University of Groningen Institute of Mathematics and Computing Science

57
University of Groningen Institute of Mathematics and Computing Science Universitá degli Studi di Roma Tre Dipartimento di Elettronica Applicata Well Posed non-Iterative Well Posed non-Iterative Edge and Corner Preserving Smoothing Edge and Corner Preserving Smoothing For Artistic Imaging For Artistic Imaging Giuseppe Pápari, Nicolai Petkov, Patrizio Campisi

description

Well Posed non-Iterative Edge and Corner Preserving Smoothing For Artistic Imaging. Giuseppe Pápari, Nicolai Petkov, Patrizio Campisi. University of Groningen Institute of Mathematics and Computing Science. Universit á degli Studi di Roma Tre Dipartimento di Elettronica Applicata. - PowerPoint PPT Presentation

Transcript of University of Groningen Institute of Mathematics and Computing Science

Page 1: University of Groningen Institute of Mathematics and Computing Science

University of Groningen Institute of Mathematics and Computing Science

Universitá degli Studi di Roma TreDipartimento di Elettronica Applicata

Well Posed non-IterativeWell Posed non-Iterative

Edge and Corner Preserving SmoothingEdge and Corner Preserving Smoothing

For Artistic ImagingFor Artistic Imaging

Giuseppe Pápari, Nicolai Petkov, Patrizio Campisi

Page 2: University of Groningen Institute of Mathematics and Computing Science

Photographical image

Page 3: University of Groningen Institute of Mathematics and Computing Science

Output of the proposed operator

Page 4: University of Groningen Institute of Mathematics and Computing Science

Smoothing out texture while preserving edges

Input image Gaussian smoothing Proposed operator

CoContentsntents•Kuwahara Filter and Generalizations•Limitations•Proposed Operator•Results and Comparison•Discussion

Page 5: University of Groningen Institute of Mathematics and Computing Science

Kuwahara Filter and GeneralizationsKuwahara Filter and Generalizations

•Four local averages: ,

1, , d di

iQ x y

m x y IQ

22

,

1, , , d di

i iQ x y

s x y I m x yQ

•Four local standard deviations:

Kuwahara outputFor each pixel, value of mi that corresponds to the minimum standard deviation

Q1 Q2

Q3 Q4 2a

x

y

Generic pixel of the input image

Page 6: University of Groningen Institute of Mathematics and Computing Science

Kuwahara Filter and GeneralizationsKuwahara Filter and Generalizations

Edge

Only the most homogeneous region is taken into account.

No smoothing across the edge

(x,y) = 1

Central pixel on the white side of the edge

(x,y) = 0

Central pixel on the black side of the edge

Page 7: University of Groningen Institute of Mathematics and Computing Science

Kuwahara Filter and GeneralizationsKuwahara Filter and Generalizations

• Local averaging Smoothing

• Flipping due to Minimum Variance Criterion Edge Preserving

Page 8: University of Groningen Institute of Mathematics and Computing Science

Kuwahara Filter and GeneralizationsKuwahara Filter and GeneralizationsAn example

Input image

Kuwahara output

Artifactson texture

Page 9: University of Groningen Institute of Mathematics and Computing Science

Kuwahara Filter and GeneralizationsKuwahara Filter and GeneralizationsGeneralizations

•Number and shape of the sub-regions

»Pentagons, hexagons, circles»Overlapping

•Weighted local averages (reducing the Gibbs phenomenon)

»Gaussian-Kuwahara

• New class of filters (Value and criterion filter structure)

»N local averages and local standard deviations (computed as convolutions)

2 2 2,i i i i im I w s I m w

»Criterion: minimum standard deviation

•Connections with the PDEs theory and morphological analysis

Page 10: University of Groningen Institute of Mathematics and Computing Science

•Kuwahara Filter and Generalizations

•LimitationsLimitations•Proposed Operator

•Results and Comparison

•Discussion

Page 11: University of Groningen Institute of Mathematics and Computing Science

LimitationsLimitations• Artifacts

(partially eliminable with weighted averages)

