You are on page 1of 93

Table of Content

Digital Image Processing Image Transform (Fourier) / Quantization / Quality


O. Le Meur

olemeur@irisa.fr
http://www.irisa.fr/temics/staff/lemeur/

Univ. of Rennes 1

December 6, 2010

Table of Content

1 2 3 4 5

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

Introduction

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

1 2 3 4 5

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

Introduction

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is a signal? This is a function of one or more variables (time, distance, temperature...).

A signal carries information!!


Signal processing is required to obtain a more desirable form. Types of signals: Natural (lighting) or synthetic (from a lab); one dimension or more; Continuous/discrete.

Introduction

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Classication of signals
Continuous-time vs Discrete-time signals: Analog vs Digital:
A continuous-time signal contains a value for all real numbers along the time axis; A discrete-time signal only has values at equally spaced intervals along the time axis. Analog signal corresponds to a continuous set of possible function values; Digital signal corresponds to a discrete set of possible function values. The past, present and future of a deterministic signal are known with certainty (its

Deterministic vs Random Signals:

values can be determined by a mathematical expression, rule or table). Otherwise, it is called Random and its properties are explained by statistical techniques.

Periodic vs Aperiodic: periodic signals repeat with some period T ; Finite vs Innite length.

Introduction

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Classication of signals

Sampling = discrete time (or space)


6

Quantization = discrete amplitude

Introduction

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

There are many dierent ways to describe or represent a signal mathematically. Most of the time a linear combination of elementary time functions. These functions are called basis functions. Let the set of basis functions be dened by: {p }, (dierent posssibilities for the range of p )

x (t )

an n (t )

Notations and reminder

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

a continuous signal x (.); a discrete signal x [.]; a matrix A.

Notations and reminder


(s);

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

f is the frequency (Hz); w = 2f is the pulsation (radian/s); T x Lp (R) means that


+

1 = f

is the period

For instance, the space L2 (R) is + composed of the functions x with a nite energy |x (t )|2 dt < +;

|x (t )|p dt < +.

Euler's formulas: exp (j ) = cos + j sin ;

exp(j ) = cos + j sin exp(j ) = cos j sin exp(j ) + exp(j ) = 2 cos exp(j ) exp(j ) = 2j sin

From Wikepedia.
9

Reminder

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

x sinc (x ) = sin ; x

Inner product:

x, y x, y

=
+

x (t ) y (t )dt x (n) y (n)

n=

Condition of orthogonality for real basis function: t2 0 n (t )m (t )dt = m t1

n=m n=m

where m is a constant (m = 1, orthonormal). Condition of orthogonality for complex basis function: t2 0 n=m n (t ) m (t )dt = m n = m t1 where m is a constant (m = 1, orthonormal).
10

Image transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

1 2

Introduction Image transformation Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation Fourier transformation Time sampling Discrete Fourier Transform

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

3 4 5

11

Image transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

What is a transformation? T
im x y

[ , ] IM [u , v ]

im is the original image; IM is the transformed image; x , y (or u, v ) represents the spatial coordinates of a pixel.
Goal of a transformation The goal of a transformation is to get a new representation of the incoming picture. This new representation can be more convenient for a particular application or can ease the extraction of particular properties of the picture.

12

Image transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

There exist 3 types of transformation: Point to point transformation: The output value at a specic coordinate is dependent only on one input value but not necessarily at the same coordinate; Local to point transformation: The output value at a specic coordinate is dependent on the input values in the neighborhood of that same coordinate; Global to point transformation: The output value at a specic coordinate is dependent on all the values in the input image.

Note that the complexity increases with the size of the considered neighborhood...
13

Point to point transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

Geometric transformation such as rotation, scaling... Example of rotation:

Scalar quantization LUT (Look-Up-Table)

Remark: a point to point transformation doesn't require extra memory...


14

Example of local to point transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

Note that local to point transformation is also called neighborhood operators. Examples of local to point transformation: Linear transformation: For instance, a tranverse lter (FIR) IM [x , y ] = i j h(i , j )im[x i , y j ], where h is the 2D impulse response. Non-linear transformation such as adaptive linear ltering and rank ltering (median)... Morphological ltering (erosion, dilatation...) based on the denition of a structuring element. Remark: Non linear lters are not reversible in general.

15

Global to point transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

One of the most important transformation is the Fourier transform that gives a frequential representation of the signal. Fourier series (development in real term): Every signal x (x L2 (t1 , t1 + T )) can be decomposed into a linear combination of sinusoidal and cosinusoidal component functions.

x (t )

b0 +

n =1

an cos (2nt ) +

n =1

bn sin(2nt )

an and bn are used to weight the inuence of each frequency component.

