non-parametric bayesian dictionary learning for sparse image...
TRANSCRIPT
Non-parametric Bayesian techniques are considered for learningdictionaries for sparse image representations, with applications indenoising, inpainting and compressive sensing. The four maincontributions of our paper are:
The dictionary is learned using a beta process construction, andtherefore the number of dictionary elements and their relativeimportance may be inferred non-parametrically.For the denoising and inpainting applications, we do not have toassume a priori knowledge of the noise variance or sparsity level.The spatial inter-relationships between different components inimages are exploited by use of the Dirichlet process and a probitstick-breaking process.Using learned dictionaries, inferred off-line or in situ, theproposed approach yields CS performance that is markedly betterthan existing standard CS methods as applied to imagery.
Non-Parametric Bayesian Dictionary Learning for Sparse Image RepresentationsMingyuan Zhou, Haojun Chen, John Paisley, Lu Ren, 1Guillermo Sapiro and Lawrence Carin
Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA1Department of Electrical and Computer Engineering, University of Minnesota, Minneapolis, MN, USA
{mz1, hc44, jwp4, lr, lcarin}@ee.duke.edu, {guille}@umn.edu
Introduction
Model and Inference
Full likelihood of the Model
Gibbs Sampling Inference
Image denoising
Image inpainting
Applications
Compressive sensing
Corrupted image, 10.9475dB
spectr
al band_100
Original imageRestored image,25.8839dB
Corrupted image, 10.9475dB
spectr
al band_1
Original imageRestored image,25.8839dB
80
%P
ixe
ls M
issi
ng
50
%P
ixe
ls M
issi
ng
480*321 RGB80% Pixels Missing
150*150*210 Hyperspectral
95% Pixels Missing
Image size: 480 by 320, 2400 8 by 8 patches, 153600 coefficients are estimated
3 3.5 4 4.5 5 5.5 6 6.5 7 7.5
x 104
0.2
0.25
0.3
0.35
0.4
0.45
0.5
Number of Measurements
Rela
tive R
econstr
uction E
rror
PSBP BP
DP BP
BP
BCS
Fast BCS
Basis Pursuit
LARS/Lasso
OMP
STOMP-CFAR
3 3.5 4 4.5 5 5.5 6 6.5 7 7.5
x 104
0
0.05
0.1
0.15
0.2
0.25
0.3
Number of Measurements
Rela
tive R
econstr
uction E
rror
PSBP BP
DP BP
Online BP
BP
BCS
Fast BCS
Basis Pursuit
LARS/Lasso
OMP
STOMP-CFAR
Over-complete DCT Learned Dictionary
100 200 300 400 500
100
200
300
400
500
600
700
800
100 200 300 400 500
100
200
300
400
500
600
700
800
100 200 300 400 500
100
200
300
400
500
600
700
800
100 200 300 400 500
100
200
300
400
500
600
700
800
Spectral band 50 Spectral band 90
Original Restored
845*512*106 Hyperspectral, 98% pixels missing
Original Scene
Original Restored