ).
– entropy, a measure of variability
(0 for homogeneous areas ).(0 for homogeneous areas ).
Texture Average Deviation R µ3 U Entropy
SmoothSmooth
Rough
Regular
8/51
Texture featuresTexture features
Grey Level Co-occurrence Matrices (GLCM):Grey Level Co-occurrence Matrices (GLCM):
GLCM - matrix of frequencies at which two pixels, separated by a certain vector, occur in the image.
∑∑ =∆+∆+=
=N M
otherwise
jyqxpIiqpIifjiC
0
1
,
),(,),(,),( ∑∑
= = =
p q otherwisejiC
1 1 0,),(
),( yx ∆∆ – separation vector;
I(p,q) – intensity of a pixel in position (p, q).
9/51
GLCM – an exampleGLCM – an example
10/51
)0,1(),( =∆∆ yx
GLCM – descriptorsGLCM – descriptors
Statistical parameters calculated from GLCM values:
∑∑= jiCEnergy ),(2
Statistical parameters calculated from GLCM values:
– is minimal when all elements are equal∑∑=i j
jiCEnergy ),(
∑∑−= jiCjiCEntropy ),(log),(
– is minimal when all elements are equal
– a measure of chaos, is maximal when all elements are equal
∑∑i j
∑∑ −=i j
jiCjiContrast ),()( 2
is maximal when all elements are equal
– has small values when big elements are near the main diagonal
∑∑i j
∑∑ −+=
i j ji
jiCMomentDifferenceInverse
2)(1
),(
near the main diagonal
−+i j ji )(1
– has small values when big elements are far from the main diagonal
11/51
Texture features: Tamura featuresTexture features: Tamura features
Features, which are important for visual perception:Features, which are important for visual perception:
� Coarseness Tamura image:Coarseness
� Contrast
� Directionality
Coarseness-coNtrast-Directionality –
points in 3-D space CND
� Directionality
� Line-likeness
Regularity
� Euclidean distance in 3D (QBIC)
� 3D histogram (Mars)
Features:
� Regularity
� Roughness
� 3D histogram (Mars)
12/51
Texture features: spectralTexture features: spectral
Texture features
General statistics parameters
Haralick’s co-occurrence matrices
Statistical
PWT
Spectral
Model basedTamura features
Geometrical
TWT
DCT, DST, DHT
Complex wavelets
Gabor filters
ICA filters
Markov random fields
Fractals
Model-based
Voronoi tesselation features
Structural methods
GeometricalICA filters
13/51
Texture features: wavelet basedTexture features: wavelet basedWavelet analysis – decomposition of a signal:
)()( ,
,
xxf kj
kj
kψα∑=
Basis functions: Basis functions:
)()(,,
)2(2
2
2/
,
RLxkj
kxjj
kj
∈Ζ∈
−=
ϕ
ϕψ – scaling function
– mother wavelet
A set of basis functions – filters bank
Image
– mother wavelet
Filter 1
Image
Energy 1Filter 1
Filter 2
Filter N
Energy 1
Energy 2
Energy N
Feature vector
14/51
Filter N Energy N
Texture features: Gabor filtersTexture features: Gabor filters
Mother wavelet: Gabor functionMother wavelet: Gabor function
+
+−
= jWxyx
yxg π21exp
1),(
22
+
+−
= jWx
yxyxg
yxyx
πσσσπσ
22
1exp
2
1),(
22
1,-S0,1,...,minteger,,,1),,(),( ==>′′= − nmayxgayxg m
mn
Filters bank:
),cossin(
),sincos(
1,-S0,1,...,minteger,,,1),,(),(
Θ+Θ−=′Θ+Θ=′
==>′′=
−
−
yxay
yxax
nmayxgayxg
m
m
mn
),cossin( Θ+Θ−=′ yxay
Kn /π=Θ)1/(1)/( −−= S
lh UUa
К – a number of directions, S – a number of scales,Uh, Ul – max and min of frequencies taken into consideration.