16

Global to point transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

Fourier series (development in complex term):

x (t )
with,

Cn exp(j 2n t ) T n=
+

Cn
Remark:

T t1

t 1 +T

x (t )exp(j 2n t )dt T

an +jbn , n 1 ; C = C ; jbn Cn = an n n 2 , n 1 ; Cn = 2 Note that if the function to be represented is also T -periodic, then t1 is an arbitrary choice. Two popular choices are t1 = 0, and t1 = T /2.

17

Global to point transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Presentation 3 kinds of transformation Point to point transformation Local to point transformation Global to point transformation

18

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

1 2 3

Introduction Image transformation Fourier transformation Denition Trivial properties Parseval Formula Properties Some examples Time sampling Discrete Fourier Transform

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

4 5

19

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

Fourier transformation in L2 (R)


x t

( ) X (f )

Inverse Fourier Transform:


+
x t

( )

X f

( )exp (j 2 ft )df

Fourier Transform:
+
X f

( )

x t exp

( )

(j 2 ft )dt

provided that x L2 (R) and X L2 (R).


20

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

From the output of the Fourier transform, we dene: The frequency spectrum: Real (X (f )) + jImg ((X (f ))) The Fourier transform of a function produces a frequency spectrum which contains all of the information about the original signal, but in a dierent form. Magnitude spectrum: |Real (X (f )) + jImg (X (f ))|

Img (X (f )) Phase spectrum: Arctg Real (X (f ))


Power spectrum: Real (X (f ))2 + Img (X (f ))2

21

Fourier transformation
Linearity:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

ax (t ) ax1 (t ) + bx2 (t )

aX (f ) aX1 (f ) + bX2 (f )

F Complex conjugate: x (t ) X (f )

x (t ) x (t )

=
+

X (f )exp(j 2ft )df

X (f )exp(j 2ft )df

F F In the same way, x (t ) X (f ) and x (t ) X (f ).

22

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

Hermitian symmetry: if x (t ) R, we deduce X (f ) = X (f )

X (f ) X (f ) X (f )

= =

+ +

x (t )exp(j 2ft )dt x (t )exp(j 2ft )dt

x (t )exp(j 2ft )dt

Given that x (t ) R, we have x (t ) = x (t ) that implies X (f ) = X (f ).

23

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

Note that, as we use the space L2 (R), the inner product of x L2 (R) and y L2 (R) is given by:

x, y
The norm is then given by:
+

x (t )y (t )dt
|x (t )|2 dt

x
=

2=

x, x
+

|x (t )|2 dt

x (t )x (t )dt x (t ) X (f )
+ +

=
+

X (f )exp(j 2ft )df dt x (t )exp(j 2ft )dt df

=
+

=
+

|x (t )|2 dt

X (f )X (f )df
|X (f )|2 df

24

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

|x (t )|

+
dt

|X (f )|

df

This formula (Parseval's theorem or energy conservation) proves that the energy is conserved by the Fourier transform. Loosely, the sum (or integral) of the square of a function is equal to the sum (or integral) of the square of its transform.

25

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

F Translation in time/space domain: x (t t0 ) exp(j 2t0 f )X (f )

x (t t0 ) x (t t0 ) x (t t0 )

+ +

x (t t0 )exp(j 2tf )dt x (k )exp(j 2(k + t0 )f )dk


+

k = t t0

F F

exp(j 2t0 f )

x (k )exp(j 2fk )dk

Displacement in time or space induces a phase shift proportional to frequency and to the amount of displacement.
F Frequency shift: x (t )exp (j 2f0 t ) X ( f f0 )

Displacement in frequency multiplies the time/space function by a unit phasor which has angle proportional to time/space and to the amount of displacement. Amplitude modulation.
26

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

F Convolution in time/space domain: x1 (t ) x2 (t ) X1 (f ) X2 (f ) + + +

x1 (t ) x2 (t ) x1 (t ) x2 (t ) x1 (t ) x2 (t )

x1 ( )x2 (t )d exp(j 2ft )dt


+

x1 ( )

x2 (t )exp(j 2ft )dt d


X2 (f )

Translation

x1 ( )exp(j 2f )d

F Convolution in frequency domain: x1 (t ) x2 (t ) X1 (f ) X2 (f )

These properties mean that if two functions are multiplied in one domain, then their Fourier transforms are convolved in the other domain.

