patch-based image deconvolution via joint modeling of sparse priors

22
Patch-based Image Deconvolution via Joint Modeling of Sparse Priors Chao Jia and Brian L. Evans The University of Texas at Austin 12 Sep 2011 1

Upload: amy

Post on 22-Feb-2016

96 views

Category:

Documents


0 download

DESCRIPTION

Patch-based Image Deconvolution via Joint Modeling of Sparse Priors . Chao Jia and Brian L. Evans The University of Texas at Austin 12 Sep 2011. Non-blind Image Deconvolution. Reconstruct natural image from blurred version Camera shake; astronomy; biomedical image reconstruction - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Patch-based Image Deconvolution via Joint

Modeling of Sparse Priors

Chao Jia and Brian L. EvansThe University of Texas at Austin

12 Sep 2011

1

Page 2: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Non-blind Image Deconvolution Reconstruct natural image from blurred version

Camera shake; astronomy; biomedical image reconstruction

2D convolution matrix H and Gaussian additive noise vector n

Maximum a-posteriori (MAP) estimation for vector X

Prior model for p(X) for natural images? [Elad 2007] Optimization method?

2

Page 3: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Analysis-based modeling [Krishnan 2009]

Prior based on hyper-Laplacian distribution of the spatial derivative of natural images

Linear filtering to compute spatial derivative Fit (0.5-0.8) and (normalization factor) to empirical data

3

Page 4: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Patch-based modeling Sparse coding of patches

Spatial receptive fields of visual cortex [Olshausen 1997] For 10 10 patches Learn an overcomplete dictionary from natural images.

Application in image restoration Denoising, superresolution

[Yang 2010] Localized algorithm: patches can

overlap Use this model in deconvolution?

[Lee 2007]

4

Page 5: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Prior model in natural images From local to global

Slow convergence (EM Algorithm)

Patches should not overlap (Why?) boundary artifacts

5

Page 6: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Joint modeling Take advantage of patch-based sparse representation

while resolving the problems in? Combine analysis-based prior and synthesis-based prior

Sparse spatial gradient

Patch-based sparse coding

Accelerate convergence

Keep consistency on the boundary of adjacent patches

Keep details and textures

6

Page 7: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Joint modeling Discard the generative model

Prior probability

After training, we fix the parameters for all images

sparsity of representation

coefficients

compatibility term

sparsity of gradients

7

Page 8: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

MAP estimation using the joint model Problem:

Iteratively updating w and X until convergence w sub-problem small-scale L1 regularized square loss minimization X sub-problem Half-quadratic splitting [Krishnan 2009]

likelihood prior

8

Page 9: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Experimental results Initialization: Wiener estimates / blurred images Dictionary: learned from Berkeley Segmentation database

Patch size 12 12 Prior parameters: Runtime: (Matlab) 16s with Intel Core2 Duo CPU @2.26GHz Experiment settings:

9

Page 10: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Experimental results

test 1

test 2

test 3

test 4

2 3 4 5 6 7 8

ISNR comparison

proposed

[Portilla 2009]

[Krishnan 2009]

test 1

test 2

test 3

test 4

0.8 0.82 0.84 0.86 0.88 0.9 SSIM comparison

proposed

[Portilla 2009]

[Krishnan 2009]

PASCAL Visual Object Classes

Challenge (VOC) 2007 database

10

Page 11: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Experimental results

11

Page 12: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Experimental results

keeps more brick

textures

[Krishnan 2009]

Original image

Blurred image Proposed12

[Portilla 2009]

Page 13: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Experimental results

13

Textures zoomed in

[Krishnan 2009]

Original image

Proposed[Portilla 2009]

Page 14: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Conclusions Global model for MAP estimation

Able to solve general non-blind image deconvolution Joint model of image pixels and representation

coefficients Sparsity of spatial derivative (analysis-based) Sparsity of representation of patches in overcomplete

dictionary (synthesis-based) Iterative algorithm

converges in a few iterations Matlab code for the proposed method is available at http://users.ece.utexas.edu/~bevans/papers/2011/sparsity/

14

Page 15: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

References [Elad 2007] M. Elad, P. Milanfar and R. Rubinstein, “Analysis versus

synthesis in signal priors”, Inverse Problems, vol. 23, 2007. [Krishnan 2009] D. Krishnan and R. Fergus, “Fast image deconvolution using

hyper-Laplacian priors,” Advances in Neural Information Processing Systems, vol. 22, pp. 1-9, 2009.

[Olshausen 1997] B.A. Olshausen and D.J. Field, “Sparse coding with an overcomplete basis set: a strategy employed by V1,” Vision Research, vol. 37, no. 23, pp. 3311-3325, 1997.

[Portilla 2009] J. Portilla, “Image restoration through L0 analysis-based sparse optimization in tight frames,” in Proc. IEEE Int. Conf. on Image Processing, 2009, pp. 3909-3912.

[Yang 2010] J. yang, J. Wright, T.S. Huang and Y. Ma, “Image super-resolution via sparse representation,” IEEE Trans. on Image Processing, vol. 19, no. 11, pp. 2861-2873, 2010.

15

Page 16: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Thank you!

16

Page 17: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

w sub-problem

patches do not overlap

small-scale l1 regularized square loss minimization

17

Page 18: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

X sub-problem

Conjugate gradientiteratively reweighted least squares

Half-quadratic splitting [Krishnan 2009]

auxiliary variable

No need to solve the equationcomponent-wise

quartic function18

Page 19: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

MAP estimation using the joint modelblurred image;

noise level; blurring kernel; initialization of recovered image

Update the coefficient of patches

(w sub-problem)

Set α=α0 α>αmax ?

Update auxiliary variable Y

(quartic equation)

Update image X (FFT)

α=kα

X converges?

finish

X sub-problem

19

Page 20: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Image Quality Assessment

20

Full reference metric ISNR -- increment in PSNR (peak signal-to-noise ratio)

SSIM -- structural similarity [Wang 2004]

Page 21: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Prior model of natural images Analysis-based prior

Fast convergence Over smooth the images

Synthesis-based prior (patch-based sparse representation) Dictionary well adapted to nature images Captures textures well Slow convergence Boundary artifacts

21

Page 22: Patch-based Image Deconvolution via Joint Modeling of Sparse Priors

Computational complexity

22

Computational complexity For each iteration: N is the total number of pixels in the image

Average runtime comparison[Krishnan

2009][Portilla 2009]

Proposed

2s 15s 16s