Short Time Fourier Transform Time/Frequency localization depends on - - PowerPoint PPT Presentation

Short Time Fourier Transform

• Time/Frequency localization depends on window size.
• Once you choose a particular window size, it will be the same

for all frequencies.

• Many signals require a more flexible approach - vary the

window size to determine more accurately either time or frequency.

The Wavelet Transform

• Overcomes the preset resolution problem of the STFT by

using a variable length window:

– Narrower windows are more appropriate at high frequencies (better time localization) – Wider windows are more appropriate at low frequencies (better frequency localization)

The Wavelet Transform (cont’d)

Wide windows do not provide good localization at high frequencies.

The Wavelet Transform (cont’d)

A narrower window is more appropriate at high frequencies.

The Wavelet Transform (cont’d)

Narrow windows do not provide good localization at low frequencies.

The Wavelet Transform (cont’d)

A wider window is more appropriate at low frequencies.

What is a wavelet?

• It is a function that “waves” above and below the x-axis;

it has (1) varying frequency, (2) limited duration, and (3) an average value of zero.

• This is in contrast to sinusoids, used by FT, which have

infinite energy.

Sinusoid Wavelet

Wavelets

• Like sine and cosine functions in FT, wavelets can define

basis functions ψk(t):

• Span of ψk(t): vector space S containing all functions f(t)

that can be represented by ψk(t).

( ) ( )

k k k

f t a t ψ =∑

Wavelets (cont’d)

• There are many different wavelets:

Morlet Haar Daubechies

Basis Construction - Mother Wavelet

( )

jk t

ψ =

Basis Construction - Mother Wavelet (cont’d)

scale =1/2j (1/frequency)

( )

/2

( ) 2 2

j j jk t

t k ψ ψ = −

j k

Discrete Wavelet Transform (DWT)

( ) ( )

jk jk k j

f t a t ψ = ∑∑

( )

/2

( ) 2 2

j j jk t

t k ψ ψ = −

(inverse DWT) (forward DWT) where

*

( ) ( )

jk

jk t

a f t t ψ =∑

DFT vs DWT

• FT expansion:
• WT expansion
• r
• ne parameter basis

( ) ( )

l l l

f t a t ψ =∑ ( ) ( )

jk jk k j

f t a t ψ =∑∑

two parameter basis

Multiresolution Representation using

high resolution

(more details)

low resolution

(less details)

…

( ) ( )

jk jk k j

f t a t ψ =∑ ∑

( ) f t

1

ˆ ( ) f t

2

ˆ ( ) f t ˆ ( )

s

f t ( )

jk t

ψ

j

Prediction Residual Pyramid - Revisited

• In the absence of quantization errors, the approximation

pyramid can be reconstructed from the prediction residual pyramid.

• Prediction residual pyramid can be represented more

efficiently.

(with sub-sampling)

Efficient Representation Using “Details”

details D2

L0

details D3 details D1

(without sub-sampling)

Efficient Representation Using Details (cont’d)

representation: L0 D1 D2 D3

A wavelet representation of a function consists of (1) a coarse overall approximation (2) detail coefficients that influence the function at various scales.

in general: L0 D1 D2 D3…DJ

Reconstruction (synthesis)

H3=H2+D3 details D2 L0

details D3

H2=H1+D2 H1=L0+D1 details D1

(without sub-sampling)

Example - Haar Wavelets

• Suppose we are given a 1D "image" with a resolution
• f 4 pixels:

[9 7 3 5]

• The Haar wavelet transform is the following:

L0 D1 D2 D3

(with sub-sampling)

Example - Haar Wavelets (cont’d)

• Start by averaging the pixels together (pairwise) to get

a new lower resolution image:

• To recover the original four pixels from the two

averaged pixels, store some detail coefficients.

1 [9 7 3 5]

Example - Haar Wavelets (cont’d)

• Repeating this process on the averages (i.e., low

resolution image) gives the full decomposition:

1 Harr decomposition:

Example - Haar Wavelets (cont’d)

• We can reconstruct the original image by adding or

subtracting the detail coefficients from the lower- resolution versions.

2 1 -1

Example - Haar Wavelets (cont’d)

Detail coefficients become smaller and smaller as j increases.

Dj Dj-1 D1 L0

How to compute Di ?