27

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

F Time and frequency scaling: x (t )

||

f) X(

x (t ) x (t ) x (t )

+ +

x (t )exp(j 2tf )dt x (k )exp(j 2k f ) dk


+

k = t

F F

| |

1
| |

x (k )exp(j 2k f )dk

Multiplication of the scale of the time/space reference frame changes by the factor scales the frequency axis of the spectrum of the function.

All the previous properties are what makes Fourier transforms so useful and practical.
28

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

Translated Dirac, x (t ) = (t T )

x (t ) = Arect (t |T )

X (f ) X (f )

x (t )exp(j 2tf )dt

exp(j 2Tf )

X (f )

x (t )exp(j 2tf )dt exp(j 2tf )dt

+T 2

If T = 0, X (f ) = 1. The Fourier transform is constant whatever the frequency. x (t ) = 1, X (f ) = (f )


29

X (f )

A [exp(jfT ) exp(jfT )] j 2f ATsinc (fT )

T 2

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

Ideal low-pass lter in frequency domain: X (f ) = rect (f | F 2 ) selects low F F frequencies over 4 , 4 . The impluse response of the lter is calculated with the inverse Fourier integral:

x (t )

X (f )exp(j 2ft )df exp(j 2ft )df

= =

+F 4

sin x (t )

= =

1 exp(j 2 F t ) exp(j 2 F t ) j 2t 4 4 exp(j ) exp(j ) 2j F sinc ( tF ) 2 2

F 4

30

Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition Trivial properties Parseval Formula Properties Some examples

31

Time sampling

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

1 2 3 4

Introduction Image transformation Fourier transformation Time sampling What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice Discrete Fourier Transform

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

32

Time sampling

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

To be more exible and to gain in accuracy... Digital algorithms are now ... everywhere from television, CD, embedded electronic (car...). Whether sound recordings or images, most discrete signals are obtained by sampling an analog signal. But, be careful, we have to follow some constraints in order to be able to reconstruct the analog signal. The simplest way to discretize an analog signal x is to record its sample values at intervals T : x (t ) can be approximated by {x (nT )}nZ .

33

Denition

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

Diracs are essential to make the transition from functions of real variable to discrete sequences. A Dirac can be dened as:
+

(t )x (t )dt = x (0)

(t ) =

0
+
+

if t = 0 for t = 0

(t )dt = 1.

Remark: { : (t )} Z is an orthonormal basis:


(t n), (t m) =

0 1

if m = n for m = n .

34

Properties of a Dirac

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

(t ) Scaling and symmetry: (t ) = || +

(t )dt (t )

= = (t ) ||

(k )

dk
| |

Note that if = 1, (t ) = (t ). Product: x (t ) (t T ) = x (T ) (t T ) Convolution: x (t ) (t T )

x (t ) (t T )

=
+

x ( )(t T )d x ( )( (t T ))d

x (t T )

The eect of convolving with a translated Dirac is to time-delay x (t ), by the same amount (copy of the signal x (t ) at the position T ).
35

Dirac comb: x (t ) =

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

nZ (t nT ). Notation: T (t )

The product of a comb with a signal x (t ) conducts to a sampling:

The convolution of a comb with a signal x (t ) conducts to a periodization of x (t ):

36

FT of a Dirac

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

F Dirac: x (t ) = (t ), (t ) 1, constant function whatever the frequency

X (f ) X (f )

(t )exp (j 2 tf )dt

exp(j 20f ) = 1

F Translated Dirac: x (t ) = (t T ), (t T ) exp(j 2Tf )

X (f ) X (f )

(t T )exp (j 2 tf )dt

exp(j 2Tf )

F Constant signal: x (t ) = 1, 1 (f ). The FT of a constant signal is a weight at the null frequency.

37

Dirac comb: x (t ) = nZ (t nT ) This is a periodic function, since x (t + T ) = x (t ), t . Its Fourier serie is given by:

x (t ) Cn x (t )

n= T 1 2

Cn exp(j 2n t ) T
(t )exp (j 2 n

T
1

T 2 +

nt )dt = 1 T T

T n=

exp(j 2n t ) T exp(j 2n t ) T n= T
1
+

By using the Fourier Trasnform,


F (x (t )) F (x (t )) F (x (t )) (t nT ) = F

T n= T n= T nZ
1 1
+

exp(j 2n t ) T
+

=
F

exp(j 2t (f

nZ

(f

n) T

n ))dt T

Sampling

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

Sampling a signal is equivalent to multiplying it by a grid of impulses.