• Not mathematically well defined

Equal standard deviations si • Devastating instability in presence of noise

Page 12: University of Groningen Institute of Mathematics and Computing Science

LimitationsLimitationsSimple one-dimensional example

Input signal I(t)•I(t) =

kt

Local averagesI

ttT t+T

21 d

t

t T

k ktT

kT

Negative offset

• 1D Kuwahara filtering Two sub-windows w1 and w2 t

w2

t*

I(t)

w1

Page 13: University of Groningen Institute of Mathematics and Computing Science

LimitationsLimitationsSimple one-dimensional example

Input signal I(t)•I(t) =

kt

Local averagesI

ttT t+T

21 d

t

t T

k ktT

kT

Negative offset

21 dt T

t

kt Tk kT

Positive offset

• 1D Kuwahara filtering Two sub-windows w1 and w2 t

w2

t*

I(t)

w1

Page 14: University of Groningen Institute of Mathematics and Computing Science

•I(t) = kt

Local standard deviations

2 2

1 2 4k Ts t s t

•Equal standard deviations

I

ttT t+T

LimitationsLimitationsSimple one-dimensional example

Local averages m1(t), m2(t)

Input signal I(t) Local std. dev. s1(t), s2(t)

• 1D Kuwahara filtering Two sub-windows w1 and w2 t

w2

t*

I(t)

w1

Page 15: University of Groningen Institute of Mathematics and Computing Science

LimitationsLimitations

Input image Kuwahara filtering Proposed approach

Synthetic two-dimensional example

Page 16: University of Groningen Institute of Mathematics and Computing Science

Kuwahara

LimitationsLimitationsNatural image example

Input image Gauss-Kuwahara

Shadowed area

Depleted edge

Our approach

Page 17: University of Groningen Institute of Mathematics and Computing Science

LimitationsLimitations

Ill-posedness of the minimum variance criterion.

Devastating effects in presence of noisy shadowed areas.

We propose

• Different weighting windows wi

• A different selection criterion instead of the minimum standard deviation

Page 18: University of Groningen Institute of Mathematics and Computing Science

•Kuwahara Filter and Generalizations

•Limitations

•Proposed OperatorProposed Operator•Results and Comparison

•Discussion

Page 19: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed Operator

• Gaussian mask divided in N sectors N weighting windows

• N local averages and local standard deviations computed as convolutions

2 2 2,i i i i im I w s I w m

Weighting windows

2 2

222

1, ,2

x y

iiw x y g x y e

Page 20: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed OperatorSelection criterion

• q Only the minimum si survives Criterion and value

• Output:

ii

qi

qi

is

ms

» Weighted average of mi

» Weights equal to proportional to (si)q (q is a parameter)

Normalization

• High variance small coefficient (si)q

No undetermination in case of equal standard deviations!

Page 21: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed OperatorParticular cases

• Equal standard deviations: s1 = s2 = … = sN

1i

im

N

Gaussian smoothing

I g

• One standard deviation is equal to zero: sk = 0

km

• Several values of si are equal to zero

= Arithmetic mean of the corresponding values of mi.

Page 22: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed Operator

• EdgeHalf of the sectors have si = 0. The other ones are not considered

An example

• Edgeless areas:

All std. dev. similar Gaussian smoothing (no Gibbs phenomenon)

• Corner preservation

Automatic selection of the prominent sectors

Page 23: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed OperatorColor images

• 3 sets of local averages and local standard deviations, one for each color component

with

qi i

iqi

i

s

s

m

• Same combination rule

3 21 2 3

1, , ,

T ci i i i i i

cm m m s s

m

Not equivalentto apply the operator

to each color component separately

Page 24: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed OperatorIndependence on the color space

Input image RGB YCrCb L*a*b*

Page 25: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed OperatorWhy independence?

qi i

iqi

i

s

s

m

Linear transform. independentNonlinear transf. almost independent for homogeneous regions

• Local averages

Page 26: University of Groningen Institute of Mathematics and Computing Science

Proposed OperatorProposed OperatorWhy independence?

Linear transform. independentNonlinear transf. almost independent for homogeneous regions

qi i

iqi

i

s

s

m

• Local averages

Low for homogeneous regions.The degree of homogeneity of a region does not depend on the color space.

