time frequency analysis_journey
TRANSCRIPT
We value our relationshipWe value our relationships.
7 January 2016
© 2014We value our relationship7 January 2016
Chandrashekhar Padole
Title for Presentation Journey from
Vector Algebra to Wavelet Transform
(Time-Frequency Analysis)
@CopyrightPresenter: Chandra7 January 2016 2
Objective
Topics to be covered (Various Transforms):
• Vector Algebra
•Discrete Fourier Transform
•KL transform
•PCA
•Wavelet Transform
•Wigner Distribution
@CopyrightPresenter: Chandra7 January 2016 3
Applications
•Pattern Recognition Problems
• Biometrics: Face, Fingerprint, IRIS etc
• Automotive: Lane , Vehicle, Pedestrian, Sleeping Pattern, Signal
etc
• Manufacturing: object, documents etc
• GIS, Healthcare, Military and so on
•Analysis-Synthesis (Graphic Equalizer, noise removal,
image restoration etc)
•Data-Mining (content based image retrieval, audio mining
based on emotions etc)
•Transmission and Compression of data
@CopyrightPresenter: Chandra7 January 2016 4
Agenda
• Vector Algebra
•QA
•Fourier Transform
•QA
•K-L Transform
•PCA
•QA
•Wavelet Transform
•QA
•Wigner Distribution
•QA
@CopyrightPresenter: Chandra7 January 2016 5
Take-off for Vector Algebra
@CopyrightPresenter: Chandra7 January 2016 6
Preliminaries in Vector Algebra
•Vector P=3 i + 4 j
•P quantity to be analyzed or represented as P
•Projection or mapping of P onto unit vectors i and j will be
3 and 4 resp.
•Measurement of projection is obtained by using
mathematical tool or operator.
•In this case, operator is dot
product (inner product)
@CopyrightPresenter: Chandra7 January 2016 7
Contd..
• e.g. P.i=[ 3i+4j] . [ i]=[3i+4j].[1i+0j]=3
•And P.j=[ 3i+4j] . [ j]=[3i+4j].[0i+1j]=4
•Significance of i and j-
•In 2D space problem
• i is the vector which represents horizontal property
(feature/dimension/axis),
• j is for vertical property
• k- depth property in 3D
@CopyrightPresenter: Chandra7 January 2016 8
Let’s Look at some Real-life Problems
Two Cupboards
@CopyrightPresenter: Chandra7 January 2016 9
Contd..
Chemist
@CopyrightPresenter: Chandra7 January 2016 10
Contd..
Carpenter
@CopyrightPresenter: Chandra7 January 2016 11
Rewind
• P=3i + 4j
•Px = P. i= 3
•Py= P . j= 4
•Operator we used for measurement of projection is dot
product or inner product
•[ 3 4 2][2 1 2]T = 3.2 + 4.1+ 2.2=6 +4+4= 14
•It is also a correlation (zero-shift) between two vectors
• P � properties� component value/coefficient
•It’s a analysis process
@CopyrightPresenter: Chandra7 January 2016 12
Contd..
•To synthesize combine component values along with
property vectors to get original P
i.e. combine ( 3 and 4) as 3i+4j= P
•What we had till now , is 2D problem. Same terminology and
process can be extended for 3 D space problem….
@CopyrightPresenter: Chandra7 January 2016 13
3D Space Problem
• P= ai+bj+ck
• analysis : find { a,b,c}
• synthesis : combine a,b,c
•In 2D space problem – 2 analysis vectors
•In 3D space problem – 3 analysis vectors
@CopyrightPresenter: Chandra7 January 2016 14
Investigation in A-S Problem
• Can there be more than 3 analysis vectors ..?
if so , give examples..
•Can each analysis vector be multi valued ( as opposed to
double/triple valued vector in 2D /3D analysis) ?
•If both answers are positive, can vector algebra analysis
theory be extended to those problems?
@CopyrightPresenter: Chandra7 January 2016 15
Stop for Queries
@CopyrightPresenter: Chandra7 January 2016 16
Replacing VA by DFT
•Use of DFT – frequency analysis
•Why it is needed?