39

Sampling

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

Let x (t ) be a continous function, its sampled version y (t ) is given by:

y (t )

= = =

x ( t ) T ( t ) x (t ) (t nT )
nZ nZ

x (t )(t nT ) x (nT )(t nT )

nZ

y (t )

1 ). where T (t ) is the sampling function (sampling frequency T

40

Fourier Transform of a sampled signal


Let x (t ) be a continous function, its sampled version y (t ), y (t ) = x (t ) T (t )

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

y (t )
F (y (t )) F (y (t )) F (y (t )) F (y (t )) F (y (t ))

x (t )

nZ

(t nT ) (t nT ) (t nT ) (f

F x (t )

nZ nZ nZ

X (f ) F

X (f ) T T
1 1
nZ nZ

n ) T

= =

X (f ) (f X (f

n ) T

n ) T

n . For n = 0, the spectrum is the primary spectrum and X (f ) and all its replicas at positions T for n = 0 we have copies of the spectrum.
41

The Fourier representation of the sampled function is a superposition of the Fourier transform

Fourier Transform of a sampled signal


Let x (t ) be a continous function, its sampled version y (t ), y (t ) = x (t ) T (t ) Let X (f ) be the Fourier transform of x (t ) and Y (f ) the Fourier transform of its sampled version y (t ). 1 n

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

Y (f ) =

T nZ

X (f

The sampling operation replicates the primary spectrum periodically in the Fourier domain.

42

Fourier Transform of a sampled signal

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

43

Fourier Transform of a sampled signal

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

Aliasing eect... We won't be able to recover the original signal with a good quality. The transformation is no more inversible! The sampling frequency is too small.

44

Shannon-Nyquist theorem

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

To ensure a perfect reconstruction of a signal x (t ) (to avoid aliasing eect), the sampling frequency fe must verify:
f

e > 2 fmax

where fmax is the maximal frequency of the spectrum X (f ).


The periodization doesn't provoke an aliasing eet.

45

Whittaker-Shannon theorem

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

If x (t ) is bandlimited to a frequency range [fmax , fmax ], that is, its Fourier transform X (f ) is zero for |f | > fmax , then x (t ) can be completely specied by samples taken at the Nyquist sample rate of 2fmax .

46

Whittaker-Shannon theorem

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

To recover the signal x (t ), we can use a low-pass lter in frequency domain. For instance, the ideal low-pass lter in the frequency domain is given by G (f ) = Te rect (f /fe )

X (f ) x (t )

= =

Y (f ) rect (f /fe )
+

k =

x (kTe )sinc ((t kTe )fe )

47

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

What is the objective? Reminder on Dirac Sampling Sampling Theorem In practice

The sampling of a physical signal always induces an aliasing eect, even with an ideal sampler. Paley-Wiener theorem indeed showed that there is no bandlimited signal having a nite energy... In many cases, the spectrum of the signal is not perfectly known and can be corrupted by noise. It is thus required to lter the analog signal before the sampling operation.
1 2 3

Pre-ltering to obtain a bandlimited signal; Sampling by checking the value of the sampling frequency; Low-pass ltering to retrieve the signal.

48

Discrete Fourier Transform

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

1 2 3 4 5

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

49

Denition of 1D DFT

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

The discrete Fourier transform (DFT) is the transform of a nite sequence. Since the time-domain is discrete the spectrum is periodic The DFT of a sequence of N complex numbers x0 , . . . , xN 1 is given by: N 1 x [n]exp (j 2 nk ), k = 0, . . . , N 1 1 X [k ] = N n =0 N
X [k ] = x , ek , where ek (n) = exp(j 2N nk ). The family ek (n) = exp (j 2 N nk ) 0k <N is an orthogonal basis of the space of signals of period N .

The IDFT of a sequence of N complex numbers X0 , . . . , XN 1 is given by: 1 2 x [n ] = N n=0 X [k ]exp (j N nk ), n = 0, . . . , N 1 The DFT computes the X [k ] from the x [n], while the IDFT shows how to compute the x [n] as a sum of sinusoidal components X [k ]exp ( j 2 N kn) with frequency k /N cycles per sample.
50

Denition of 1D DFT

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

X [k ]

N n =0

1 N 1

x [n]exp(j 2 nk ) N

1 can be removed or replaced by 1 . In the rst case, the The weighting coecient N N 1 whereas in the second case it is equal weighting coecient of the IDFT is equal to N 1 . to N

What is important is that DFT IDFT = Id .

