Spectral Representation Bandlimited Signals/Systems ...bioucas/IP/files/Image Deconvolution.pdf · Convolution operators Spectral Representation Bandlimited Signals/Systems ... Approximate

IP, José Bioucas Dias, IST, 2015 1

Convolution operators

Spectral Representation

Bandlimited Signals/Systems

Inverse Operator

Null and Range Spaces

Sampling, DFT, and FFT

Tikhonov Regularization/Wiener Filtering

IP, José Bioucas Dias, IST, 2015 2

Convolution Operators

Definition:

Spectral representation of a convolution operator:

FT

IP, José Bioucas Dias, IST, 2015 3

• A is linear and bounded

• A is bounded:

Let

is continuous

Adjoint of a convolution operator

Properties

Parseval’s Theorem

IP, José Bioucas Dias, IST, 2015 4

Adjoint of convolution operator (cont.)

since

Inverse of a convolution operator or has isolated zeros

as

is not bounded

is defined only if

IP, José Bioucas Dias, IST, 2015 5

Bandlimited convolution operators/systems

is bandlimited with band B, i.e.,

are orthogonal

IP, José Bioucas Dias, IST, 2015 6

Sampling 2D signals

IP, José Bioucas Dias, IST, 2015 7

Sampling 2D signals

IP, José Bioucas Dias, IST, 2015 8

Convolution of bandlimited 2D signals

Approximate using periodic sequences

IP, José Bioucas Dias, IST, 2015 9

From continuous to discrete representation

Assume that

Let

Let is N-periodic sequences such that

Discrete Fourier Transform (DFT)

IP, José Bioucas Dias, IST, 2015 10

Fast fourier transform (FFT)

Efficient algorithm to compute

When N is a power of 2

IP, José Bioucas Dias, IST, 2015 11

Vector space perspective

Let vectors defined in Euclidian vector space with inner product

Parseval generalized equality

Basis

IP, José Bioucas Dias, IST, 2015 12

2D Periodic convolution

2D N-periodic signals (images)

Periodic convolution

DFT of a convolution

Hadamard product

IP, José Bioucas Dias, IST, 2015 13

Spectral Representation of 2D Periodic Signals

Can be represented as a block cyclic matrix

Spectral Representation of A

eingenvalues of A

IP, José Bioucas Dias, IST, 2015 14

Adjoint operator

Operator

IP, José Bioucas Dias, IST, 2015 15

Inverse operator

Let

IP, José Bioucas Dias, IST, 2015 16

Deconvolution Examples

Imaging Systems

Linear Imaging

System

System noise + Poisson noise

Impulsive Response function

or

Point spread function (PSF)

Invariant systems

is the transfer function (TF)

IP, José Bioucas Dias, IST, 2015 17

Example 1: linear motion blur

lens plane

Let a(t)=ct for , then

target velocity

IP, José Bioucas Dias, IST, 2015 18

Example 1: Linear motion blur

IP, José Bioucas Dias, IST, 2015 19

Example 1: Linear Motion Blur

IP, José Bioucas Dias, IST, 2015 20

Example 2: out of focus blur

lens plane

Circle of confusion COC

Geometrical optics

0 5 10 15 20 25 30-0.1

0

0.1

0.2

0.3

0.4

0.5

0.6

zeros

IP, José Bioucas Dias, IST, 2015 21

Deconvolution of linear motion blur

Let and

IP, José Bioucas Dias, IST, 2015 22

Deconvolution of Linear Motion Blur

IP, José Bioucas Dias, IST, 2015 23

Deconvolution of linear motion blur (TFD)

-4 -3 -2 -1 0 1 2 3 40

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2-8

-6

-4

-2

0

2

4

6

8

ISNR

IP, José Bioucas Dias, IST, 2015 24

Deconvolution of Linear Motion Blur (Tikhonov regularization)

Assuming that D is cyclic convolution operator

Wiener filter

Regularization

filter

IP, José Bioucas Dias, IST, 2015 25

Deconvolution of linear motion blur (Tikhonov regularization)

Regularization

filter

Effect of the regularization filter

is a frequency selective threshold

IP, José Bioucas Dias, IST, 2015 26

Deconvolution of Linear Motion Blur

IP, José Bioucas Dias, IST, 2015 27

Deconvolution of linear motion blur (Total Variation )

Iterative Denoising algorithm

where solves the denoising optimization problem

IP, José Bioucas Dias, IST, 2015 28

Deconvolution of Linear Motion Blur

TFD Tikhonov (D=I)

TV