• to measure/analyze the variation of physical quantity
(such as pressure, intensity, voltage , current etc) with
respect to one or more independent variables
•Representation of physical quantity with respect to
independent variables is called as function or signal
•Variation need not to be periodic and mostly it is aperiodic
in real life problems
• For aperiodic DFT will analyze variations in signal
@CopyrightPresenter: Chandra7 January 2016 17
Take-off for Vector Algebra
@CopyrightPresenter: Chandra7 January 2016 18
Fourier Idea
@CopyrightPresenter: Chandra7 January 2016 19
Frequency Analysis
• Definition of DFT
• Computations steps in DFT
• Its physical significance
•Definition of DFT
for N-point DFT
∑−
=
−=
1
0
2
)()(N
n
N
nkj
enxkX
π
@CopyrightPresenter: Chandra7 January 2016 20
Computations in DFT
• Signal to be analyzed is 64 points in length
x(n) �
analog discrete
0 10 20 30 40 50 60 70-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
0 10 20 30 40 50 60 70-0 .5
-0 .4
-0 .3
-0 .2
-0 .1
0
0 .1
0 .2
0 .3
0 .4
0 .5
@CopyrightPresenter: Chandra7 January 2016 21
K=0
)/2sin()/2cos(
2
NnkjNnke N
nkj
πππ
+=−
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 700
0.5
1COSINE SEQUENCE
Am
plitu
de
n
@CopyrightPresenter: Chandra7 January 2016 22
K=1
)/2sin()/2cos(
2
NnkjNnke N
nkj
πππ
+=−
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ude
n
@CopyrightPresenter: Chandra7 January 2016 23
K=2
)/2sin()/2cos(
2
NnkjNnke N
nkj
πππ
+=−
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ude
n
@CopyrightPresenter: Chandra7 January 2016 24
K=3
)/2sin()/2cos(
2
NnkjNnke N
nkj
πππ
+=−
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ude
n
@CopyrightPresenter: Chandra7 January 2016 25
K=31
)/2sin()/2cos(
2
NnkjNnke N
nkj
πππ
+=−
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ude
n
@CopyrightPresenter: Chandra7 January 2016 26
DFT- Overall Physical Significance
0 1 0 2 0 3 0 4 0 5 0 6 0 7 0-0 . 5
-0 . 4
-0 . 3
-0 . 2
-0 . 1
0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 700
0.5
1
COSINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1
SINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1
COSINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ud
e
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1
SINE SEQUENCE
Am
plitu
de
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ud
e
n
a1+jb1
a5+jb5
Dot
Product
a2+jb2
a3+jb3
a4+jb4
X(1)
X(31)
X(2)
X(3)
X(4)
∑−
=
−=
1
0
2
)()(N
n
N
nkj
enxkX
π
@CopyrightPresenter: Chandra7 January 2016 27
DFT- Phase Information
clear all;
close all;
N=64;
t=0:N-1;
f=1;
ws=sin(2*pi*f*t/N);
figure,
subplot(2,1,1)
stem(ws);
wc=cos(2*pi*f*t/N);
subplot(2,1,2)
stem(wc);
%ps=sin(2*pi*f*t/N);
%ps=sin(2*pi*f*t/N + (pi/8));
%ps=sin(2*pi*f*t/N + (pi/4));
ps=sin(2*pi*f*t/N + (3*pi/8));
%ps=sin(2*pi*f*t/N + (pi/2));
figure,
stem(ps);
a=sum(ps.*ws);
b=sum(ps.*wc);
ph_ang=atan(b/a);
@CopyrightPresenter: Chandra7 January 2016 28
Contd..
0 10 20 30 40 50 60 70-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1SINE SEQUENCE
Am
plit
ude
n
0 10 20 30 40 50 60 70-1
-0.5
0
0.5
1COSINE SEQUENCE
Am
plit
ude
n
Dot
Product
a=32
b= 0
ph_ang=0
a=0
b= 32
ph_ang=pi/2
a= 22.67
b= 22.67
ph_ang=pi/4
0 10 20 30 40 50 60 70-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 10 20 30 40 50 60 70-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
ph=0
ph=3*pi/8
ph=pi/2
0 10 20 30 40 50 60 70-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
ph=pi/8
a= 29.56
b= 12.24
ph_ang=pi/8
0 10 20 30 40 50 60 70-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
ph=pi/4
a= 12.24
b= 29.56
ph_ang=pi/4
@CopyrightPresenter: Chandra7 January 2016 29
Relation bet VA and DFT
Vector algebra -2/3D space DFT
No of analysis vectors 2/3 N
No of elements in each
vector
2/3 N
Projection measurement
method
Inner Product Inner Product
@CopyrightPresenter: Chandra7 January 2016 30
Fourier Idea
Jean Baptiste Joseph
Fourier (1768 - 1830).