51

Discrete bi-dimensional Fourier transformation


im k l

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

[ , ] is a discrete picture having a size

Inverse Fourier Transform:


N 1 M 1
IM u v exp j

M.

im k l

[ , ]

=
u=0 v =0

[ , ]

( 2 (

k N

l M

))

Fourier Transform:
IM u v

[ , ]

1
NM

N 1 M 1
im k l exp

[ , ]

k =0 l =0

(j 2 (

k N

l M

))

u v

The unit is cycles per unit length (pel 1 ).


52

u , [0, 1]; = 0, . . . , N 1, k = N k v , [0, 1]. = 0, . . . , M 1, l = M l

Discrete bi-dimensional Fourier transformation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

A frequential decomposition in terms of exponential basis images ...


im k l

[ , ]

im k l

[ , ]

[0, 0] exp (j 2 (0 k + 0 l )) + [0, 1] exp (j 2 (0 k + l )) + IM [1, 0] exp (j 2 (k + 0 l )) + . . . = IM [0, 0] + IM [0, 1] (cos (2 l ) + jsin (2 l )) + IM [1, 0] (cos (2 k ) + jsin (2 k )) + IM [1, 1] (cos (2 (k + l )) + jsin (2 (k + l ))) + . . . =
IM IM

IM [u , v ] are the weighting coecients of the spatial frequencies present in the signal.

53

Some remarks:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

1 A picture having a size of N M is a linear combination of N M exponential basis images!!!! 2 The DFT of a picture yields a complex picture of the same size;

54

Some remarks:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

3 IM [u , v ] and im[k , l ] are periodic of innite extent. Period = N for k , u . Period = M for l , v :

The image imt is bandlimited and is spatially sampled, for example, with a rectangular array of N M pels:

im[k , l ]

N 1 M 1
=

i =0 j =0

imt (k , l )(k iTK )(l jTL )

55

Some remarks:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

56

Translation

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

Translation in the frequency domain


im k l exp

[ , ]

2
N

(u0 k + v0 l )

DF T

IM u

[ u0 , v v0 ]

Translation in the spatial domain


im k

[ k0 , l l 0 ]

DF T

IM u v exp

[ , ]

2
N

(k0 u + l0 v )

The multiplication by an exponential function in a representation modies the origin in the other representation.

57

Average value

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

im

1 N 1 M 1
NM

im k l

[ , ] (j 2 (
k N l M

IM u v

[ , ] [0, 0]

1 1

k =0 l =0 N 1 M 1
im k l exp

[ , ]

))

NM

k =0 l =0 N 1 M 1
im k l

IM

[ , ]

NM

k =0 l =0

im

1
NM

IM

[0, 0]

IM

[0, 0] is proportional to the average value of the picture.

58

Separable transformation = separable basis decomposition


The 2D DFT can be computed as the separable product of two one-dimensional discrete Fourier transforms. N 1 M 1 IM [u, v ] = 1 im[k , l ]exp(j 2( k u + l v )) N M NM k =0 l =0

k u + l v )) Initialy the family ek ,l [u , v ] = exp (j 2( N is used, but we M 0k <N ,0l <M can also use the family {ek (u )el (v )}0k <N ,0l <M .

IM [u, v ] IM [u, v ] IM [u, v ]


{el }0l <M .

1 N 1

= =

exp(j 2 k u) MDFT1D (im[k , l ]) N N k =0 NDFT1D MDFT1D (im[k , l ])

1 N 1

exp(j 2 k u) N N k =0

1 M 1

M l =0

im[k , l ]exp(j 2 l v ) M

We start by computing the Fourier coecients on the image rows int the basis A second transform is applied on the columns of the transformed images in the basis {ek }0k <N .

Separable transformation = separable basis decomposition

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

60

Separable transformation = separable basis decomposition


For a picture having a size N N , there are three ways to compute its DFT: DFT 1D: N 2 multiplications and N (N 1) additions, O(N 2 )

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

x [n]exp(j 2 nk ) N n =0 N DFT 2D: N 2 N 2 multiplications and N 2 (N 1)2 additions, O(N 4 ) X [k ]


=

1 N 1

im[k , l ]exp(j 2( k u + l v )) N k =0 l =0 N N DFT 2D separable transform: 2N N 2 multiplications and 2N N (N 1)2 additions, O(N 3 ) IM [u, v ]
=

1 N 1 N 1

61

Examples

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

Remark: the spatial frequency amplitude spectra of natural images are well described by a power law X (f ) = fk . Estimated average values of are ranged over 0.9 and 1.2. The power always falls with spatial frequency.