Multiresolution Conditions

• If a set of functions can be represented by a weighted

sum of ψ(2jt - k), then a larger set, including the

• riginal, can be represented by a weighted sum of ψ(2j

+1t - k):

low resolution high resolution

j

Multiresolution Conditions (cont’d)

• If a set of functions can be represented by a weighted

sum of ψ(2jt - k), then a larger set, including the

• riginal, can be represented by a weighted sum of ψ(2j

+1t - k):

Vj: span of ψ(2jt - k):

( ) ( )

j k jk k

f t a t ψ =∑

Vj+1: span of ψ(2j+1t - k):

1 ( 1)

( ) ( )

j k j k k

f t b t ψ

+ +

=∑

1 j j

V V + ⊆

Nested Spaces Vj

ψ(t - k) ψ(2t - k) ψ(2jt - k) … V0 V1 Vj Vj : space spanned by ψ(2jt - k) Multiresolution conditions à nested spanned spaces:

( ) ( )

jk jk k j

f t a t ψ =∑ ∑

f(t) ϵ Vj Basis functions: i.e., if f(t) ϵ V j then f(t) ϵ V j+1

1 j j

V V + ⊂

How to compute Di ? (cont’d)

( ) ( )

jk jk k j

f t a t ψ =∑ ∑

f(t) ϵ Vj IDEA: define a set of basis functions that span the difference between Vj+1 and Vj

Orthogonal Complement Wj

• Let Wj be the orthogonal complement of Vj in Vj+1

Vj+1 = Vj + Wj

How to compute Di ? (cont’d)

If f(t) ϵ Vj+1, then f(t) can be represented using basis functions φ(t) from Vj+1:

1

( ) (2 )

j k k

f t c t k ϕ

+

= −

∑

( ) (2 ) (2 )

j j k jk k k

f t c t k d t k ϕ ψ = − + −

∑ ∑

Vj+1 = Vj + Wj Alternatively, f(t) can be represented using two sets of basis functions, φ(t) from Vj and ψ(t) from Wj: Vj+1

Think of Wj as a means to represent the parts of a function in Vj+1 that cannot be represented in Vj

1

( ) (2 )

j k k

f t c t k ϕ

+

= −

∑

( ) (2 ) (2 )

j j k jk k k

f t c t k d t k ϕ ψ = − + −

∑ ∑

Vj Wj

How to compute Di ? (cont’d)

differences between Vj and Vj+1

How to compute Di ? (cont’d)

• à using recursion on Vj:

( ) ( ) (2 )

j k jk k k j

f t c t k d t k ϕ ψ = − + −

∑ ∑∑

V0 W0, W1, W2, …

basis functions basis functions

Vj+1 = Vj-1+Wj-1+Wj = …= V0 + W0 + W1 + W2 + … + Wj if f(t) ϵ Vj+1 , then: Vj+1 = Vj + Wj

Summary: wavelet expansion (Section 7.2)

• Wavelet decompositions involve a pair of waveforms

(mother wavelets):

φ(t) ψ(t)

encode low resolution info encode details or high resolution info

( ) ( ) (2 )

j k jk k k j

f t c t k d t k ϕ ψ = − + −

∑ ∑∑

scaling function wavelet function

1D Haar Wavelets

• Haar scaling and wavelet functions:

computes average computes details

φ(t) ψ(t)

1D Haar Wavelets (cont’d)

• V0 represents the space of one pixel images
• Think of a one-pixel image as a function that is constant
• ver [0,1)

Example:

0 1

width: 1

1D Haar Wavelets (cont’d)

• V1 represents the space of all two-pixel images
• Think of a two-pixel image as a function having 21

equal-sized constant pieces over the interval [0, 1).

• Note that

Examples:

0 ½ 1

1

V V ⊂

= +

width: 1/2

e.g.,

1D Haar Wavelets (cont’d)

• V j represents all the 2j-pixel images
• Functions having 2j equal-sized constant pieces over

interval [0,1).

• Note that

Examples: width: 1/2j

ϵ Vj ϵ Vj

1 j j

V V

− ⊂

e.g.,

1D Haar Wavelets (cont’d)

V0, V1, ..., V j are nested

i.e.,

VJ … V1 V0 coarse details fine details

1 j j

V V + ⊂

Define a basis for Vj

• Mother scaling function:
• Let’s define a basis for V j :

0 1

note alternative notation:

( ) ( )

j i ji

x x ϕ ϕ ≡

1

Define a basis for Vj (cont’d)

Define a basis for Wj (cont’d)

• Mother wavelet function:
• Let’s define a basis ψ j

i for Wj :

1

• 1

0 1/2 1

( ) ( )

j i ji

x x ψ ψ ≡

note alternative notation:

Define a basis for Wj

Note that the dot product between basis functions in Vj and Wj is zero!