Fourier was a French
mathematician, who was
taught by Lagrange and
Laplace
@CopyrightPresenter: Chandra7 January 2016 31
Application –Example
Music Equalizer
Audio file recorded
with studio settingN-point FFT
N-point
User settingModified FFT
IFFT
@CopyrightPresenter: Chandra7 January 2016 32
Stop for Queries and Discussion
@CopyrightPresenter: Chandra7 January 2016 33
New terminology:
•quantity to be analyzed: signal /function
•Analysis vectors : basis vectors
• Thus, collection of basis vectors forms basis set
•Properties of basis set
• Completeness
• Orthogonal
• orthonormal
@CopyrightPresenter: Chandra7 January 2016 34
Completeness
• e.g. we have a function P=3i+4j+2k
• while analyzing we used only two basis vectors , i and k;
• using projection over these two vectors will not give
proper reconstruction
• But having so , we can achieve dimensionality reduction
which useful in data compression
•Dimensionality extension is possible?
XOR Problem
@CopyrightPresenter: Chandra7 January 2016 35
Orthogonal
•Information carried by one basis function should not be included in any other basis function from basis set
that’s a orthogonality
•Mathematically, if i.j=0, then I and j are orthogonal .
• In Fourier transform , and are orthogonal if
•In Fourier transform within basis function , two sub basis function ,sin and cosine, are also orthogonal and used for determining the phase of that frequency
N
nkj
e
π2
N
mkj
e
π2 nm ≠
∑∞
−∞=
=n
wmwn 0sinsin ∑∞
−∞=
=n
wnwn 0cossin
@CopyrightPresenter: Chandra7 January 2016 36
2D Fourier Transform
2D FT basis images for 8x8
@CopyrightPresenter: Chandra7 January 2016 37
Thank you
Queries?
@CopyrightPresenter: Chandra7 January 2016 38
Objective
Topics to be covered (Various Transforms):
• Vector Algebra( part I)
•Discrete Fourier Transform( part I)
•KL transform (part II)
•PCA ( part II)
•Wavelet Transform( part II)
•Wigner Distribution
@CopyrightPresenter: Chandra7 January 2016 39
Take-off for KL-Transform
@CopyrightPresenter: Chandra7 January 2016 40
Agenda
• Vector Algebra
•QA
•Fourier Transform
•QA
•K-L Transform
•PCA
•QA
•Wavelet Transform
•QA
•Wigner Distribution
•QA
@CopyrightPresenter: Chandra7 January 2016 41
KL Transform
• Basis vectors adopt to the data
•Powerful tool when reducing the dimensionality
•de-correlates the data => less redundancy
•Key idea: Represent the data in a more compact manner
•Used for PCA (to be discussed later)
•Thus , it provides the good and compact representation in
transformed domain
•Basis functions are derived from data itself
@CopyrightPresenter: Chandra7 January 2016 42
Contd..
Procedure to obtain KL transform basis functions
•Collect data
• subtract mean from the data (optional)
•Organize the data in matrix
• If sources are M and elements/dimensions in each
source data are N
• Then place a data in NxM matrix
•Calculate covariance of the matrix to get NxN matrix
•Calculate eigen vectors and eigen values
• N eigen vector and each vector will have N elements
• N eigen values corresponding to each eigen vector
@CopyrightPresenter: Chandra7 January 2016 43
PCA –Example
@CopyrightPresenter: Chandra7 January 2016 44
Contd..
@CopyrightPresenter: Chandra7 January 2016 45
Contd..
calculating covariance
Calculating eigen vectors and values
@CopyrightPresenter: Chandra7 January 2016 46
Contd..
@CopyrightPresenter: Chandra7 January 2016 47
Contd..
Reconstruction/restoration
@CopyrightPresenter: Chandra7 January 2016 48
Contd..