From E.P. Simoncelli and B.A.Olhausen, Natural Image Statistics and Neural Representation, Annun. Rev. Neuroscience, 2001. 62

Power spectrum of a natural image (solid line) compared with f12 (dashed line).

Examples

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

From R.M. Balboa, N.M. Grzywacz, Power spectra and distribution of contrasts of natural images from dierent habitats, Vision Research, 43, pp. 2527-2537, 2003.

The spectra may be more or less anisotropic.

63

Examples

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Denition of 1D DFT Denition of 2D DFT Remarks Properties Separable transformation Examples

From left hand-side to right: original picture, spectrum, contour lines.

Remark: For display purpose, a logarithm is applied on the spectrum.

64

Scalar Quantization

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

1 2 3 4 5

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform

Scalar Quantization Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer Distortion/quality assessment Conclusion

7 8

65

Principle

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

The quantization is a process to represent a large set of values with a smaller set.
Scalar quantization:

Q:X x
X C

C = {yi , i = 1, 2, ...N } Q ( x ) = yi

N is the number of quantization level;

could be continue (ex: R) or discret; is always discret (codebook,dictionnary); card (X ) > card (C ); As x = Q (x ), we will lost some information (lossy compression).

66

Uniform quantization

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

In the uniform quantization, the quantization step size is xed, no matter what the signal amplitude is. Denition: output y The quantization thresholds are uniformly distributed: i {1, 2, ..., N }, ti ti +1 = {yi }, representative levels The output values are the center of the quantization interval: t +t i {1, 2, ..., N }, yi = i 2i +1 input x t 1 t2 t3 t4 t 6 t7 t8 t9 Example of the nearest neighborhood quantizer (see on the left): {ti }, decision levels x + 0 .5 . Q (x ) = y2 y1
67

Uniform quantization

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

A mid-step quantizer Zero is one of the representative levels yk

A uniform quantization is completely dened by the number of levels, the quantization step and if it is a mid-step or mid-riser quantizer.

output y

A mid-riser quantizer Zero is one of the decision levels tk

output y

t1 t2 t3 t 4 y2 y1

t6 t7 t 8 t 9

input x

t1 t2 t 3 t4 y2 y1

t6 t7 t 8 t9

input x

Usually, a mid-riser quantizer is used if the number of representative levels is even and a mid-step quantizer if the number of level is odd.
68

Uniform quantization with dead zone


Uniform quantization output y

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

Uniform quantization with dead zone output y

t1 t2 t3 t 4 y2 y1

t6 t7 t 8 t 9

input x

t1 t2 t 3 t4 y2 y1

t6 t7 t 8 t9

input x

Deadzone

Interest: To remove small coecients by favouring the zero value. Increase the coding eciency with a small visual impact.
69

Example of an uniform quantization


Example (Original picture quantized to 8, 7, 6, 4 bits/pixels)

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

(a) Original (8 bits per pixel)

(b) 7 bpp ( = 2)

(c) 6 bpp ( = 4)

(d) 4 bpp ( = 8)

70

Optimal quantization

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

Optimal quantization An optimal quantization of a random variable X having a probability distribution p (x ) is obtained by a quantizer that minimises a given metric: Linf -norm: D = max |X Q (X )|; L1 -norm: D = E [|X Q (X )|]; L2 -norm: D = E (X Q (X ))2 called the Mean-Square Error (MSE). This is the most used. Considering the MSE, we have: max p (x )(x Q (x ))2 dx ; if the random variable X is continue: D = xxmin n if the random variable X is discret: D = k =1 p (xk )(xk Q (xk ))2 .

71

Example

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

Example (Quantization error) Hypothesis: Uniform quantization with a quantization step ; p(x ) is a uniform probability distribution of a random variable X ; N is the number of representative levels. xmax xmin

xmin

xmax xmax xmin N

Quantization step: =

1 1 Probability distribution: p (x ) = xmax xmin = N

72

Example

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

Example (Quantization error)

= = = =

xmax xmin

p(x )(x Q (x ))2 dx


/2

N N
2

/2 /2

p(x )(x 0)2 dx , 0 the mid-point


1

/2 N

x 2 dx

12

73

Optimal quantizer in a nutshell

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

Uniform quantizer is not optimal if the source is not uniformly distributed. Optimal quantizer To nd the decision levels {ti } and the representative levels {yi } to minimize the distortion D . To reduce the MSE, the idea is to decrease the bin's size when the probability of occurrence is high and to increase the bin's size when the probability is low. For N representative levels and with a probability density p (x ), the distortion is given by:

xmax

xmin N 1 tk +1 k =1 t k

p(x )(x Q (x ))2 dx p(x )(x yk )2 dx

74

Optimal quantizer in a nutshell

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

The optimal {ti } and {yi } satisfy:


