Transcript
Page 1: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Learning sparse representationsLearning sparse representationsto restore, classify, and sense images and videos

Guillermo Sapirop

University of Minnesota

Supported by NGA, ONR, DARPA, ARO, NSF, NIH

Page 2: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Compressive LecturingCompressive Lecturing

Learning Sparsity 2

Page 3: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Ramirez

Martin Duarte

Lecumberry

Rodriguez

Learning Sparsity 3

Page 4: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

For the past ~2 years …• Introduce and Extend the K-SVD

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

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

• Incoherent dictionaries– Ramirez, Lecumberry, Sapiro, January 2009, pre-print

• Learning to classify– Mairal, Bach, Ponce, Sapiro, Zisserman, CVPR 2008, NIPS 2008Mairal, 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

Duarte and Sapiro, pre print, May 2008, IEEE TIP to appear

Page 5: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

The Sparseland Model for Images

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

NN The vector α

contains very few (say L) non-zeros.

=x

N

DA fixed Dictionary A sparse

& random vector

xD α

Learning Sparsity 5

Page 6: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

What Should the Dictionary D Be?

α=≤α−α=α ˆx̂L.t.sy21

minargˆ 00

22

DDα 2

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

Learn D :

Multiscale LearningOne approach to choose D is from a known set of transforms (Steerable

Color Image Examples

Task / sensing adapted

Internal structure

wavelet, Curvelet, Contourlets, Bandlets, …)

Learning Sparsity 6

Internal structure

Page 7: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Dictionaries for Reconstruction

Learning Sparsity 7

Page 8: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Color multiscale dictionaries

Learning Sparsity 8

Page 9: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Example: Non-uniform noise

Learning Sparsity 9

Page 10: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Example: Inpainting

Learning Sparsity 10

Page 11: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Learning multiple reconstructive and discriminative dictionaries

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

Learning Sparsity 11

, , , , ,

Page 12: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Semi-supervised detection learningp g

MIT -- Learning Sparsity 12

Page 13: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Supervised Dictionary Learning

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

Learning Sparsity 13

, , , ,

Page 14: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Learning Incoherent Dictionaries

• Optimization depends on the incoherence• Improved generalization properties• Improved classificationImproved classification

Learning Sparsity 14

Page 15: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Results

Learning Sparsity 15

Page 16: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Results

Learning Sparsity 16

Page 17: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Learning for Compressed Sensing

+ “RIP (Identity Gramm Matrix)”

Learning Sparsity 17

Page 18: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Design the dictionary and sensing togethertogether

Learning Sparsity 18

Page 19: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Just Believe the PicturesJust Believe the Pictures

Learning Sparsity 19

Page 20: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Just Believe the PicturesJust Believe the Pictures

Learning Sparsity 20

Page 21: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Just Believe the PicturesJust Believe the Pictures

Learning Sparsity 21

Page 22: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Conclusions• State-of-the-art denoising results for still

(shared with Dabov et al ) and video(shared with Dabov et al.) and video• Vectorial and multiscale learned

dictionariesdictionaries• Dictionary learning with internal structure• Dictionary learning for classification• Dictionary learning for classification• Dictionary learning for sensing

• Dictionary learning for the task• Optimization is dictionary dependent

Learning Sparsity 22

p y p

Page 23: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Please do not use the wrong dictionaries thanksdictionaries… thanks

• 12 M pixel image• 7 million patches7 million patches• LARS+online

learning:learning: ~8 minutes

Learning Sparsity 23

Page 24: to restore, classify, and sense images and videospeople.ee.duke.edu/~lcarin/sapiro.pdfto restore, classify, and sense images and videos ... Bach,Mairal, Bach, Ponce, ... – Duarte

Learning Sparsity 24