21
29.10.2014 1 1/20 Department of Electrical Engineering University of Isfahan Introduction to Compressive Sensing Section I: Sparsity 2 References Y. C. Eldar and G. Kutyniok, Compressed Sensing: Theory and Application, Cambridge University Press, 2012. Fundamental of Image and Video Processing Online course

Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

1

1/20

Department of Electrical EngineeringUniversity of Isfahan

Introduction to

Compressive

SensingSection I: Sparsity

2

References

Y. C. Eldar and G. Kutyniok, Compressed Sensing: Theory

and Application, Cambridge University Press, 2012.

Fundamental of Image and Video Processing Online course

Page 2: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

2

3

Outlines

4

Introduction

Nyquist theory Vs. Compressive Sensing

DSP Reconstruction filter

Analog Sampling Kernels

DSP

Analog Interpolation

Kernels

Page 3: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

3

5

Introduction

Nyquist theory Vs. Compressive Sensing

DSP Reconstruction filter

Analog Sampling Kernels

DSP

Analog Interpolation

Kernels

Nyquist Theory CS Theory

infinite-length, continuous-time Sig

Sampling Measuring

Simple recovery Special and nonlinear recovery

6

Introduction

Nyquist

sampling:

Compressive

Sampling:

Many beautiful papers covering theory, algorithms, and applications

b A x

b = Ax

b A x

=

=

Page 4: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

4

7

Outlines

8

Review of Vector Spaces

𝒙 =𝟑𝟎𝟒

𝒙 2 = 3 2 + 4 2 = 25 = 5

𝒙 1 = 3 + 4 = 7

𝒙 0 = 2

𝒒𝒖𝒂𝒔𝒊𝒏𝒐𝒓𝒎

Page 5: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

5

9

Review of Vector Spaces

𝒒𝒖𝒂𝒔𝒊𝒏𝒐𝒓𝒎

10

Review of Vector Spaces

Page 6: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

6

11

Outlines

12

Signal Modeling

Low Dimensional Signal Models

Sparse ModelsLow-Rank Matrix Models

Parametric Models

Finite Union of Subspaces

Sparsity…

Page 7: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

7

13

Outlines

14

What is Sparsity?

A vector is said to be sparse if it only has “a few” non-zero components.

The vector can represent a signal, witch my be sparse in its native domain (e.g., image of sky at night) or can be made sparse in another domain (e.g., natural images in the DCT domain)

A sparse vector may originate in numerous applications

Page 8: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

8

15

What is Sparsity?

A vector is said to be sparse if it only has “a few” non-zero components.

The vector can represent a signal, witch my be sparse in its native domain (e.g., image of sky at night) or can be made sparse in another domain (e.g., natural images in the DCT domain)

A sparse vector may originate in numerous applications

=

16

Outlines

Page 9: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

9

17

Applications

Compressive Sensing

Image and Video Processing

Machine learning

Statistics

Genetics

Econometrics

Neuroscience

18

Econometrics

sparse

Page 10: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

10

19

Robust Regression

Least Square Method:

20

Robust Regression

Least Square Method:

Page 11: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

11

21

Recommender Systems

Matrix Completion problem

Rank Minimization Problem

Low Rank Matrix Many of Singular Values are zero

22

Image Denoising

DCT Domain Bases Sparse

Smooth

• Image Inpating• Super Resolution Image• Face Rrecognition• Video Surveillance• …

Page 12: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

12

23

Compressive Sensing

24

Outlines

Page 13: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

13

25

Linear Inverse Problems

26

Linear Inverse Problems

Page 14: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

14

27

Linear Inverse Problems

It depends on the application

Regularization Principle: Adding priori knowledge to problem

Sparse

28

Unit Sphere

Page 15: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

15

29

30

It’s not a sparse solution

Derivation of closed form solution

Page 16: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

16

31

Solutions are sparse

32

Solution is sparse

Basis Pursuit Problem

Page 17: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

17

33

Solution is sparse

34

On Convexity

Non-convex functionConvex function

Convex set Non-convex set

Page 18: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

18

35

Convex Optimization Problem

Convex function Convex set

36

A

Page 19: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

19

37

Small solutionClosed form

Sparse solutionNon- convex

Sparse solutionconvex Sparse solution

NP-Hard

Greedy approaches (Matching Pursuit) approximate the solution

Can be solved via convex optimization algorithms

38

Outlines

Page 20: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

20

39

Problem Reformulation

Noise in Observation

Swapping the Constraint and Objective

40

Problem Reformulation

Noise in Observation

Swapping the Constraint and Objective

convex convex

Page 21: Exploiting Structure in Data - University of Isfahanengold.ui.ac.ir/~sabahi/Advanced digital signal... · 2016-05-08 · DSP Reconstruction filter Analog Sampling Kernels DSP Interpolation

29.10.2014

21

41

Problem Reformulation

Bring constraint to objective

LASSO Problem (least absolute shrinkage and selection operator)