50
1 International Workshop on Com puter Vision April 26-30, 200 4 Tehran,Iran Singular Value Decompositions with Singular Value Decompositions with applications to applications to 1. Texture differentiation 2. Detection of an extraneous object in a texture environment 3. Segmentation of images 4. Locating eyes in facial images Alireza Tavakoli Targhi Alireza Tavakoli Targhi [email protected] Institute for Studies in Theoretical Physics and Mathematics (IPM), Iran and Royal Institute of Technology (KTH), Sweden

1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

  • View
    218

  • Download
    1

Embed Size (px)

Citation preview

Page 1: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

1International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Singular Value Decompositions with Singular Value Decompositions with applications toapplications to

1. Texture differentiation2. Detection of an extraneous object in a texture

environment3. Segmentation of images

4. Locating eyes in facial images

Alireza Tavakoli TarghiAlireza Tavakoli [email protected]

Institute for Studies in Theoretical Physics and Mathematics (IPM), Iran

and

Royal Institute of Technology (KTH), Sweden

Page 2: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

2International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

IntroductionIntroductionWe propose new measures for texture classification based on a local version of Singular Value Decomposition (SVD) .

The proposed measures classify textures by their roughness and structure.

Experimental results show that these measures are suitable for texture clustering and image segmentation and they are robust to changes in local lighting, orientation etc.

Page 3: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

3International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Singular Value DecompositionSingular Value Decomposition

A=UA=U11*D*U*D*U22

UUi i Orthogonal Matrix , D Diagonal Matrix with Orthogonal Matrix , D Diagonal Matrix with Diagonal Entries in descending order: Diagonal Entries in descending order:

dd1 1 >d>d2 2 > …>=0> …>=0

Overview of SVDOverview of SVD

Page 4: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

4International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

To find the SVD curve of an arbitrary row of To find the SVD curve of an arbitrary row of the texture:the texture:

1.1. Scan a row with overlapping w*w Scan a row with overlapping w*w windows Wwindows Wa a , a=1,2,3,….. ; w ≈ 32 , a=1,2,3,….. ; w ≈ 32

2.2. Calculate the SVD Decomposition Calculate the SVD Decomposition

3.3. WWaa =U =U1,a1,a*D*Daa *U *U2,a2,a

4.4. As windows scan the image we As windows scan the image we obtain w curves (obtain w curves (ii) ) corresponding to diagonal entries corresponding to diagonal entries dd1,a 1,a >d>d2,a2,a>….>d>….>dw,aw,a

SVD CurvesSVD Curves

Page 5: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

5International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Classification Classification

Diagonal entries reflect image Diagonal entries reflect image characteristics .characteristics .

Page 6: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

6International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

The sizes of the first few The sizes of the first few coefficients are considerably coefficients are considerably

larger than the remaininglarger than the remaining

Page 7: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

7International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

SVD Curve Classifiers

We introduce two measures, obtained from SVD curves, which capture some of the perceptual and conceptual features in an image.

SVD Curve mean classifier SVD Curve distance classifier

Page 8: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

8International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

SVD Curve Mean Classifier Our experiments show that the smaller coefficients d a,j ,

i.e., da,j with 23<j<32, are more representative of the structure of the texture and less dependent on inessential features.

In practice, we set l = 22 and k = 10.

Page 9: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

9International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Page 10: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

10International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

SVD-Distance Classifier

Page 11: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

11International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Detection of Extraneous Object:Detection of Extraneous Object:

Page 12: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

12International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Even Small objects Even Small objects

Page 13: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

13International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Interest Point Detector

Page 14: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

14International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

The technique identifies The technique identifies the bug even the location the bug even the location

of its legsof its legs

Page 15: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

15International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Added two coinsAdded two coins

Page 16: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

16International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

No DifferentiationNo Differentiation

Page 17: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

17International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

SVD TransformSVD Transform

We slide a w*w window across the We slide a w*w window across the image. image.

We identify each window by its upper We identify each window by its upper left corner coordinates (x,y).left corner coordinates (x,y).

Let F be a function of w variables. Let F be a function of w variables. The SVD surface is the graph of the The SVD surface is the graph of the

function (x,y) → the value of F on Dfunction (x,y) → the value of F on D(x,y )(x,y ).. The SVD transform is the representation The SVD transform is the representation

of the of the SVD surface as a 2D image. of the of the SVD surface as a 2D image.

Page 18: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

18International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

SVD TransformSVD Transform

Page 19: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

19International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Original Image Original Image SVD SVD TransformTransform

SVD surface

Page 20: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

20International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 21: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

21International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 22: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

22International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 23: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

23International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 24: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

24International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD Segmentation via SVD TransformTransform

Page 25: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

25International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 26: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

26International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Application of Feature Vector IIApplication of Feature Vector IISVD Transform, Segmentation SVD Transform, Segmentation

Page 27: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

27International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 28: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

28International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Sensitivity to TextureSensitivity to Texture

Page 29: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

29International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Sensitivity to TextureSensitivity to Texture

The Berkeley Segmentation Data Base. Computer Vision Group.

