lecture unitarytransforms
TRANSCRIPT
-
7/27/2019 Lecture UnitaryTransforms
1/32
Unitary Trans form s
Borrowed from UMD ENEE631Spring04
-
7/27/2019 Lecture UnitaryTransforms
2/32
Image Trans form : A Revis i t
With A Coding Perspect ive
UM
CPENEE631Slides(cr
eatedbyM.Wu2004)
-
7/27/2019 Lecture UnitaryTransforms
3/32
Why Do Transform s?
Fast computation
E.g., convolution vs. multiplication for filter with wide support
Conceptual insights for various image processing
E.g., spatial frequency info. (smooth, moderate change, fast change, etc.)
Obtain transformed data as measurement
E.g., blurred images, radiology images (medical and astrophysics) Often need inverse transform
May need to get assistance from other transforms
For efficient storage and transmission Pick a few representatives (basis)
Just store/send the contribution from each basis
UM
CPENEE631Slides(cr
eatedbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
4/32
Basic Process o f Trans form Coding
UM
CPENEE631Slides(cr
eatedbyM.Wu2004)
Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 8)
-
7/27/2019 Lecture UnitaryTransforms
5/32
1-D DFT and Vecto r Form
{ z(n) }
{ Z(k) }n, k = 0, 1, , N-1
WN= exp{ - j2/ N }~ complex conjugate of primitive Nth root of unity
Vector form and interpretation for inverse transform
z = kZ(k) akak= [ 1 WN
-k WN-2k WN
-(N-1)k]T/N
Basis vectors
akH= ak
* T= [ 1 WNk WN
2k WN(N-1)k] /N
Use akHas row vectors to construct a matrix F
Z = F zz = F*TZ = F* Z
F is symmetric (FT=F) and unitary (F-1 = FHwhere FH= F*T)
1
0
1
0
)(1)(
)(1
)(
N
k
nk
N
N
n
nk
N
WkZN
nz
WnzN
kZ
UMCP
ENEE631Slides(createdbyM.Wu2001/2004)
-
7/27/2019 Lecture UnitaryTransforms
6/32
Basis Vecto rs and Basis Images
A basis for a vector space ~ a set of vectors
Linearly independent ~ ai vi = 0 if and only if all ai=0 Uniquely represent every vector in the space by their linear combination
~ bi vi ( spanning set {vi} )
Orthonormal basis
Orthogonality ~ inner product = y*Tx= 0
Normalized length ~ ||x ||2 = = x*Tx= 1
Inner product for 2-D arrays
= mn f(m,n) g*(m,n) = G1
*TF1 (rewrite matrix into vector)
!! Dont do FG ~ may not even be a valid operation for MxN matrices! 2D Basis Matrices (Basis Images)
Represent any images of the same size as a linear combination of basisimages
A vector space consists of a set of vectors, a
field of scalars, a vector addition operation,
and a scalar multiplication operation.
UM
CPENEE631Slides(cr
eatedbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
7/32
1-D Unitary Trans form
Linear invertible transform
1-D sequence { x(0), x(1), , x(N-1) } as a vector
y = A x andA is invertible
Unitary matrix ~A-1 = A*T
DenoteA*T
asAH
~ Hermitianx = A-1 y = A*Ty = ai
*Ty(i)
Hermitian of row vectors of A form a set of orthonormal basis vectorsai
*T = [a*(i,0), , a*(i,N-1)] T
Orthogonal matrix ~A-1 = AT Real-valued unitary matrixis also an orthogonal matrix
Row vectors of real orthogonal matrix A form orthonormal basis vectors
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
8/32
Propert ies o f 1-D Unitary Transfo rmy = A x
Energy Conservation
||y ||2 = ||x ||2
|| y ||2 = || Ax ||2= (Ax)*T(Ax)= x*TA*TA x = x*Tx = || x ||2
Rotation
The angles between vectors are preserved
A unitary transformation is a rotation of a vector in an
N-dimension space, i.e., a rotation of basis coordinatesUMCP
ENEE631Slides(createdbyM.Wu2001/2004)
-
7/27/2019 Lecture UnitaryTransforms
9/32
Propert ies o f 1-D Unitary Transfo rm(contd)
Energy Compaction
Many common unitary transforms tend to pack a large fraction of signalenergy into just a few transform coefficients
Decorrelation
Highly correlated input elements quite uncorrelated output coefficients
Covariance matrix E[ ( yE(y) ) ( yE(y) )*T
] small correlation implies small off-diagonal terms
Example: recall the effect of DFT
Question: What unitary transform gives the best compaction and decorrelation?