Matlab function used in PCA
• cov
• eig
@CopyrightPresenter: Chandra7 January 2016 49
Stop for Queries and Application
@CopyrightPresenter: Chandra
1st Person
7 January 2016 50
PCA- Face Recognition
1 2 7
MxNAveraging
2D-> 1D
Matrix
MNx40
2nd Person
1 2 7
40th Person
1 2 3
MNxMNCov Eig
MNxMN
MN
X
1
KL Transform
(Eigen Vectors)
Eigen Values
PCA1D-2D
@CopyrightPresenter: Chandra7 January 2016 51
WAVELET TRANSFORM
@CopyrightPresenter: Chandra7 January 2016 52
Need of WT
• Gives good resolution in time as well as frequency domain.
•It also gives locations of different frequency spectral
components during that particular instant of time.
•This is the main advantage of wavelet Transform over FT &
STFT.
•WT is used to mainly analyze non stationary signals, i.e.,
whose frequency response varies in time.
@CopyrightPresenter: Chandra7 January 2016 53
Take-off for Wavelet Transform
@CopyrightPresenter: Chandra7 January 2016 54
What is wavelet ?
• The term wavelet means a small wave , i.e. a window
function of finite length.
• It a oscillatory in which high frequency components exist
only for a short duration of time and low frequency exist
throughout the signal.
• The main feature of wavelet is that it can be of finite or
infinite duration but most of the energy of wavelet is
confined to a particular interval of time thus making it time
limited window function.
@CopyrightPresenter: Chandra7 January 2016 55
Properties of Wavelet
For any function to be a wavelet function must satisfy
following properties:
1.The function integrates to zero:
∞-∞∫Ψ(t) dt = 0.
2.It is square integrable or, equivalently has finite energy:
∞-∞∫Ψ(t)2 dt < ∞.
3.The admissibility condition:
∞C ≡ -∞∫ (Ψ(ω)2) / (ω) dω
Such that 0<C<∞.
@CopyrightPresenter: Chandra7 January 2016 56
Contd..
• Property 1 is suggestive of a function that is oscillatory or
that has a wavy appearance. Thus, in contrast to a sinusoidal
function, it is a “small wave” or a wavelet.
•Property 2 implies that most of the energy inΨ(t) is
confined to a finite duration.
•Property 3 is known as admissibility condition and is
sufficient condition that leads to the Inverse CWT but it is
not a necessary condition to obtain a mapping from the set
of CWTs back to L2R.
These properties are easily satisfied and there are an infinite
number of functions that qualify as mother wavelets.
@CopyrightPresenter: Chandra7 January 2016 57
Some Mother Wavelets
Mother wavelet: It’s a basic wavelet function at t=0 without
performing any scaling and shifting operation i.e.Ψ(t).
Morlet wavelet Haar
DB4
@CopyrightPresenter: Chandra7 January 2016 58
Definition of Wavelet Transform
∫∞
∞
−= dt
a
bt
atfbaW )(
1)(),( ψ
Where,
x(t) = Input signal which is to be
transformed
Ψ(t) = Mother wavelet
b = Time shift parameter
a = Scaling parameter
1/sqrt(a)=normalizing factor which ensures
that energy of remains constant for all values of
a & b.
@CopyrightPresenter: Chandra7 January 2016 59
MRA
Signal to be analyzed
@CopyrightPresenter: Chandra7 January 2016 60
Contd..
@CopyrightPresenter: Chandra7 January 2016 61
Contd..
@CopyrightPresenter: Chandra7 January 2016 62
Contd..
@CopyrightPresenter: Chandra7 January 2016 63
Chirp function
@CopyrightPresenter: Chandra7 January 2016 64
Time-Frequency Resolution
@CopyrightPresenter: Chandra7 January 2016 65
DWT Calculation
Mallat Algorithm:
@CopyrightPresenter: Chandra7 January 2016 66
Stop for Calculation of WT
@CopyrightPresenter: Chandra7 January 2016 67
2D DWT
Music Equalizer
@CopyrightPresenter: Chandra7 January 2016 68
Matlab Code
[im1,n]=imread( ‘lena.jpg‘);
im=double(im1);
wname='haar';
[A_1,H_1,V_1,D_1] = DWT2(im,wname);
[A_2,H_2,V_2,D_2] = DWT2(A_1,wname);
[A_3,H_3,V_3,D_3] = DWT2(A_2,wname);
[A_4,H_4,V_4,D_4] = DWT2(A_3,wname);
@CopyrightPresenter: Chandra7 January 2016 69
Thank you
Queries?