Page 30: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

30International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD TransformSegmentation via SVD Transform

Page 31: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

31International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation via SVD Segmentation via SVD transformtransform

The Berkeley Segmentation Data Base. Computer Vision Group.

Page 32: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

32International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

The Berkeley Segmentation Data Base. Computer Vision Group.

Page 33: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

33International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Effect of change of a Effect of change of a parameterparameter

The Berkeley Segmentation Data Base. Computer Vision Group.

Page 34: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

34International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Effect of change of a Effect of change of a parameterparameter

Page 35: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

35International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation (cont.)Segmentation (cont.)

Page 36: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

36International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation (cont.)Segmentation (cont.)

Page 37: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

37International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation (cont.)Segmentation (cont.)

Page 38: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

38International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Segmentation (cont.)Segmentation (cont.)

Page 39: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

39International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Images show the effect of Images show the effect of substituting the diagonal substituting the diagonal part or the orthogonal part or the orthogonal parts from the SVD parts from the SVD decomposition of an decomposition of an image into that of another image into that of another image.image.

ws=5ws=5

Understanding Understanding SVDSVD

Page 40: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

40International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Images show the effect of Images show the effect of substituting the diagonal substituting the diagonal part or the orthogonal part or the orthogonal parts from the SVD parts from the SVD decomposition of Lena decomposition of Lena into that of a randomly into that of a randomly generated image.generated image.

ws=5ws=5

Understanding Understanding SVDSVD

Page 41: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

41International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Understanding Understanding SVDSVD

Images show the effect of Images show the effect of substituting the diagonal substituting the diagonal part or the orthogonal part or the orthogonal parts from the SVD parts from the SVD decomposition of an decomposition of an image into that of another image into that of another image.image.

ws=32ws=32

Page 42: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

42International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Images show the effect of Images show the effect of substituting the diagonal substituting the diagonal part or the orthogonal part or the orthogonal parts from the SVD parts from the SVD decomposition of Lena decomposition of Lena into that of a randomly into that of a randomly generated image.generated image.

ws=32ws=32

Understanding Understanding SVDSVD

Page 43: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

43International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Invariance relative to Invariance relative to inversioninversion

These images are negatives of each other.

Their SVD surfaces are identical.

Page 44: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

44International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

An SVD SurfaceAn SVD Surface

Page 45: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

45International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Detecting cracks and Detecting cracks and defectsdefects

Page 46: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

46International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Detecting cracks and Detecting cracks and defectsdefects

Page 47: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

47International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Detecting cracks and Detecting cracks and defectsdefects

Page 48: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

48International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Detecting cracks and defectsDetecting cracks and defects

Page 49: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

49International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Theoretical FrameworkTheoretical Framework We do not have a definitive answer why SVD works to the extent We do not have a definitive answer why SVD works to the extent

that it does. However, on the basis of our experimentations we that it does. However, on the basis of our experimentations we can assert the following:can assert the following:

1. The diagonal entries of SVD capture some features of an image 1. The diagonal entries of SVD capture some features of an image which are not encoded by the correlations of nearby pixels in an which are not encoded by the correlations of nearby pixels in an image.image.

2. This may explain why images constructed on the basis local 2. This may explain why images constructed on the basis local correlations virtually never exhibit features similar to ones in real correlations virtually never exhibit features similar to ones in real images. Images constructed on the basis of Markov random images. Images constructed on the basis of Markov random fields or similar procedures generally look very random.fields or similar procedures generally look very random.

3. As the size of the sliding window increases the importance of 3. As the size of the sliding window increases the importance of the orthogonal parts of SVD decomposition relative to the the orthogonal parts of SVD decomposition relative to the diagonal part increases.diagonal part increases.

Page 50: 1 International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran Singular Value Decompositions with applications to Singular Value Decompositions

50International Workshop on Computer Vision April 26-30, 2004 Tehran,Iran

Work in ProgressWork in Progress We are in the process of applying SVD for We are in the process of applying SVD for

material/surface classification.material/surface classification. The SVD transform can be applied to the test The SVD transform can be applied to the test

case of differentiating between a dog and a cat. case of differentiating between a dog and a cat. The results are preliminary and require further The results are preliminary and require further tests. tests.

SVD transforms are also being tested on movie SVD transforms are also being tested on movie images.images.

Our methods are being tested on medical Our methods are being tested on medical images by IRMA (Image Retrieval in Medical images by IRMA (Image Retrieval in Medical Applications) of Institut fuer Medizinische Applications) of Institut fuer Medizinische Informatik Universitaetklinikum der RWTH in Informatik Universitaetklinikum der RWTH in Aachen, Germany.Aachen, Germany.