15/51
lh
Texture features: ICA filtersTexture features: ICA filtersFilters are obtained using Independent Component Analysis
I1
dist(I1,I2) = KLH(H1i , H2i)Σi=1
N
H. Borgne, A. Guerin-Dugue, A. Antoniadis.
Representation of images for classification
with independent features. Pattern with independent features. Pattern
Recognition Letters, vol. 25, p. 141-154,
2004I2
…
N filters
16/51
N filters
ICA FiltersICA Filters
17/51
Lecture 3: Outline
• Texture features
Lecture 3: Outline
• Texture features
− Statistical
− Spectral− Spectral
− Comparison
• Shape features
− Boundary based− Boundary based
− Region based
− Comparison− Comparison
• Fusion methods
18/51
Texture features: comparisonTexture features: comparison
In the context of In the context of image retrieval!
P. Howarth, S. Rüger. Robust texture features for still image retrieval.
19/51
P. Howarth, S. Rüger. Robust texture features for still image retrieval. In Proc. IEE Vis. Image Signal Processing, vol. 152, No. 6, December 2006
Texture features: comparison (2)Texture features: comparison (2)
Gabor filters v. s. ICA filtersGabor filters v. s. ICA filters
Image classification task:
� Collection of angiographic images
� ICA filters performs better by 13%� ICA filters performs better by 13%
� Brodatz texture collection
� ICA filters perform better by 4%
Snitkowska, E. Kasprzak, W. Independent Component Analysis of Textures in Angiography Images. Computational Imaging and Vision, vol. 32, pages 367-372, 2006.
20/51
Lecture 3: Outline
• Texture features
Lecture 3: Outline
• Texture features
− Statistical
− Spectral− Spectral
− Comparison
• Shape features
− Boundary based− Boundary based
− Region based
− Comparison− Comparison
• Fusion methods
21/51
Shape featuresShape features
Shape featuresShape features
Boundary-based methodsRegion-based methods
Perimeter
EccentricityCentroid Distance
Complex Coordinates
Gometrical Signatures
Geometrical
Moment invariants
Global
Curvature
Axes directionality
Complex Coordinates
Curvature signature
Turning Angle
Area
Compactness
Euler number
Moment invariants
Zernike moments
Pseudo Zernike moments
Grid method
Others
Fourier Descriptors
UNL-Fourier
NFD
Wavelet Descriptors
Signature descriptors
Triangulation
Medial Axis Transform
(Skeleton Transform)
Decompasition
Chain codes
Others
Wavelet Descriptors
B-Splines
( )
Spectral descriptors
22/51
Requirements to the shape featuresRequirements to the shape features
� Translation invariance
� Scale invariance� Scale invariance
� Rotational invariance� Rotational invariance
� Stability against small form changes
Low computation complexity� Low computation complexity
� Low comparison complexity� Low comparison complexity
23/51
Boundary-based featuresBoundary-based features
Shape features
Boundary-based methodsRegion based methodsRegion-based methods
Geometrical
Signatures GeometricalGlobal
Signature desciptors
Geometrical
Decomposition Others
Fourier Descriptors
NFD
...
Chain codes
24/51
Chain codesChain codesDirections for 4-connected and 8-connected chain codes:
A: 03001033332322121111 A: 03001033332322121111
B: 70016665533222
Example:Example:
0 0
0 03 1 3
31
10 0
1 6
62
2 7
Starting point invariance: minimal code
70016665533222 -> 00166655332227
22
22
3
3
31
1
1 6
5
53
3
270016665533222 -> 00166655332227
Rotation invariance: codes subtractiona) б) в)
A B
Rotation invariance: codes subtraction
00166655332227 -> 01500706070051
25/51
Fourier descriptorsFourier descriptors1. Signature calculation (2D -> 1D):
� Centroid – contour distance� Centroid – contour distance
� Complex coordinates: z(t) = x(t) + iy(t)
� ...
