Subsampling Method for Robust Estimation of Regression Models
New Multiresolution and subband coding - Stanford University · 2000. 10. 26. · Filtering...
Transcript of New Multiresolution and subband coding - Stanford University · 2000. 10. 26. · Filtering...
-
1
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 1
Multiresolution and subband coding
n Predictive (closed-loop) pyramids
n Open-loop (“Laplacian”) pyramids
n Subband coding
n Perfect reconstruction filter banks
n Quadrature mirror filter banks
n Discrete Wavelet Transform (DWT)
n Embedded zerotree wavelet (EZW) coding
n Transform coding as a special case of subband coding
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 2
Interpolation error coding, I
Interpolator
InterpolatorSubsampling
• • •
• • •
Input picture
Reconstructed picture
Q
Q
+
+
-
- +
+
+
+
InterpolatorSubsampling
• • • • • •
Coder includes
Decoder
Sample encoded in current stage
Previously coded sample
-
2
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 3
Interpolation error coding, II
signals to be encoded
original image
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 4
Predictive pyramid, I
Interpolator
Subsampling
Q
Q
+
+
-
- +
+
+
+
InterpolatorInterpolator
Filtering
Filtering
Subsampling
Input picture
Reconstructed picture
• • •
• • •• • • • • •
Coder includes
Decoder
Sample encoded in current stage
-
3
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 5
Predictive pyramid, II
Number of samples to be encoded =
1+1N
+1
N 2+ ...
=
N
N −1 x number of original image samples
subsampling factor
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 6
Predictive pyramid, III
original image
signals to be encoded
-
4
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 7
Comparison:interpolation error coding vs. pyramid
n Resolution layer #0, interpolated to original size for display
Interpolation Error Coding Pyramid
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 8
Comparison:interpolation error coding vs. pyramid
n Resolution layer #1, interpolated to original size for display
Interpolation Error Coding Pyramid
-
5
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 9
Comparison:interpolation error coding vs. pyramid
n Resolution layer #2, interpolated to original size for display
Interpolation Error Coding Pyramid
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 10
Comparison:interpolation error coding vs. pyramid
n Resolution layer #3
=(original)
Interpolation Error Coding Pyramid
-
6
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 11
Open-loop pyramid (Laplacian pyramid)
Interpolator
Interpolator
Subsampling
Q+
Q+ -
-
Filtering
Filtering
Interpolator
InterpolatorSubsampling
Input picture
Reconstructed picture
• • • • • • • • • • • •
Receiver Transmitter
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 12
When multiresolution coding was a new idea . . .
This manuscript is okay if compared to some of the weaker papers.[. . .] however, I doubt that anyone will ever use this algorithm again.
Anonymous reviewer of Burt and Adelson‘s original paper, ca. 1982
-
7
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 13
Embedded multiresolution coding
Φ (ω)uu
ω
preserved spectrum
reconstruction errorspectrum low rate
Φ (ω)vv
reconstruction errorspectrum high rate
Low frequency components have to berefined for optimal resolution/noise balance
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 14
Optimum bit allocation for open-loop pyramid
i=0
i=1
i=2
variance
number of samples
ri =12
log2σi
2
Ni+ f (R) bits /sample
depends on bit-rate R, but not not layer i
Assume layer m.s.e distortion-rate function:
di ri
= σ i2 2
−2ri
-
8
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 15
ri =
12
log2
ασi2(N
i−1− Ni)σ i−1
2 (Ni − Ni+1)
bits/sample
i=0
i=1
i=2 variance number of samples
power transfer
factor
does not depend on R !!
