Minimizing Relative Entropy in Hierarchical Predictive Coding
predictive coding 07 - HHIiphome.hhi.de/wiegand/assets/pdfs/DIC_predictive_coding_07.pdf · Thomas...
Transcript of predictive coding 07 - HHIiphome.hhi.de/wiegand/assets/pdfs/DIC_predictive_coding_07.pdf · Thomas...
Thomas Wiegand: Digital Image Communication Predictive Coding - 1
Predictive CodingPredictive Coding
Prediction
Prediction in Images
Principle of Differential Pulse Code Modulation
(DPCM)
DPCM and entropy-constrained scalar
quantization
DPCM and transmission errors
Adaptive intra-interframe DPCM
Conditional Replenishment
Thomas Wiegand: Digital Image Communication Predictive Coding - 2
PredictionPrediction
Prediction is difficult – especially for the future. Mark Twain
Prediction: Statistical estimation procedure where
future random variables are estimated/predicted from
past and present observable random variables.
Prediction from previous samples:
Optimization criterion
Optimum predictor:
ˆ S 0 = f (S1,S2,...,SN ) = f (S)
min})],...,,({[})ˆ{( 2
210
2
00 == NSSSfSESSE
)},...,(|{ˆ2100 N
SSSSES =
Thomas Wiegand: Digital Image Communication Predictive Coding - 3
The optimum predictor
can be stored in a table (Pixels: 8 bit size 28N)
Optimal linear prediction (zero mean, Gaussian RVs)
Optimization criterion
Optimum linear predictor is solution of
In case RS=E{SSt} is invertible
StructureStructure
)},...,(|{ˆ2100 N
SSSSES =
Sat
NNSaSaSaS =+++= ...ˆ
22110
}){(})ˆ{( 2
0
2
00 Sat
SESSE =
}{ 0
t
S
tSE SRa =
}{ 0
1SRa SE
S=
Thomas Wiegand: Digital Image Communication Predictive Coding - 4
Prediction in Images: Intra-frame PredictionPrediction in Images: Intra-frame Prediction
Past and present observable random variables are
prior scanned pixels within that image
When scanning from upper left corner to lower right
corner:
1-D Horizontal prediction: A only
1-D Vertical prediction: C only
Improvements for 2-D approaches (requires line store)
B
A
DC
X
),(),(),(ˆ
)0,0(),(
0
2
1
qypxsqpayxs
qp
P
Pp
Q
q
== =
43421
Thomas Wiegand: Digital Image Communication Predictive Coding - 5
Prediction Example: Test PatternPrediction Example: Test Pattern
s[x,y]
uV[x,y]=
s[x,y]-0.95 s[x,y-1]
uH[x,y]=s[x,y]-0.95 s[x-1,y]
uD[x,y]=s[x,y]-0.5(s[x,y-1]+ s[x-1,y])
Thomas Wiegand: Digital Image Communication Predictive Coding - 6
Prediction Example: CameramanPrediction Example: Cameraman
s[x,y] uH[x,y]=s[x,y]-0.95 s[x-1,y]
uV[x,y]=
s[x,y]-0.95 s[x,y-1]
uD[x,y]=s[x,y]-0.5(s[x,y-1]+ s[x-1,y])
Thomas Wiegand: Digital Image Communication Predictive Coding - 7
Change of Histograms: CameramanChange of Histograms: Cameraman
0 50 100 150 200 2500
500
1000
1500
2000
2500
3000
Image signal Prediction error
signal (diag. pred.)
Can we use prediction for compression ?
Yes, if we reproduce the prediction signal at the decoder
-50 0 500
0.5
1
1.5
2x 104
Thomas Wiegand: Digital Image Communication Predictive Coding - 8
Differential Pulse Code ModulationDifferential Pulse Code Modulation
Reconstruction error =
Prediction error Reconstruction quantization error
Coder
Decoder
quantizerentropy
coder
predictor
input
predictor
entropy
decoder output
channel
+
+
-s
's
e
s 's
s
ses ˆ'' += qeess == ''
'e
'e
sse ˆ=
Thomas Wiegand: Digital Image Communication Predictive Coding - 9
DPCM and QuantizationDPCM and Quantization
Prediction is based on quantized samples
Stability problems for large quantization errors
Prediction shapes error signal (typical pdfs: Laplacian,
generalized Gaussian)
Simple and efficient: combine with entropy-
constrained scalar quantization
Higher gains: Combine with block entropy coding
Use a switched predictor
• Forward adaptation (side information)
• Backward adaptation (error resilience, accuracy)
DPCM can also be conducted for vectors
• Predict vectors (with side information)
• Quantize prediction error vectors
Thomas Wiegand: Digital Image Communication Predictive Coding - 10
0 1 2 3 4 5 6 70
5
10
15
20
25
30
35
Panter & Dite App
Entropy-Constrained Opt.