=> Will revisit this issue in a few lectures
UMCP
ENEE631Slides(crea
tedbyM.Wu2001/2004)
-
7/27/2019 Lecture UnitaryTransforms
10/32
Review : 1-D Discrete Cos ine Transfo rm(DCT)
Transform matrix C
c(k,n) = (0) for k=0
c(k,n) = (k) cos[(2n+1)/2N] for k>0
Cis real and orthogonal
rows ofCform orthonormal basis Cis not symmetric!
DCT is not the real part of unitary DFT! See Assignment#3
related to DFT of a symmetrically extended signal
Nk
N
N
knkkZnz
N
knknzkZ
N
k
N
n
2)(,
1)0(
2
)12(cos)()()(
2
)12(cos)()()(
1
0
1
0
UMCP
ENEE631Slides(crea
tedbyM.Wu2001/20
04)
-
7/27/2019 Lecture UnitaryTransforms
11/32
Period ic i ty Impl ied by DFT and DCT
UM
CPENEE631Slides(createdbyM.Wu2004)
Figure is from slides at Gonzalez/ Woods DIP book website (Chapter 8)
-
7/27/2019 Lecture UnitaryTransforms
12/32
Examp le of 1-D DCT
100
50
0
-50
-100
0 1 2 3 4 5 6 7
n
z(n)
100
50
0
-50
-100
0 1 2 3 4 5 6 7
k
Z(k)
DCT
From Ken Lams DCT talk 2001 (HK Polytech)
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
13/32
UM
CPENEE631Slides(createdbyM.Wu2001)
Examp le of 1-D DCT (contd)
k
Z(k)
Transform coeff.
1.0
0.0
-1.0
1.0
0.0
-1.0
1.0
0.0
-1.0
1.0
0.0
-1.0
1.0
0.0
-1.0
1.0
0.0
-1.0
1.0
0.0
-1.0
1.0
0.0
-1.0
Basis vectors
100
0
-100
100
0
-100
100
0
-100
100
0
-100
100
0
-100
100
0
-100
100
0
-100
100
0
-100u=0
u=0 to 1
u=0 to 4
u=0 to 5
u=0 to 2
u=0 to 3
u=0 to 6
u=0 to 7
Reconstructions
n
z(n)
Original signal
From Ken Lams DCT talk 2001 (HK Polytech)
-
7/27/2019 Lecture UnitaryTransforms
14/32
2-D DCT
Separable orthogonal transform
Apply 1-D DCT to each row, then to each columnY = C X CTX = CTY C = mn y(m,n) Bm,n
DCT basis images:
Equivalent to representan NxN image with a setof orthonormal NxNbasis images
Each DCT coefficient
indicates the contributionfrom (or similarity to) thecorresponding basis image
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
15/32
2-D Transfo rm : General Case
A general 2-D linear transform {ak,l(m,n)}
y(k,l) is a transform coefficient for Image {x(m,n)}
{y(k,l)} is Transformed Image
Equiv to rewriting all from 2-D to 1-D and applying 1-D transform
Computational complexity N2 values to compute
N2 terms in summation per output coefficient
O(N4) for transforming an NxN image!
