The Compressive Sensing

Oct. 29th,2012

Contents

Compressive Sensing overview The Single-pixel camera Conclusions

Compressive Sensing overview

Sampling theorem

Shannon stated that

If a function x(t) contains no frequencies higher than B hertz, it is completely determined by giving its ordinates at a series of points spaced 1/(2B) seconds apart.
f s f max completely determined f s : sampling frequency f max : highest frequency component of signal
Claude Elwood Shannon (19162001)

Claude E. Shannon, Communication in the Presence of Noise, Proceeding of the I.R.E., vol. 37, pp. 10-21, January, 1949.

Dimensionality reduction by Shannon

2TW samples of a signal 2TW dimensions

Message bearing signals: bear several points are very important

can treat a group of dimensions as one dimensional reduction

Example: speech signal

Human ear is insensitive to a certain amount of phase distortion Messages differing only in phases sound the same All the points which differ only in phase can be grouped together and treated as one point
In the case of sounds, if the ear were completely insensitive to phase, then the number of dimensions would be reduced by one-half due to this cause alone. The sine and cosine components an and bn for a given frequency 2 2 would not need to be specified independently, but only an bn that is, the total amplitude for this frequency.
5

Compressible signal

Compressible signal is signal which sparse in a certain transform domain

Sparse: almost energy of x is distributed in some (K) coefficients almost coefficients (N - K) are very small or zero

x s

Where x is signal in time/space domain, size Nx1 s is signal in transform domain, size Nx1 is transform matrix, size NxN
Power spectrum density

2D-DCT transform

Width

Height

Compressive Sensing in a Nutshell

Directly acquire compressed samples: sampling and compression as a whole Compressively measure via linear projections

y x s s
Where x is signal in time/space domain, size Nx1 s is signal in transform domain, size Nx1 is transform matrix, size NxN is measurement matrix

Compressive Sensing in a Nutshell

Objectives
Under

Sparse signal to be sensed Good measurement matrix: compressive and robust

Satisfies Null Space condition Satisfies Restricted Isolated Property Incoherent Property

we can use

Good signal recovery algorithm: fast and robust

to perfectly reconstruct sparse signal

An application: The single pixel camera

Why single-pixel camera

CCD/CMOS technology is much cheaper over time

A semiconductor material, named Silicon, is widely used in large scale circuit Silicon also readily converts photon in visual wavelength to electrons.

http://photographythebest.blogspot.kr/2011/07/mekanisme-asas-kamera-digital.html 10

Why single-pixel camera

Imaging at wavelengths where silicon is blind is considerably more complicated and expensive

E.g., for comparable resolution, 500$ visible-wavelength camera becomes 50,000$ for the infrared camera.

HOW ?

If photodiode is very expensive

Single-pixel photodiode

http://photographythebest.blogspot.kr/2011/07/mekanisme-asas-kamera-digital.html 11

Architecture

A single-pixel camera model

y x
Image to be sent x Measurements y
DMD: digital micro-mirror device RNG: random number generator A/D: analog to digital converter DSP: digital signal processor

Lens: focuses light to DMD

Lens: acts as accumulor

DMD: acts as measurement matrix

http://dsp.rice.edu/cscamera 12

Architecture

A practical single-pixel camera

http://dsp.rice.edu/cscamera 13

Measurement matrix
Matrix : - Store in SRAM and control Digital Micro-mirror Device - Matrix is Gaussian matrix - Generate by Random Number Generator

Reflect light in one of two directions (20 optical) depending upon the state of the underlying memory cell (SRAM). SRAM value 1': reflect light into the optical system SRAM value '0': reflect light out of the optical system

http://www.opticalsciences.com/dmd.html 14

Experimental results (1)

(a) (b) (c)

(d)

(e)

(f)

Examples (a, d) original, (b, e) 4096 Pixels 800 Measurements (20%), (c, f) 4096 Pixels 1600 Measurements (40%)
15

Conclusions

16

Advance on Recovery Algorithms

Some advance techniques

Convex optimization-based methods Greedy algorithms Combinatorial algorithms Bayesian methods

Need more research on fast and robustness techniques

17

Other promising applications

The most promising applications

Sparse error correction Group testing and data stream algorithms Compressive medical imaging Single-pixel camera Hyper-spectral imaging Compressive sensing of manifold-model data Inference using compressive measurements Compressive sensor networks

Still very opened and promising area for researching

18

Thank you for your attention!

19