Beruflich Dokumente
Kultur Dokumente
2/55
The Quantization Problem
Encoder mapping
Map a range of values to a codeword
Irreversible mapping
If source is analog → A/D converter
Decoder mapping
Map the codeword to a (fixed) value representing the range
Knowledge of the source distribution can help us pick a
better value representing each range
If output is analog → D/A converter
Informally, the encoder mapping is called the
quantization process, and the decoder mapping is
called the inverse quantization process
3/55
Quantization Examples
3-bit Quantizer
Encoder (A/D) Decoder (D/A)
4/55
Quantization Function
5/55
Quantization Problem Formulation
Input:
X – random variable
fX(x) – probability density function (pdf)
Output:
{bi}i = 0..M decision boundaries
{yi}i = 1..M reconstruction levels
Discrete processes are often approximated by
continuous distributions
Example: Laplacian model of pixel differences
If source is unbounded, then the first and the last decision
boundaries = ±∞ (they are often called “saturation” values)
6/55
Quantization Error
bi −1
(x − yi )2 f X dx
i =1
8/55
Quantized Bitrate with VLC
9/55
Optimization of Quantization
Rate-optimized quantization
Given: Distortion constraint σq2 ≤ D*
Find: { bi }, { yi } binary codes
Such that: R is minimized
Distortion-optimized quantization
Given: Rate constraint R ≤ R*
Find: { bi }, { yi } binary codes
Such that: σq2 is minimized
10/55
Uniform Quantizer
11/55
Midrise vs. Midtread Quantizer
Midrise Midtread
12/55
Uniform Quantization of Uniform Source
13/55
Alternative MSQE Derivation
14/55
The SNR of Quantization
σ s2 (2 X max ) 2 12
10 log10 2 = 10 log10 ⋅ 2
σ ∆
q 12
(2 X max )
2
12 2
= 10 log10 2
= 10 log10 M
12 2 X max
M
= 20 log10 2 n = 6.02n dB.
15/55
Example: Quantization of Sena
16/55
Quantization of Non-uniform Sources
17/55
Optimal ∆ that minimizes MSQE
x – Q(x)
18/55
Granular error
Solving for Optimum Step Sizes
19/55
Overload/Granular Regions
20/55
Variance Mismatch Effects (1/2)
21/55
Variance Mismatch Effects (2/2)
22/55
Distribution Mismatch Effects
24/55
Forward Adaptive Quantization (FAQ)
N i =0
25/55
Speech Quantization Example (1/2)
26/55
Speech Quantization Example (2/2)
27/55
FAQ Refinement
29/55
Jayant Quantizer
30/55
Output Levels of 3-bit Jayant Quantizer
31/55
Example: Jayant Quantizer
M0 = M4 = 0.8, M1 = M5 = 0.9
M2= M6 = 1.0, M3= M7 = 1.2, ∆0= 0.5
Input: 0.1, –0.2, 0.2, 0.1, –0.3, 0.1, 0.2, 0.5, 0.9, 1.5
32/55
Picking Jayant Multipliers
∑l P
k =0
k k = 0, → γ and lk are chosen, Pk is known.
33/55
Example: Ringing Problem
34/55
Jayant Quantizer Performance
35/55
Non-uniform Quantization
36/55
pdf-optimized Quantization
∫ f X ( x)dx
X j–1 j
b j −1
38/55
Lloyd-Max Algorithm (2/3)
Continue the process until all {bj} and {yj} are found
39/55
Lloyd-Max Algorithm (3/3)
40/55
Example: pdf-Optimized Quantizers
(9.24) (7.05)
(12.18) (9.56)
(14.27) (11.39)
41/55
Mismatch Effects
42/55
Companded Quantization (CQ)
43/55
Example: CQ (1/2)
2x if − 1 ≤ x ≤ 1
c( x) = 23x + 43 x >1 .
2x − 4 x < −1
3 3
44/55
Example: CQ (2/2)
2x if − 2 ≤ x ≤ 2
c −1 ( x) = 32x − 2 x>2 .
3x + 2 x < −2
2
45/55
Remarks on CQ
47/55
Vector Quantization
48/55
Why Vector Quantization (1/2)?
useless
49/55
Why Vector Quantization (2/2)?
x2
Vj
Yj
x1
51/55
Codebook Design
The set of quantizer output points in VQ is called the codebook
of the quantizer, and the process of placing these output points
is often referred to as the codebook design
† The idea is the same as the scalar quantization problem in Stuart P. Lloyd, “Least Squares Quantization in PCM,” IEEE
Trans. on Information Theory, Vol. 28, No. 2, March 1982. 52/55
The Linde-Buzo-Gray Algorithm
53/55
Example: Codebook Design
Initial state Final state
54/55
Impact of Training Set
55/55