Define a basis for Wj (cont’d)

Basis functions ψ j

i of W j

Basis functions φ j

i of V j

form a basis in V j+1

Example - Revisited

f(x)= V2

Example (cont’d)

V1and W1 V2=V1+W1

φ1,0(x) φ1,1(x) ψ1,0(x) ψ1,1(x)

(divide by 2 for normalization)

Example (cont’d)

V2=V1+W1=V0+W0+W1 V0 ,W0 and W1

φ0,0(x) ψ0,0(x) ψ1,0 (x) ψ1,1(x)

(divide by 2 for normalization)

Example

Example (cont’d)

Filter banks

• The lower resolution coefficients can be calculated

from the higher resolution coefficients by a tree- structured algorithm (i.e., filter bank).

h0(-n) is a lowpass filter and h1(-n) is a highpass filter Subband encoding (analysis)

Example (revisited)

[9 7 3 5]

low-pass, down-sampling high-pass, down-sampling

(9+7)/2 (3+5)/2 (9-7)/2 (3-5)/2

V1 basis functions

Filter banks (cont’d)

Next level:

Example (revisited)

[9 7 3 5]

high-pass, down-sampling low-pass, down-sampling

(8+4)/2 (8-4)/2

V1 basis functions

Convention for illustrating 1D Haar wavelet decomposition

x x x x x x … x x

detail average

…

re-arrange: re-arrange: V1 basis functions

Examples of lowpass/highpass (analysis) filters

Haar h0 h1

Filter banks (cont’d)

• The higher resolution coefficients can be calculated from

the lower resolution coefficients using a similar structure. Subband encoding (synthesis)

Filter banks (cont’d)

Next level:

Examples of lowpass/highpass (synthesis) filters

Haar (same as for analysis):

+

g0 g1

2D Haar Wavelet Transform

• The 2D Haar wavelet decomposition can be computed

using 1D Haar wavelet decompositions.

– i.e., 2D Haar wavelet basis is separable

• Two decompositions (i.e., correspond to different

basis functions):

– Standard decomposition – Non-standard decomposition

Standard Haar wavelet decomposition

• Steps

(1) Compute 1D Haar wavelet decomposition of each row of

the original pixel values. (2) Compute 1D Haar wavelet decomposition of each column of the row-transformed pixels.

Standard Haar wavelet decomposition (cont’d)

x x x … x x x x … x … … . x x x ... x (1) row-wise Haar decomposition: …

detail average

… … … . … … … … .

re-arrange terms

Standard Haar wavelet decomposition (cont’d)

(1) row-wise Haar decomposition: …

detail average

… … … … . … … … . …

row-transformed result

Standard Haar wavelet decomposition (cont’d)

(2) column-wise Haar decomposition: …

detail average

… … … … . … … … … … . …

row-transformed result column-transformed result

Example

… … … … … .

row-transformed result

… … … .

re-arrange terms

Example (cont’d)

… … … … … .

column-transformed result

Non-standard Haar wavelet decomposition

• Alternates between operations on rows and columns.

(1) Perform one level decomposition in each row (i.e., one step of horizontal pairwise averaging and differencing). (2) Perform one level decomposition in each column from step 1 (i.e., one step of vertical pairwise averaging and differencing). (3) Repeat the process on the quadrant containing averages

• nly (i.e., in both directions).

Non-standard Haar wavelet decomposition (cont’d)

x x x … x x x x … x … … . x x x . . . x

• ne level, horizontal

Haar decomposition: … … … … . … … … … … .

• ne level, vertical

Haar decomposition: …

Non-standard Haar wavelet decomposition (cont’d)

• ne level, horizontal

Haar decomposition

• n “green” quadrant
• ne level, vertical

Haar decomposition

• n “green” quadrant

… … … … . … …

re-arrange terms

… … … … … . …

Example

… … … … . … …

re-arrange terms

Example (cont’d)

… … … … … .

2D DWT using filter banks (analysis)

LL LH HL HH

2D DWT using filter banks (analysis)

LL Hà LH Và HL Dà HH The wavelet transform is applied again on the LL sub-image

LL LL LH HL HH

2D DWT using filter banks (analysis)

Fast Multiresolution Image Querying

painted low resolution target

queries

Charles E. Jacobs Adam Finkelstein David H. Salesin, "Fast Multiresolution Image Querying", SIGRAPH, 1995.