remaining bits
r0 = R −
Ni rii=1
L−1
∑
Nii=0
L−1
∑bits /sample
Optimum bit allocation for closed-loop pyramid
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 16
Two-layer open- vs. closed-loop pyramid
30
32
34
36
38
0.2 0.4 0.6 0.8 1.0
PS
NR
[dB
]
Rate R [bits/sample]
28
30
32
34
36
0.2 0.4 0.6 0.8 1.0
PS
NR
[dB
]
Rate R [bits/sample]
Theory Actual Coder
closed loop
open loop open loop
closed loop
-
9
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 17
Subband coding
n Number of degrees of freedom is preserved:
“Critically sampled filter bank”
n Perfect reconstruction filter bank required
Input signal G (ω)k Q k
k Q k
k Q k
• • •
Analysis filterbank
Synthesis filterbank
Reconstructed signal
F (ω)
G (ω)
G (ω)
1
M
1
F (ω)0 0 0 0
1F (ω)1
MMM
+
Transmitter Receiver
1k0
+1k1
+ ...+1
kM= 1
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 18
Two-channel filterbank
n Aliasing cancellation if :
2 2
2 2
F (ω)0
F (ω)1
G (ω)
G (ω)
0
1
X(ω)X(ω)
G0 (ω ) = F1 (ω + π )
−G1(ω ) = F0 (ω + π)
ˆ X (ω) =12
F0 (ω )G0 (ω) + F1 (ω )G1(ω )[ ]X(ω )
+12
F0 (ω + π )G0 (ω) + F1(ω + π )G1 (ω)[ ]X(ω + π )Aliasing
-
10
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 19
Example: two-channel filter bank with perfect reconstruction
n Analysis filter impulse responses:
l Lowpass band:
l Highpass band:
n Synthesis filter impulse responses:
l Lowpass band:
l Highpass band:
n Frequency responses:
|G |0
|F |0
1|F |
1|G |
π2
π
1
2
00
FrequencyF
requ
ency
res
pons
e
14
−1, +2, +6, +2, −1( )
14
+1, −2, +1( )
14
+1, +2, −6, +2, +1( )
14
+1, +2, +1( )
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 20
Quadrature mirror filters (QMF)
n QMFs achieve aliasingcancellation by choosing
n Highpass band is the mirror image of the lowpass band in the frequency domain
F1(ω ) = F0 (ω + π )
= −G1 (ω ) = G0 (ω + π )
frequency ω
Example:16-tap QMF filterbank
-
11
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 21
Cascaded analysis / synthesis filterbanks
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 22
Discrete Wavelet Transform
n Recursive application of a two-band filter bank to thelowpass band of the previous stage yields octave band splitting:
n Same concept can be derived from wavelet theory: Discrete Wavelet Transform (DWT)
frequency
-
12
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 23
2-d Discrete Wavelet Transform
ωx
ωy
ωx
ωy
ωx
ωy
ωx
ωy
ωx
ωy
ωx
ωy
ωx
ωy
ωx
ωy
ωx
ωy
...etc
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 24
2-d Discrete Wavelet Transform example
-
13
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 25
Embedded zero-tree wavelet algorithm
X X X XX X X XX X X XX X X X
X X X X
X
X
X X X X
X X X XX X X XX X X XX X X X
X
X X X X
X X X XX X X XX X X XX X X X
n Idea: Conditional coding of all descendants (incl. children)
n Coefficient magnitude > threshold: significant coefficients
n Four cases
l ZTR: zero-tree, coefficient and all descendants are not significant
l IZ: coefficient is not significant, but some descendants are significant
l POS: positive significant
l NEG: negative significant
„Parent“
„Children“
„Descendants“
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 26
Embedded zero-tree wavelet algorithm (cont.)
n For the highest bands, ZTR and IZ symbols are merged into one symbol Z
n Successive approximation quantization and encodingl Initial „dominant“ pass
• Set initial threshold T, determine significant coefficients• Arithmetic coding of symbols ZTR, IZ, POS, NEG
l Subordinate pass• Refine magnitude of coefficients found significant so far by one bit
(subdivide magnitude bin by two)• Arithmetic coding of sequence of zeros and ones.
l Repeat dominant pass• Set previously found significant coefficients to zero• Decrease threshold by factor of 2, determine new significant
coefficients• Arithmetic coding of symbols ZTR, IZ, POS, NEG
l Repeat subordinate and dominate passes, until bit budget is exhausted.
-
14
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 27
Embedded zero-tree wavelet algorithm (cont.)
n Decoding: bitstream can be truncated to yield a coarser approximation: „embedded“ representation
n Further details: J. M. Shapiro, „Embedded image coding using zerotrees of wavelet coefficients,“ IEEE Transactions on Signal Processing, vol. 41, no. 12, pp. 3445-3462, December 1993.
n Enhancement SPIHT coder: A. Said, A., W. A. Pearlman, „A new, fast, and efficient image codec based on set partitioning inhierarchical trees,“ IEEE Transactions on Circuits and Systemsfor Video Technology, vol. 63 , pp. 243-250, June 1996.
n JPEG-2000 standard similar to SPIHT
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 28
Subband coding vs. transform coding, I
n Transform coding is a special case of subband coding with:l Number of bands = order of transform N
l Subsampling factor K = N
l Length of impulse responses of analysis/synthesis filters � N
n Filters used in subband coders are not in general orthogonal.
-
15
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 29
Subband coding vs. transform coding, II
re-order
Original image
8x8 DCT
8-channel subband decomposition (using DCT filters)
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 30
Frequency response of a DCT of order N=8
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 0
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 1
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 2
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 3
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 4
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 5
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 6
0 π/2 π −30
−20
−10
0
10
Frequency
Fre
quen
cy r
espo
nse
[dB
]
i = 7
-
16
Bernd Girod: EE368b Image and Video Compression Multiresolution & Subband no. 31
Summary:multiresolution and subband coding
n Resolution pyramids with subsampling 2:1 horizontally and vertically
n Predictive pyramids: quantization error feedback („closed loop“)
n Transform pyramids: no quantization error feedback („open loop“)
n Pyramids: overcomplete representation of the image
n Critically sampled subband decomposition: number of samples not increased
n Quadrature mirror filters: aliasing cancellation
n Discrete Wavelet Transform = cascaded 2:1 subband splits
n Exploit statistical dependencies across subbands by zero-trees
n Transform coding is a special case of subband coding