D ti

=0

D and yi = 0

Lloyd-Max quantizer y +y D = 0 ti = i 2 i +1 . ti ti is the midpoint of yi and yi +1 . ti ti 1 p (x )xdx D = 0 y = . ti i yi ti 1 p (x )dx yi is the centroid of the interval [ti 1 , ti ]. given the {ti }, we can nd the corresponding optimal {yi }. given the {yi }, we can nd the corresponding optimal {ti }. How can we nd the optimal {ti } and {yi } simultaneously?

75

Lloyd-Max algorithm in a nutshell

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

The Lloyd-Max algorithm is an algorithm for nding the representative levels {yi } and the decision levels {ti } to meet the previous conditions, with no prior knowledge. Principle of the iterative process
1 The iterative process starts for (0) ( 0) levels y1 , ..., yN .

k = 0 with a set of initial values for the representative


(k ) +y

y 2 New values for decision levels are determined ti(k +1) = i


3 Compute the distortion

4 Depending on the stopping criteria ( (k ) < ), stop the process or update the ti(k +1) xp(x )dx (k +1) t 1 representative levels yi(k +1) = i and go back to step 2. ti(k +1) p(x )dx (k +1) ti 1

D (k ) and the relative errors (k ) ;

(k ) i +1

76

Uniform, Semi-uniform and optimal quantizer


Suppose we have the following 1D discrete signal:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

= {0, 0.01, 2.8, 3.4, 1.99, 3.6, 5, 3.2, 4.5, 7.1, 7.9}

Example (Uniform quantizer: N = 4, Mid-riser and = 2) output y ti T = {t0 = 0, t1 = 2, t2 = 4, t3 = 6, t4 = 8} ri R = {r0 = 1, r1 = 3, r2 = 5, r3 = 7} 7


5 3 1 0 2 4 6 8 Check the denition of a uniform quantizer...

input x

Quantized vector X = {1, 1, 3, 3, 1, 3, 5, 3, 5, 7, 7} (MSE = 0.42)


77

Uniform, semi-uniform and optimal quantizer


Suppose we have the following 1D discrete signal:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

= {0, 0.01, 2.8, 3.4, 1.99, 3.6, 5, 3.2, 4.5, 7.1, 7.9}

Example (Semi-uniform quantizer: N = 4, Mid-riser and = 2) output y ti T = {t0 = 0, t1 = 2, t2 = 4, t3 = 6, t4 = 8} ri R = {r0 = 2/3, r1 = 3.25, r2 = 4.75, r3 = 7.5} 7
5 3 1 0 2 4 6 8

input x

Quantized vector X = {2/3, 2/3, 3.25, 3.25, 2/3, 3.25, 4.75, 3.25, 4.75, 7.5, 7.5} (MSE = 0.31)
78

Uniform, semi-uniform and optimal quantizer


Suppose we have the following 1D discrete signal:

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

= {0, 0.01, 2.8, 3.4, 1.99, 3.6, 5, 3.2, 4.5, 7.1, 7.9}

Example (Lloyd-Max Algorithm) Given that T = {0, 1.5, 3.87, 6.125, 8} and R = {0.005, 2.998, 4.75, 7.5}, the quantized vector is X = {0.005, 0.005, 2.998, 2.998, 2.998, 2.998, 4.75, 2.998, 4.75, 7.5, 7.5} (MSE=0.18).

output y
Uniform Semi-uniform Optimal 7 5 3 1 0
79

xi

input x

Uniform vs optimal quantizer


Example (N=5)

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Principle Uniform quantization Optimal quantization, Lloyd-Max algorithm Examples: uniform, semi-uniform and optimal quantizer

(a) Original

(b) Uniform

(c) Optimal

(d) Histo.
80

(e) Decision levels

(f) Decision levels

Distortion/quality assessment

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

1 2 3 4 5

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform

6 7

Scalar Quantization Distortion/quality assessment Taxonomy Signal delity Perceptual metric Examples Conclusion

81

Taxonomy

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

Distortion/quality metrics can be divided into 3 categories:


1 2 3

Full-Reference metrics (FR) for which both the original and the distorted images are required (benchmark, compression) ; Reduced-Reference metrics (RR) for which a description of the original and the distorted image is required (network monitoring); No-Reference (NR) metrics for which the original image is not required (network monitoring).