2. Perform the discrete Fourier transform, take coefficients (s(t) – signature):
∑−
−=1
/2)(1 N
Nntjetsu π∑=
−=0
/2)(1
t
Nntj
n etsN
u π
3. Normalization (NFD – Normalized Fourier Descriptors):
0
1
0
2
0
1,...,,
u
u
u
u
u
u N−
2
12
)(∑ −=cN
n
J
n
I ffd
4. Comparison:
26/51
0
)(∑=
−=n
JI ffd
Region-based featuresRegion-based features
Shape featuresShape features
Boundary-based methodsRegion-based methods
Geometrical
Signatures
Geometrical
Global
Signature representation
Moment invariants
Zernike moments
Pseudo Zernike moments
Grid method
DecompositionOthers
27/51
Grid-methodGrid-methodА: 001111000 011111111 111111111 111111111 111110111 0111000011А
Б: 001100000 011100000 111100000 111101111 111111110 001111000Б
Invariance:
Normalization by major axe:
� direction;
� scale;� scale;
� position.
28/51
Moment invariantsMoment invariantsThe moment of order (p+q) for a two-dimension continuous function:
∫∫= dxdyyxfyxm qp
pq ),(
Central moments for f(x,y) – discrete image:
mmqp ==−−=∑∑µ00
01
00
10 ,),,()()(m
my
m
mxyxfyyxx
x y
qp
pq ==−−=∑∑µ
Seven scale, translation and rotation invariant moments were derived based
Feature vector:
Seven scale, translation and rotation invariant moments were derived based
on central normalized moments of order p + q = 2; 3.
29/51
Lecture 3: Outline
• Texture features
Lecture 3: Outline
• Texture features
− Statistical
− Spectral− Spectral
− Comparison
• Shape features
− Boundary based− Boundary based
− Region based
− Comparison− Comparison
• Fusion methods
30/51
Shape features comparisonShape features comparison
Mehtre B. M., Kankanhalli M. S., Lee W. F. Shape measures for content based image retrieval: a comparison. Inf. Processing and Management, vol. 33, No. 3, pages 319-337, 1997.
31/51
comparison. Inf. Processing and Management, vol. 33, No. 3, pages 319-337, 1997.
Lecture 3: Outline
• Texture features
Lecture 3: Outline
• Texture features
− Statistical
− Spectral− Spectral
− Comparison
• Shape features
− Boundary based− Boundary based
− Region based
− Comparison− Comparison
• Fusion methods
32/51
Data fusion in CBIRData fusion in CBIR
color (2)annotations
texture shapecolor
� Combined search
fusion
� Combined search
(different features)
� Refine search results fusion� Refine search results
(different algorithms for the same
feature)
result� Supplement search results
(different datasets)
33/51
Fusion of retrieval result sets
Fusion of weighted lists with ranked elements:
(x11, r
11), (x
12, r
12), … , (x1
n, r1n)ω1
(x21, r
21), (x
22, r
22), … , (x2
k, r2n)ω2
…?
(xm1, r
m1), (x
m2, r
m2), … , (xm
l, rm
l)ωm
Existing approaches in text retrieval:Existing approaches in text retrieval:
� CombMax, CombMin, CombSum
� CombAVG� CombAVG
� CombMNZ = CombSUM * number of nonzero similarities
� ProbFuse
34/51
� ProbFuse
� HSC3D
Fusion function: propertiesFusion function: properties
1) Depend on both weight and rank1) Depend on both weight and rank
2) Symmetric
3) Monotony by weight and rank
},...,,max{},...,,min{)()()()0()()()( 2121 NN rrrrrrr
αααααα ≤≤
3) Monotony by weight and rank
4) MinMax condition /CombMin, CombMax, CombAVG/:
},...,,max{},...,,min{)()()()0()()()( 2121 NN
xxxxxxx rrrrrrrαααααα ≤≤
5) Additional property – “conic” property: non-linear dependency from 5) Additional property – “conic” property: non-linear dependency from
weight and rank; high weight, high rank – influence bigger to the
result than several inputs with low weight, low rank.
35/51
Weighted Total with Gravitation Function
CombAVG as a base, but use gravitation function instead of weight:CombAVG as a base, but use gravitation function instead of weight:
where
36/51
WTGF: some resultsWTGF: some results
• Experiments on search in semi annotated collections and • Experiments on search in semi annotated collections and of color and texture fusion (compare with CombMNZ)
• WTGF is good when:• WTGF is good when:
− There are a lot of viewpoints.