R(D*), =0.9 DPCM & ECSQ
R [bits]
Comparison for Gauss-Markov Source: Comparison for Gauss-Markov Source: =0.9=0.9
Linear predictor orderN=1, a=0.9Entropy-Constrained
Scalar Quantizer with
Huffman VLC
Iterative design
algorithm applied
SNR [dB]
=10 log10
2
D
Thomas Wiegand: Digital Image Communication Predictive Coding - 11
K=511, H=4.79 b/p K=15, H=1.98 b/p K=3, H=0.88 b/pK...number of reconstruction levels, H...entropy
DPCM with Entropy-Constrained Scalar QuantizationDPCM with Entropy-Constrained Scalar Quantization
from: Ohm
Example: Lena, 8 b/p
Thomas Wiegand: Digital Image Communication Predictive Coding - 12
Transmission Errors in a DPCM SystemTransmission Errors in a DPCM System
For a linear DPCM decoder, the transmission error
response is superimposed to the reconstructed signalS'
For a stable DPCM decoder, the transmission error
response decays
Finite word-length effects in the decoder can lead to
residual errors that do not decay (e.g., limit cycles)
from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 13
Error rate p=10-3.1D pred., hor. aH=0.95 1D pred., ver. aV=0.95 2D pred.*,aH=aV=0. 5
from: Ohm
Transmission Errors in a DPCM System IITransmission Errors in a DPCM System II
Example: Lena, 3 b/p (fixed code word length)
Thomas Wiegand: Digital Image Communication Predictive Coding - 14
Inter-frame Coding of Video SignalsInter-frame Coding of Video Signals
Inter-frame coding exploits:
• Similarity of temporally successive pictures
• Temporal properties of human vision
Important inter-frame coding methods:
• Adaptive intra/inter-frame coding
• Conditional replenishment
• Motion-compensating prediction (in Hybrid Video Coding)
• Motion-compensating interpolation
from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 15
Principle of Adaptive Intra/Inter-Principle of Adaptive Intra/Inter-FFrame DPCMrame DPCM
Predictor is switched between two states:
for moving or changed areas.
Intra-frame prediction Inter-frame prediction (previous frame
for moving or changed areas. prediction) for still areas of the picture.
S 22 S23 S 24
S 21 S20 S 25
FRAME N - 1
S 2 S 3 S4
S 1 S 0
FRAME N
S 22 S 23 S24
S 21 S 20 S25
FRAME N - 1
S 2 S3 S 4
S 1 S0
FRAME N
40 ms
''''ˆ44332211
SaSaSaSaS +++=20'ˆ SS
inter= from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 16
Intra/Inter-Intra/Inter-FFrame DPCM: Adaptation Strategies, Irame DPCM: Adaptation Strategies, I
Feedforward Adaptation
Interframe
predictor
Intraframe
predictor
VLC VLCQuantizer
Interframe
predictor
Intraframe
predictor
Predictor
adaptation
+
-
Intra/inter-frame
Switching
information
Coder Decoder
s e
s
inters
intras
's
s
's'e 'e
inters
intras
from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 17
Intra/Inter-Intra/Inter-FFrame DPCM: Adaptation Strategies, IIrame DPCM: Adaptation Strategies, II
Feedback Adaptation
Inter-frame
predictor
Intra-frame
predictor
VLCVLCQuantizer+
-
Coder Decoder
Predictor
adaptation
Inter-frame
predictor
Intra-frame
predictor
Predictor
adaptation
from: Girod
s e
s
inters
intras
's
'e 'e
s
inters
intras
's
Thomas Wiegand: Digital Image Communication Predictive Coding - 18
Principle of a Conditional Replenishment CoderPrinciple of a Conditional Replenishment Coder
Still areas: repeat from frame store
Moving areas: transmit address and waveform
SIGNAL
INPUT
SEGMENTER
(MOVEMENT
DETECTOR)
CODING,
ADDRESSING,
BUFFERING
BUFFERING,
DECODING,
ADDRESSING
FRAME DELAY
(1 PICTURE MEMORY)FRAME DELAY
(1 PICTURE MEMORY)
TRANSMISSION
CHANNEL
SIGNAL
OUTPUT
Coder Decoder
from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 19
Change DetectionChange Detection
Example of a pel-oriented change detector
Example of a block-oriented change detector
ABS
Average
of 3x3
window
Eliminate isolated
points or pairs of
points
+
-
Current
frame
Previous
frame
Decision
changed/
unchanged
Threshold
ABSAccumulate
over NxN
blocks
+
-
Current
frame
Previous
frame
Decision
changed/
unchanged
Threshold
from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 20
The "Dirty Window" EffectThe "Dirty Window" Effect
Conditional replenishment scheme with change detection threshold
set too high leads to the subjective impression of looking through a
dirty window.
BackgroundMoving
Area
picked up
by change
detector
Moving areas
missed by
change detector
from: Girod
Thomas Wiegand: Digital Image Communication Predictive Coding - 21
SummarySummary
Prediction: Estimation of random variable from past or
present observable random variables
Optimal prediction
Optimal linear prediction
Prediction in images: 1-D vs. 2-D prediction
DPCM: Prediction from previously coded/transmitted
samples (known at coder and decoder)
DPCM and quantization
DPCM and transmission errors
Adaptive Intra/Inter-frame DPCM: forward adaptation
vs. backward adaptation
Conditional Replenishment: Only changed areas of
image are transmitted