• Local standard deviations

Page 27: University of Groningen Institute of Mathematics and Computing Science

•Kuwahara Filter and Generalizations

•Limitations

•Proposed Operator

•Results and ComparisonResults and Comparison•Discussion

Page 28: University of Groningen Institute of Mathematics and Computing Science

Results and comparisonResults and comparisonExisting algorithm for comparison

• Kuwahara filter and generalizations

• Bilateral filtering

• Morphological filters

• Median filters

Page 29: University of Groningen Institute of Mathematics and Computing Science

Input image

Page 30: University of Groningen Institute of Mathematics and Computing Science

Proposed approach

Page 31: University of Groningen Institute of Mathematics and Computing Science

Gauss-Kuwahara filter

Page 32: University of Groningen Institute of Mathematics and Computing Science

Input image (blurred)

Page 33: University of Groningen Institute of Mathematics and Computing Science

Proposed approach (deblurred)

Page 34: University of Groningen Institute of Mathematics and Computing Science

Bilateral filtering (not deblurred)

Page 35: University of Groningen Institute of Mathematics and Computing Science

Input image

Page 36: University of Groningen Institute of Mathematics and Computing Science

Proposed approach

Page 37: University of Groningen Institute of Mathematics and Computing Science

Morphological closing (Struct. elem.: Disk of radius 5px)

Page 38: University of Groningen Institute of Mathematics and Computing Science

Morphological area open-closing

Page 39: University of Groningen Institute of Mathematics and Computing Science

Input image

Page 40: University of Groningen Institute of Mathematics and Computing Science

Morphological area open-closing

Page 41: University of Groningen Institute of Mathematics and Computing Science

Proposed approach

Page 42: University of Groningen Institute of Mathematics and Computing Science

Input image

Page 43: University of Groningen Institute of Mathematics and Computing Science

Proposed approach

Page 44: University of Groningen Institute of Mathematics and Computing Science

Kuwahara Filter

Page 45: University of Groningen Institute of Mathematics and Computing Science

Morphological area open-closing

Page 46: University of Groningen Institute of Mathematics and Computing Science

Input image

Page 47: University of Groningen Institute of Mathematics and Computing Science

Proposed approach

Page 48: University of Groningen Institute of Mathematics and Computing Science

Bilateral Filtering

Page 49: University of Groningen Institute of Mathematics and Computing Science

Input image

Page 50: University of Groningen Institute of Mathematics and Computing Science

proposed aproach

Page 51: University of Groningen Institute of Mathematics and Computing Science

55 median filter

Page 52: University of Groningen Institute of Mathematics and Computing Science

Results and comparisonResults and comparisonLarger set of results and Matlab implementation available at

http://www.cs.rug.nl/~imaging/artisticsmoothing

• Graphical interface

Page 53: University of Groningen Institute of Mathematics and Computing Science

•Kuwahara Filter and Generalizations

•Limitations

•Proposed Operator

•Results and Comparison

•DiscussionDiscussion

Page 54: University of Groningen Institute of Mathematics and Computing Science

DiscussionDiscussion• Edge/corner preserving smoothing

• Undetermination for equal standard deviation

» Instability in presence of noise

» Discontinuities in presence of shadowed areas

• Criterion and value filter structure» Local averaging Smoothing

» Minimum variance criterion Edge preserving

Page 55: University of Groningen Institute of Mathematics and Computing Science

DiscussionDiscussion• Proposed approach

» Different windows

» Different criterion

Mathematically well defined operator

Adaptive choice of the most appropriate sub-regions.

Our approachGauss-KuwaharaKuwahara

Page 56: University of Groningen Institute of Mathematics and Computing Science

DiscussionDiscussion• Limitations

» Lines are thinned

» Small objects are not preserved

Page 57: University of Groningen Institute of Mathematics and Computing Science

ReferencesReferences•G. Papari, N. Petkov, P. CampisiArtistic Edge and Corned Preserving SmoothingTo appear on IEEE Transactions on Image Processing, 2007

•G. Papari, N. Petkov, P. CampisiEdge and Corned Preserving Smoothing for Artistic ImagingProceedings SPIE 2007 Image Processing: Algorithms and Systems, San Jose, CA