Each category can be divided into two subcategories: metrics based on signal delity and metrics based on properties on the human visual system.

82

Peak Signal to Noise Ratio (PSNR)


The PSNR is the most popular quality metric. This simple metric just calculates the mathematical dierence between each pixel of the degraded image and the original image. PSNR: Let I and D the original and impaired images, respectively. These images having a size of M pixels are coded with n bits. n 1)2 ) PSNR = 10log10 ((2MSE dB, with the Mean Squared Error MSE =
( x ,y ) (

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

I (x ,y )D (x ,y ))2 . M

A high value indicates that the amount of impairment is small. A small value indicates that there is a strong degradation.

83

Peak Signal to Noise Ratio (PSNR)


The PSNR is not always well correlated with the human judgment (MOS Mean Opinion Score). The reason is simple: this metric does not take into account the properties of the human visual system. Example

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

(a)

(b)

(c) Original

(d) Original+uniform noise

(a) and (b) from [?]: impact of Gabor patch on our perception.

84

Peak Signal to Noise Ratio (PSNR)


Example (These three pictures have the same PSNR...)

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

(a) Original

(b) Contrast stretched

(c) Blur
85

(d) JPEG

Metric based on the error visibility

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

For this type of metric, the behavior of the visual cells are simulated: I Perceptual Color Space PSD CSF Masking

Pooling

Quality

Perceptual Color Space

PSD

CSF

Masking

PSD: Perceptual Subband Decomposition (Wavelet, Gabor, Fourier); CSF: Contrast Sensitivity Function.

86

Metric based on the error visibility

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

Example VDP (Visible Dierences Predictor) [Daly,93]:

WQA (Wavelet-based Quality Assessment) [Ninassi et al.,08a]

VQM (Video Quality Model) [Pinson et al.,04]...


87

Metric based on the structural similarity


SSIM standing for Structural Similarity index. Image degradations are considered here as perceived structural information loss instead of perceived errors [Wang et al.,04a]. Let I and D the original and the degraded images, respectively.

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

S (x , y ) = l (x , y )

c (x , y )

s (x , y )

(1)

The luminance comparison measure l (x , y ) = The contrast comparison measure c (x , y ) =

2x y 2 + 2 x y
xy x y

2x y 2 2 x + y

The structural comparison measure s (x , y ) =

SSIM (x , y ) =

(2x y + C1 )(2xy + C2 ) 2 2 2 (2 x + y + C1 )(x + y + C2 )

SSIM 1, the best quality and SSIM 0 indicates a poor quality.


88

Metric based on the structural similarity

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

All pictures have the same MSE. (a) Original image; (b) Contrast-stretched image, MSSIM = 0.9168; (c) Mean-shifted image, MSSIM = 0.99; (d) JPEG, MSSIM = 0.6949; (e) Blurred image, MSSIM = 0.7052; (f) Impluse noise, MSSIM = 0.7748. Extracted from [Wang et al.,04a].
89

Example of distortion maps


Example

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Taxonomy Signal delity Perceptual metric Examples

(a) Original

(b) Degraded

(c) MSE
90

(d) WQA

(e) SSIM

Distortion/quality assessment

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

1 2 3 4 5

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform

6 7 8

Scalar Quantization Distortion/quality assessment Conclusion

91

Conclusion

Introduction Image transformation Fourier transformation Time sampling Discrete Fourier Transform Scalar Quantization Distortion/quality assessment Conclusion

Image Transformation: global to point; Fourier transformation; Time sampling with Shannon's theorem; Scalar quantization, a lossy process; Quality and distortion.

92

Suggestion for further reading... S. Daly. The visible dierences predictor: An algorithm for the assessment of image delity. Digital Images and Human Vision, pp. 179-206, 1993, MIT Press. A. Ninassi, O. Le Meur, P. Le Callet, and D. Barba. On the performance of human visual system based image quality assessment metric using wavelet domain. Proc. SPIE Human Vision and Electronic Imaging XIII., Vol. 6806, pp. 680610-12, 2008. M. Pinson and S. Wolf. A new standardized method for objectively measuring video quality. IEEE Trans. Broadcasting, Vol. 50, N. 3, pp. 312-322, 2004. Z. Wang, A. C. Bovik, H.R. Sheikh and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. on Image Processing, Vol. 13, pp. 600-612, 2004.

92

You might also like