Learning sparse representations to restore, classify, and sense images and videos Guillermo Sapiro...

Learning sparse representations to restore, classify, and sense images and videos

Guillermo Sapiro

University of Minnesota

Supported by NSF, NGA, NIH, ONR, DARPA, ARO, McKnight FoundationSupported by NSF, NGA, NIH, ONR, DARPA, ARO, McKnight Foundation

Learning Sparsity 2

Martin Duarte

Rodriguez

Ramirez

Lecumberry

Learning Sparsity 3

Overview

• Introduction– Denoising, Demosaicing, Inpainting– Mairal, Elad, Sapiro, IEEE-TIP, January 2008

• Learn multiscale dictionaries – Mairal, Elad, Sapiro, SIAM-MMS, April 2008

• Sparsity + Self-similarity– Mairal, Bach, Ponce, Sapiro, Zisserman, pre-print.

• Incoherent dictionaries and universal coding– Ramirez, Lecumberry, Sapiro, June 2009, pre-print

• Learning to classify– Mairal, Bach, Ponce, Sapiro, Zisserman, CVPR 2008, NIPS 2008– Rodriguez and Sapiro, pre-print, 2008.

• Learning to sense sparse signals– Duarte and Sapiro, pre-print, May 2008, IEEE-TIP to appear

Learning Sparsity 4

Introduction I: Sparse and

Redundant Representations

Webster Dictionary: Of few and scattered elements

Learning Sparsity 5

Relation to measurements

Restoration by Energy Minimization

Thomas Bayes 1702 -

1761

Prior or regularizationy : Given measurements

x : Unknown to be recovered

xPryx21

xf22

Restoration/representation algorithms are often related to the minimization of an energy function of the form

Bayesian type of approach

What is the prior? What is the image model?

Learning Sparsity 6

The Sparseland Model for Images

MK

N

DA fixed Dictionary

Every column in D (dictionary) is a prototype signal (Atom).

The vector contains very few (say L) non-zeros.A sparse

& random vector

αx

N

Learning Sparsity 7

What Should the Dictionary D Be?

ˆx̂L.t.sy21

minargˆ00

22

DD

D should be chosen such that it sparsifies the representations (for a given task!)

Learn D :

Multiscale Learning

Color Image Examples

Task / sensing adapted

Internal structure

One approach to choose D is from a known set of transforms

(Steerable wavelet, Curvelet, Contourlets, Bandlets, …)

Learning Sparsity 8

Introduction II: Dictionary

Learning

Learning Sparsity 9

Each example is a linear combination of atoms from

D

Measure of Quality for D

DX A

Each example has a sparse representation with no more than L

atoms

L,j.t.sxMin0

0j

P

1j

2

2jj,

DAD Field & Olshausen (‘96)

Engan et. al. (‘99)Lewicki & Sejnowski (‘00)

Cotter et. al. (‘03)Gribonval et. al. (‘04)

Aharon, Elad, & Bruckstein (‘04) Aharon, Elad, & Bruckstein (‘05)

Ng et al. (‘07)Mairal, Sapiro, Elad (‘08)

Learning Sparsity 10

The K–SVD Algorithm – General

DInitialize D

Sparse Coding

Orthogonal Matching Pursuit (or L1)

Dictionary Update

Column-by-Column by SVD computation over the relevant

examples

Aharon, Elad, & Bruckstein (`04)

XT

Learning Sparsity 11

Show me the pictures

Learning Sparsity 12

Change the Metric in the OMP

Learning Sparsity 13

Non-uniform noise

Learning Sparsity 14

Example: Non-uniform noise

Learning Sparsity 15

Example: Inpainting

Learning Sparsity 16

Example: Demoisaic

Learning Sparsity 17

Example: Inpainting

Learning Sparsity 18

Not enough fun yet?:

Multiscale Dictionaries

Learning Sparsity 19

Learned multiscale dictionary

Learning Sparsity 20

Learning Sparsity 21

Color multiscale dictionaries

Learning Sparsity 22

Example

Learning Sparsity 23

Video inpainting

Extending the Models

Learning Sparsity 24

Universal Coding and Incoherent Dictionaries

• Consistent• Improved generalization properties• Improved active set computation• Improved coding speed• Improved reconstruction• See poster by Ramirez and Lecumberry…

Learning Sparsity 25

Sparsity + Self-similarity=Group Sparsity

• Combine the two of the most successful models for images

• Mairal, Bach. Ponce, Sapiro, Zisserman, pre-print, 2009

Learning Sparsity 26

Learning to Classify

Learning Sparsity 28

Global Dictionary

Learning Sparsity 29

Barbara

Learning Sparsity 30

Boat

Learning Sparsity 31

Digits

Which dictionary? How to learn them?

• Multiple reconstructive dictionary? (Payre)

• Single reconstructive dictionary? (Ng et al, LeCunn et al.)

• Dictionaries for classification!

• See also Winn et al., Holub et al., Lasserre et al., Hinton et al. for joint discriminative/generative probabilistic approaches

Learning Sparsity 32

Learning Sparsity 33

Learning multiple reconstructive and discriminative dictionaries

With J. Mairal, F. Bach, J. Ponce, and A. Zisserman, CVPR ’08, NIPS ‘08

Learning Sparsity 34

Texture classification

Semi-supervised detection learning

MIT -- Learning Sparsity 35

Learning Sparsity 36

Learning a Single Discriminative and Reconstructive Dictionary

• Exploit the representation coefficients for classification– Include this in the optimization

– Class supervised simultaneous OMP

With F. Rodriguez

Learning Sparsity 37

Digits images: Robust to noise and occlusions

Learning Sparsity 38

Supervised Dictionary Learning

With J. Mairal, F. Bach, J. Ponce, and A. Zisserman, NIPS ‘08

Learning to Sense Sparse Images

Motivation

• Compressed sensing (Candes &Tao, Donoho, et al.)– Sparsity

– Random sampling• Universality

• Stability

• Shall the sensing be adapted to the data type?– Yes! (Elad, Peyre, Weiss et al., Applebaum et al, this talk).

• Shall the sensing and dictionary be learned simultaneously?

Learning Sparsity 40

Some formulas….

Learning Sparsity 41

+ “RIP (Identity Gramm Matrix)”

Design the dictionary and sensing together

Learning Sparsity 42

Just Believe the Pictures

Learning Sparsity 43

Just Believe the Pictures

Learning Sparsity 44

Just Believe the Pictures

Learning Sparsity 45

Learning Sparsity 46

Conclusions• State-of-the-art denoising results for still (shared with Dabov et

al.) and video• General• Vectorial and multiscale learned dictionaries• Dictionaries with internal structure• Dictionary learning for classification

– See also Szlam and Sapiro, ICML 2009– See also Carin et al, ICML 2009

• Dictionary learning for sensing

• A lot of work still to be done!

Please do not use the wrong dictionaries…

• 12 M pixel image

• 7 million patches

• LARS+online learning:

~8 minutes

• Mairal, Bach, Ponce, Sapiro, ICML 2009

Learning Sparsity 47

Learning Sparsity 48