− Viewpoints are very different (different opinions regarding the rank − Viewpoints are very different (different opinions regarding the rank of the same element).
− Viewpoints have different reliability.
• CombMNZ is good when:
− Viewpoints have the same reliability.
Viewpoints have similar opinions.− Viewpoints have similar opinions.
Natalia Vassilieva, Alexander Dolnik, Ilya Markov. Image Retrieval. Combining multiple search
37/51
Natalia Vassilieva, Alexander Dolnik, Ilya Markov. Image Retrieval. Combining multiple search
methods’ results. In "Internet-mathematics" Collection, 46—55, 2007.
Adaptive merge: color and textureAdaptive merge: color and texture
Dist(I, Q) = α*C(I, Q) + (1 - α)*Т(I, Q),Dist(I, Q) = α*C(I, Q) + (1 - α)*Т(I, Q),
C(I, Q) – color distance between I and Q;
T(I, Q) – texture distance between I and Q; T(I, Q) – texture distance between I and Q;
0 ≤ α ≤ 1
Hypothesis:
Optimal α depends on features of query Q. It is possible Optimal α depends on features of query Q. It is possible to distinguish common features for images that have the same “best” α.
Ilya Markov, Natalia Vassilieva, Alexander Yaremchuk. Image retrieval. Optimal weights for color
and texture fusion based on query object. In Proceedings of the Ninth National Russian Research
38/51
and texture fusion based on query object. In Proceedings of the Ninth National Russian Research Conference RCDL'2007
Example: texture searchExample: texture search
39/51
Example: color searchExample: color search
40/51
Mixed metrics: semantic groupsMixed metrics: semantic groups
2a
1a
a41/51
3a
Experimental results 1Experimental results 1
• It is possible to select the best value of a• It is possible to select the best value of a
0,6
0,4
0,5
Pre
cis
ion
0,2
0,3
полнота Кластер 6
Кластер 7
Кластер 8
Pre
cis
ion
Cluster 7
Cluster 6
Cluster 8
0,1
0,2
0
0 0,25 0,5 0,75 1
коэффициент смешанной метрики (а)Value of a
42/51
Experimental results 2Experimental results 2
• Adaptive mixed-metrics increase precision • Adaptive mixed-metrics increase precision
43/51
Adaptive merge: color and colorAdaptive merge: color and color1,2
0,8
1
0,6
0,8
Pre
cis
ion
0,2
0,4
0
0 1 2 3 4 5 6 7 8 9 10
Top N
Moments, Clouds Moments, Fields
HSL, 6*2*2, Clouds HSL, 6*2*2, Fields
44/51
Adaptive merge: color and colorAdaptive merge: color and color
1,2
1
1,2
0,6
0,8
Pre
cis
ion
0,2
0,4
Pre
cis
ion
0
0 1 2 3 4 5 6 7 8 9 10
Top NTop N
Moments, People Moments, Bears
HSL, 6*2*2, People HSL, 6*2*2, Bears
45/51
Color fusionColor fusion
CombMNZ (Moments + HSL histogram)CombMNZ (Moments + HSL histogram)
0,8
0,6
0,7
0,4
0,5
Moments
HSL, 6*2*2
Color mixed
0,2
0,3
Color mixed
0
0,1
46/51
0
0 1 2 3 4 5 6 7 8 9 10
Ranked lists fusion: application area
� Search by textual query in semi annotated image � Search by textual query in semi annotated image collection
Textual query
…tr1by annotations
base
d
TextResult1, textrank1…tr2
conte
nt-
base
d
Result
TR2, tr2,
... …conte
nt
47/51
Retrieve by text: fusion resultsRetrieve by text: fusion results
∑
⋅=Moverlap
xRMxR
)()(
)0(
α∑M
i
xR i )()(α
∑
⋅=M
i
overlap
xR
xRMxR
i )(
)()(
)(
)0(
α
⋅= xNMxN
)()(
)0(
Size of input lists ∑=
M
i
overlap
xN
xN
i )(
)()(α
48/51
Lecture 3: ResumeLecture 3: Resume
• Texture features• Texture features
− Statistics (Haralik’s co-occurance matrices, Tamura features)
− Spectral features are more efficient (Gabor filters, ICA filters)− Spectral features are more efficient (Gabor filters, ICA filters)
• Shape features
− Boundary-based (Fourier descriptors)
− Region-based (Moment invariants)
• Fusion methods
Are very important− Are very important
− Need to choose based on a particular fusion task
49/51
Lecture 3: BibliographyLecture 3: Bibliography
• Haralick R. M., Shanmugam K., Dienstein I. Textural features for image classification. In IEEE Transactions on Systems, Man and Cybernetics, classification. In IEEE Transactions on Systems, Man and Cybernetics, vol. 3(6), pp. 610 – 621, Nov. 1973.