),(),(),(
),(),(),(
1
0
1
0
,
1
0
1
0,
nmhlkynmx
nmanmxlky
N
k
N
l
lk
N
m
N
nlk
UMCPENEE631Slides(crea
tedbyM.Wu2001/20
04)
-
7/27/2019 Lecture UnitaryTransforms
16/32
2-D Separab le Unitary Trans fo rms
Restrict to separable transform
ak,l(m,n) = ak(m) bl(n) , denote this as a(k,m) b(l,n)
Use 1-D unitary transform as building block
{ak(m)}kand {bl(n)}lare 1-D complete orthonormal sets of basis vectors
use as row vectors to obtain unitary matrices A={a(k,m)} & B={b(l,n)}
Apply to columns and rows Y = A X BT
often choose same unitary matrix asA andB (e.g., 2-D DFT)
For square NxN imageA: Y = A X ATX = AHY A*
For rectangular MxN imageA: Y = AMX ANTX = AM
HY AN*
Complexity ~ O(N3)
May further reduce complexity if unitary transf. has fast implementation
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
17/32
Basis Images
X = AHY A* => x(m,n) = kl a*(k,m)a*(l,n) y(k,l)
RepresentXwith NxN basis images weighted by coeff. Y
Obtain basis image by setting Y={(k-k0, l-l0)} & getting X
{ a*(k0,m)a*(l0,n) }m,n in matrix form A*
k,l= a*
ka
l
*T ~ a*k
is kth column vector of
AH
trasnf. coeff. y(k,l) is the inner product ofA*k,lwiththe image
(Jains e.g.5.1, pp137)
2
1
2
12
1
2
1
A
43
21X
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
18/32
UMCPENEE631Slides(cre
atedbyM.Wu2001)
Common Unitary Transform s and BasisImages
DFT DCT Haar transform K-L transform
See also: Jains Fig.5.2 pp136
-
7/27/2019 Lecture UnitaryTransforms
19/32
2-D DFT
2-D DFT is Separable
Y = F X FX = F* Y F*
Basis images Bk,l= (ak)(al
)T
where ak= [ 1 WN-k WN
-2k WN-(N-1)k]T/N
1
0
1
0
1
0
1
0
),(1
),(
),(1
),(
N
k
N
l
mk
N
nl
N
N
m
N
n
mk
N
nl
N
WWlkYN
nmX
WWnmXN
lkY
UMCPENEE631Slides(crea
tedbyM.Wu2001/20
04)
-
7/27/2019 Lecture UnitaryTransforms
20/32
Summary and Review (1)
1-D transform of a vector
Represent an N-sample sequence as a vector in N-dimension vector space
Transform
Different representation of this vector in the space via different basis
e.g., 1-D DFT from time domain to frequency domain
Forward transform In the form of inner product
Project a vector onto a new set of basis to obtain N coefficients
Inverse transform
Use linear combination of basis vectors weighted by transform coeff.
to represent the original signal
2-D transform of a matrix
Rewrite the matrix into a vector and apply 1-D transform
Separable transform allows applying transform to rows then columns
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
21/32
Summary and Review (1) contd
Vector/matrix representation of 1-D & 2-D sampled signal
Representing an image as a matrix or sometimes as a long vector
Basis functions/vectors and orthonomal basis
Used for representing the space via their linear combinations
Many possible sets of basis and orthonomal basis
Unitary transform on inputx ~ A-1 = A*T
y = A xx = A-1 y = A*Ty = ai*Ty(i) ~ represented by basis vectors{ai
*T}
Rows (and columns) of a unitary matrix form an orthonormal basis
General 2-D transform and separable unitary 2-D transform 2-D transform involves O(N4) computation
Separable: Y = A X AT= (A X) AT ~ O(N3) computation
Apply 1-D transform to all columns, then apply 1-D transform to rows
UM
CPENEE631Slides(createdbyM.Wu2001)
-
7/27/2019 Lecture UnitaryTransforms
22/32
Optimal Trans form
UM
CPENEE631Slides(createdbyM.Wu2004
)
-
7/27/2019 Lecture UnitaryTransforms
23/32
Optimal Transfo rm
Recall: Why use transform in coding/compression?
Decorrelate the correlated data
Pack energy into a small number of coefficients
Interested in unitary/orthogonal or approximate orthogonal transforms
Energy preservation s.t. quantization effects can be better understood
and controlled
Unitary transforms weve dealt so far are data independent
Transform basis/filters are not depending on the signals we are processing
What unitary transform gives the best energy compaction and decorrelation?
Optimal in a statistical sense to allow the codec works well withmany images
Signal statistics would play an important role
UM
CPENEE631Slides(createdbyM.Wu2004
)
-
7/27/2019 Lecture UnitaryTransforms
24/32
Review : Correlation A fter a Lin earTransform
Consider an Nx1 zero-mean random vectorx
Covariance (autocorrelation) matrixRx = E[ x xH]
give ideas of correlation between elements
Rxis a diagonal matrix for if all N r.v.s are uncorrelated
Apply a linear transform tox: y = Ax What is the correlation matrix fory ?
