image blur estimation based on the average cone of ratio in ...sanja/papers/blurestimation/...select...

Ghent University-TELIN/IPI-IBBT Image Processing and Interpretation Group

Image Blur Estimation Based on the Average Cone of Ratio in the Wavelet Domain

L. Platiša, N. Lukid, A. Pižurica, E. Vansteenkiste, W. Philips

7/10/2011 2 Blur Identification

• Introduction

• The proposed blur metric

• Experimental results

• Real-time implementation

• Conclusions

• Future research

Contents


• Blur is usually observed as a loss of image sharpness which corresponds to smoothening of the edges.

• Why blur evaluation?

– assessment of image quality

– image restoration

• Standard observation model

Introduction

),(),(*),(),( yxnyxfyxhyxg

g(x,y) - degraded image h(x,y) - blurring function (PSF)

f(x,y) - ideal image n(x,y) - noise


... characterize the regularity of the signals

Using wavelet transform to estimate local Lipschitz exponent, α

• [Jaffard, 1991]

• [Mallat and Zhong, 1992]

• [Malfait and Roose, 1997]

• [Hsung et al., 1999]

• [Pizurica et al., 2002]

Lipschitz exponents

Pointwise Lipschitz exponents

[courtesy of A. Pizurica]

Pointwise smoothness, two-microlocalization and wavelet coefficients, Jaffard, 1991 Characterization of signals from multiscale edges, Mallat and Zhong, 1992 Wavelet-based image denoising using a markov random field a priori model, Malfait and Roose, 1997 Denoising by singularity detection, Hsung et al., 1999 A joint inter- and intrascale statistical model for wavelet based Bayesian image denoising, Pizurica et al., 2002


Average cone ratio (ACR)

describes the “collective” evolution of the wavelet coefficients inside a cone of influence centred at the spatial position l.

• Properties of ACR

– good estimate of the local Lipschitz exponent α – significant robustness to noise

),(

,,

1

,

,1

2, ,1

logljCm

mjlj

k

nj lj

lj

lkn II

I

nk

Average cone ratio, ACR [Pizurica et al. 2002]

Average point ratio (APR) describes the evolution of the individual wavelet coefficients at the spatial position l.


• ACR is able to better separate the noise from the useful edges

Average Point Ratio vs. Average Cone Ratio

Conditional densities of APR and ACR, computed from scales 21 − 24.

The standard deviation of added noise is σ = 25. [courtesy of A. Pizurica]

ACR APR

noisy edges

noise noisy edges

noise


ACR metric. Probability density function

-1 0 1 2 3 40

0.2

0.4

0.6

0.8

1.0

ACR 2-4

PD

F

Degradation free

Noisy (var=25)

Blurred (r=3)

Blurred (r=3) &Noisy (var=25)

(σ2=25)

(σ2=25)


0.5 1 1.5 20.5

0.6

0.7

0.8

0.9

ACR 2-4P

DF

BL3

BL2

BL1

Ref.

0.5 1 1.5 20.5

0.6

0.7

0.8

0.9

ACR 2-4

PD

F

Ref.

BL3

BL1

BL2

0.5 1 1.5 20.5

0.6

0.7

0.8

0.9

1

1.1

ACR 2-4

PD

F

BL1

Ref.

BL3

BL2

Low sensitivity to noise

Noisy image (σ2=10)

Noise-free image

Gaussian blur Motion blur Circular blur


New blur metric: CogACR

Edge positions

CogACR

Reference

image

Degraded

image

0 2 40

0.2

0.4

0.6

0.8

ACR 2-4

PD

F

the center of gravity

of the ACR histogram

Wavelet transform

. . . . . . Scenario 1.

Reference image available

Scenario 2.

Ref. image NOT available

Image blur estimation based on the average cone of

ratio in the wavelet domain, Platisa et al., 2009


Blur model. Circular blur (pillbox)

Example test image

.


Example test images [130 real images]


Space of values for CogACR

0 2 4 6 8 100.5

0.75

1

1.25

1.5

1.75

2

2.25C

og

AC

R24

Level of blur

The results on 130 test images

Consistent behaviour for the whole range of tested blur


Extracting valid edges for ACR analysis

a) THRESHOLDING

WAVELET COEFFICIENTS

b) THRESHOLDING

INTERSCALE PRODUCTS

V H

Noisy σ2=10

Noise-free

Noisy σ2=25

Edge detection

based on:

V H


Tp=0.10

Selecting the threshold

V H

Tp=0.01

Tp=0.05

Tp=0.10

Tp=0.15

Tp=0.20

Tp=0.10

0 1 2 3 4 5 6 7 8 9 10

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

Level of blur

Cog

AC

R2

4

Tp=0.01

Tp=0.05

Tp=0.10

Tp=0.15

Tp=0.20

Tp - threshold percent coefficient



H V Tp=0.10 Tp=0.01

Tp=0.05

Tp=0.10

Tp=0.15

Tp=0.20

Tp=0.10

0 1 2 3 4 5 6 7 8 9 10

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

Level of blur

Cog

AC

R2

4

Tp=0.01

Tp=0.05

Tp=0.10

Tp=0.15

Tp=0.20



V H Tp=0.01 Tp=0.01

Tp=0.05

Tp=0.10

Tp=0.15

Tp=0.20

Tp=0.01

0 1 2 3 4 5 6 7 8 9 10

0.6

0.8

1

1.2

1.4

1.6

1.8

2

Level of blur

Cog

AC

R2

4

Tp=0.01

Tp=0.05

Tp=0.10

Tp=0.15

Tp=0.20


Texture of the content

♦ Image class A

x Image class B


Texture of the content

r=0 r=1 r=2 r=3 r=4 r=6 r=8


0 2 4 6 8 100.5

0.75

1

1.25

1.5

1.75

2

Level of blur

Cog

AC

R24


Reference image NOT available

No-Reference metric

◊ Ref. image available

x Ref. image NOT available


• based on ACR histogram similarity, using histogram intersection metric

• preliminary results

No-Reference image classification

0 2 4 6 8 100.5

0.75

1

1.25

1.5

1.75

2

Level of blur

Cog

AC

R2

4


• based on ACR histogram similarity, using histogram intersection metric

step1. compute the intersected section of the input image with the template histograms

step2. select the best matching template histogram

step3. calculate CogACR of the input image

step4. based on CogACR curve of the selected image (template histogram), estimate the level of blur in the input image




0 2 4 6 8 100.5

0.75

1

1.25

1.5

1.75

2

Level of blur

Cog

AC

R2

4

selected template

input image

Input image


0 1 2 3 4 5 6 7 8 9 10

0.6

0.8

1.0

1.2

1.4

1.6

1.8

2.0

2.2

2.4

2.6

2.8

3.0

Level of blur

Metr

ic o

f blu

r

Noisy images. Kurtosis vs CogACR

Kurtosis

CogACR

--- noise free

--- noisy, var=10

--- noisy, var=25

Li et al. Blur identification based on kurtosis minimization. ICIP 2005.


CogACR compared to other blur metrics [Image 1]

[1] Hu and Haan. Low Cost Robust Blur Estimator. ICIP 2006.

[2] Tong et al. Blur detection for digital images using wavelet transform. ICME 2004.

0 1 2 3 4 50

0.2

0.4

0.6

0.8

1.0

1.2

1.4

Radius of Gaussian blur

Est

imat

ed

lev

el o

f blu

r

Hu`s method

CogACR

Tong`s method

noise var=0

noise var=25


CogACR compared to other blur metrics [Image 2]



0 1 2 3 4 50

0.2

0.4

0.6

0.8

1.0

1.2

1.4

Radius of Gaussian blur

Est

imat

ed

lev

el o

f blu

r

CogACR

Tong`s method

Hu`s method

noise var=0

noise var=25


CogACR. Pros & Cons



Hu and Haan 2006 [1] Tong et al. 2004 [2] CogACR

Range of distinguishable levels of Gaussian blur

R~(0..1.5) R~(0..2) R~(0..3.5)

Sensitivity to noise

High High Low

Computational complexity

Low High High

Image content dependency Yes Yes Yes


Real-time implementation

Cell Broadband Engine Architecture (CBEA): microprocessors dedicated to distributed data processing

• Computationaly most expensive part is wavelet transform (~38% proc. time)

• Real-time performance requires multiple SPE cores (3 out of 8)

PPE

Processor

SPE Processor

SPE Program

SPE Processor

SPE Program

SPE Processor

SPE Program

SPE Program

Referent

image

Blurred

image

EdgesWavelet t.

ACR CogACR


CBEA performance measurements

0

20

40

60

80

100

120

140

160

Frame rate on single SPE core

TongHuCogACR

(a)

Tong

Hu

CogACRSingle SPE core

Two SPE cores

Three SPE cores

0

20

40

60

80

100

120

140

160

Frame rate of CogACR on multiple SPE cores

Single SPE core

Two SPE cores

Three SPE cores

(b) Frame rate of CogACR on multiple

SPE coresFrame rate on single SPE core


New metric of blur, CogACR

• Sensitive to a wide range of blur levels

• Nearly insensitive to noise

• Valid in reference- and no-reference scenario

• Real-time performance could be achieved

• Current investigation

– automated content based image classification

– automated selection of Tp

Conclusions


Consider using orientation sensitive wavelet-like transforms (shearlets, curvelets, ...)

Directions for further investigation

-150

-100

-50

0

50

100

150

Work with blocks Use only the strongest edge in the image

Work with the region of interest as seen by humans

Does where you gaze on an image affect your

perception of quality? Applying visual attention to

image quality metric, Ninassi et al., 2007


Contact

LJILJANA PLATISA

Ghent University - TELIN - IPI - IBBT

[email protected]

http://telin.UGent.be/

mailto:[email protected]

image blur estimation based on the average cone of ratio in ...sanja/papers/blurestimation/...select...

Documents