• Tamura H., Mori S., Yamawaki T. Textural features corresponding to visual perception. In IEEE Transactions on Systems, Man and visual perception. In IEEE Transactions on Systems, Man and Cybernetics, vol. 8, pp. 460 – 472, 1978.
• Tuceryan M., Jain A. K. Texture analysis. The Handbook of Pattern Recognition and Computer Vision (2nd Edition), by C. H. Chen, L. F. Pau, Recognition and Computer Vision (2nd Edition), by C. H. Chen, L. F. Pau, P. S. P. Wang (eds.), pp. 207-248, World Scientific Publishing Co., 1998.
• Tuceryan M., Jain A. Texture segmentation using Voronoi polygons. In IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 12, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 12, No 2, pp. 211 – 216, February 1990.
• Walker R., Jackway P., Longstaff I. D. Improving co-occurrence matrix feature discrimination. In Proc. of DICTA’95, The 3rd Conference on feature discrimination. In Proc. of DICTA’95, The 3rd Conference on Digital Image Computing: Techniques and Applications, pp. 643 – 648, 6-8 December, 1995.
50/51
Lecture 3: BibliographyLecture 3: Bibliography• Li B., Ma S. D. On the relation between region and contour representation. In Proc.
of the IEEE International Conference on Pattern Recognition, vol. 1, pp. 352 – 355, of the IEEE International Conference on Pattern Recognition, vol. 1, pp. 352 – 355, 1994.
• Lin T.-W., Chou Y.-F. A Comparative Study of Zernike Moments for Image Retrieval. In Proc. of 16th IPPR Conference on Computer Vision, Graphics and Image Processing (CVGIP 2003), pp. 621 – 629, 2003.Image Processing (CVGIP 2003), pp. 621 – 629, 2003.
• Loncaric S. A survey of shape analysis techniques. In Pattern Recognition, vol. 31(8), pp. 983 – 1001, 1998.
• Luren Y., Fritz A. Fast computation of invariant geometric moments: A new method • Luren Y., Fritz A. Fast computation of invariant geometric moments: A new method giving correct results. In Proc. of IEEE International Conference on Image Processing, 1994.
• Zakaria M. F., Vroomen L. J., Zsombor-Murray P. J. A., van Kessel J. M. H. M. Fast algorithm for the computation of moment invariants. In Pattern Recognition, vol.
• Zakaria M. F., Vroomen L. J., Zsombor-Murray P. J. A., van Kessel J. M. H. M. Fast algorithm for the computation of moment invariants. In Pattern Recognition, vol. 20(6), pp. 639 – 643, 1987.
• Zernike polynomials. Wikipedia, the free encyclopedia. http://en.wikipedia.org/wiki/Zernike_polynomialshttp://en.wikipedia.org/wiki/Zernike_polynomials
• Zhang D., Lu G. Shape-based image retrieval using generic Fourier descriptor. In Signal Processing: Image Communication, vol. 17, pp. 825 – 848, 2002.
• Zhang D., Lu G. A Comparative Study on Shape Retrieval Using Fourier
51/51
• Zhang D., Lu G. A Comparative Study on Shape Retrieval Using Fourier Descriptors with Different Shape Signatures. In Proc. of the International Conference on Multimedia, 2001.