Ry = E[ y yH] = E[ (Ax) (Ax)H] = E[ A x xHAH]
= A E[ x xH] AH= A Rx AH
Decorrelation: try to search forA that can produce a decorrelatedy (equiv. a
diagonal correlation matrixRy)
UM
CPENEE631Slides(createdbyM.Wu2004
)
-
7/27/2019 Lecture UnitaryTransforms
25/32
K-L Trans form (Princ ipal ComponentAnalys is)
Eigen decomposition of Rx: Rx uk= kuk
Recall the properties of Rx Hermitian (conjugate symmetric RH= R);
Nonnegative definite (real non-negative eigen values)
Karhunen-Loeve Transform (KLT)
y = UHxx = U y with U = [ u1, uN]
KLT is a unitary transform with basis vectors in U being the
orthonormalized eigenvectors of Rx
UH Rx
U = diag{1
, 2
, , N
} i.e. KLT performs decorrelation
Often order{ui} so that 12 N
Also known as the Hotelling transform or
the Principle Component Analysis (PCA)
UMCPENEE631Slides(crea
tedbyM.Wu2001/20
04)
-
7/27/2019 Lecture UnitaryTransforms
26/32
Propert ies o f K-L Trans form
Decorrelation
E[ y yH]= E[ (UHx) (UHx)H]= UHE[ x xH] U= diag{1, 2, , N}
Note: Other matrices (unitary or nonunitary) may also decorrelate
the transformed sequence [Jains e.g.5.7 pp166]
Minimizing MSE under basis restriction
If only allow to keep m coefficients for any 1m N, whats the
best way to minimize reconstruction error?
Keep the coefficients w.r.t. the eigenvectors of the first m largesteigenvalues
Theorem 5.1 and Proof in Jains Book (pp166)
UMCPENEE631Slides(crea
tedbyM.Wu2001/20
04)
-
7/27/2019 Lecture UnitaryTransforms
27/32
KLT Basis Restr ic t ion
Basis restriction
Keep only a subset of m transform coefficients and then performinverse transform (1 m N)
Basis restriction error: MSE between original & new sequences
Goal: to find the forward and backward transform matrices to minimize therestriction error for each and every m
The minimum is achieved by KLT arranged according to the
decreasing order of the eigenvalues of R
UM
CPENEE631Slides(c
reatedbyM.Wu2004
)
-
7/27/2019 Lecture UnitaryTransforms
28/32
K-L Trans form for Images
Work with 2-D autocorrelation function
R(m,n; m,n)= E[ x(m, n) x(m, n) ] for all 0m, m, n, nN-1
K-L Basis images is the orthonormalized eigenfunctions ofR
Rewrite images into vector form (N2x1)
Need solve the eigen problem forN2xN2 matrix! ~ O(N6)
Reduced computation for separableR
R(m,n; m,n)= r1(m,m) r2(n,n)
Only need solve the eigen problem for twoNxNmatrices ~ O(N3
) KLT can now be performed separably on rows and columns
Reducing the transform complexity from O(N4) to O(N3)
UM
CPENEE631Slides(c
reatedbyM.Wu2001
)
-
7/27/2019 Lecture UnitaryTransforms
29/32
Pros and Cons of K -L Trans form
Optimality
Decorrelation and MMSE for the same# of partial coeff.
Data dependent
Have to estimate the 2nd-order statistics to determine the transform
Can we get data-independent transform with similar performance?
DCT
Applications
(non-universal) compression
pattern recognition: e.g., eigen faces
analyze the principal (dominating) components
UM
CPENEE631Slides(c
reatedbyM.Wu2001
)
-
7/27/2019 Lecture UnitaryTransforms
30/32
Energy Compact ion o f DCT vs. KLT
DCT has excellent energy compaction for highlycorrelated data
DCT is a good replacement for K-L
Close to optimal for highly correlated data
Not depend on specific data like K-L does
Fast algorithm available
[ref and statistics: Jains pp153, 168-175]
UM
CPENEE631Slides(c
reatedbyM.Wu2004
)
-
7/27/2019 Lecture UnitaryTransforms
31/32
Energy Compaction of DCT vs. KLT (contd)
Preliminaries
The matrices R, R-1, and R-1 share the same eigen vectors
DCT basis vectors are eigenvectors of a symmetric tri-diagonal matrixQc
Covariance matrixR of 1st-order stationary Markov sequence with has an
inverse in the form of symmetric tri-diagonal matrix DCT is close to KLT on 1st-order stationary Markov
For highly correlated sequence, a scaled version ofR-1 approx. Qc
UM
CPENEE631Slides(c
reatedbyM.Wu2004
)
-
7/27/2019 Lecture UnitaryTransforms
32/32
Summary and Review on Uni tary Trans form
Representation with orthonormal basis Unitary transform
Preserve energy
Common unitary transforms
DFT, DCT, Haar, KLT
Which transform to choose?
Depend on need in particular task/application
DFT ~ reflect physical meaning of frequency or spatial
frequency
KLT ~ optimal in energy compaction
DCT ~ real-to-real, and close to KLTs energy compaction
UM
CPENEE631Slides(c
reatedbyM.Wu2001
)