Stanford EE398B April 19, 2007 Lossless VideoCoding · Better Compression. Takamura: ... – Fixed,...

Lossless Lossless VVideoideo CodingCoding

Dr. Seishi TakamuraDr. Seishi TakamuraNTT CorporationNTT Corporation

JapanJapan

StanfordEE398B

April 19, 2007

Takamura: Lossless Video Coding EE398B 2007.4.19 2/54

“Lossy” vs. “Lossless”“Lossy” vs. “Lossless”

Lossy image/video coding– Input != Output

– High compression (~1/10 - 1/1000)

– Possibly be“perceptually lossless”

Lossless image/video coding– Input == Output

– Low compression (~1/2 -1/3)

– “mathematically lossless”

– This is what’s meant “lossless” in this talk


Is It Worthy to Sacrifice Compression Ratio?Is It Worthy to Sacrifice Compression Ratio?

Yes, for many practical reasons– Medical diagnosis, fine art archive, military purposes– Image/video analysis, automatic product examination

The observer is not a human — “perceptually lossless”does not make sense

– Contribution transmission, studio applicationsSubject to be re-encoded/edited repeatedly — coding

error accumulates, if lossyand Yes for some emotional reasons– A CG artist said, “Just do not lose even one bit from my

artwork!”– Provides an upper-bound of image quality (can feel at

ease)– Fed up with drawing R-D curves, evaluating Lagrange cost

function, subjective impairment measurement, HVS…


OutlineOutline

Lossless image coding– Typical schemes and technology– FYI: Lossless image codecs comparison

Lossless video coding– Lossless variants of lossy H.264/AVC

– Using reversible transforms– Using DCT residual packing

– FYI: Lossless video codecs comparison

Summary


OutlineOutline





Summary


Typical Lossless Image CodecsTypical Lossless Image Codecs

NoteAlgorithm

Context-based linear prediction + nonlinear correction

Median prediction

reversible (5,3) DWTLinear prediction

Linear prediction——

Prediction Entropy codingMethod

Huffman / Arithmetic codingCALIC

based on “LOCO” method

context-based Golomb(run-length in flat region)

JPEG-LS

Arithmetic coding (EBCOT)JPEG 2000*lossless modeHuffmanJPEG*

DEFLATE (LZ77+Huffman)PNG

LZWTIFF

up to 256 colorsLZWGIF

Bet

ter C

ompr

essi

on


Prediction SchemesPrediction Schemes

Linear prediction– Fixed, encoder-optimized and online-update predictors– Using causal pixels (extrapolative DPCM)– Using non-causal pixels (interpolative DPCM)– Block transform

– Reversible DWT– Reversible DCT– R/G/B decorrelation

Median predictionMarkov model– Build a stochastic model of the image– Encode the pixel occurrence probability: Prob(x|c)– cf. minimum MSE prediction = Σx(Prob(x|c)*x)

Pixelwise residual (“Lossy plus lossless”) approach– Consider another lossy coder (e.g., JPEG) as a predictor– Entropy code the pixel-wise residual

(And motion compensation for video)

Decoded(lossy)

Originalimage

Residualdata– =

xc

x

x


OutlineOutline





Summary


ARTest Codec Comparison (full-color)ARTest Codec Comparison (full-color)

2.3

2.4

2.5

2.6

2.7

2.8

2.9

3

BMFRKIMUHIC NK

ERIArH

anGeL

UHArcACE

LSPSBC DC RK

PNGCrush

LOCO

Com

pres

sion

Rat

io

Top 14 codecs out of 26Average of 627 full-color (24-bit) images, weighted by image size (#pixels)Also provides encoding/decoding time comparison

25%

[The art of lossless data compression, 2002]

Base ofJPEG-LS


MRP — A Gray-Scale Lossless Image CodecMRP — A Gray-Scale Lossless Image Codec

Context-based 2D-linear prediction– M=20-56 contexts (predictors)– P=30-72 pixels from current frame– All M·P coefficients are optimized for minimum rate,

not for minimum MSE, and downloadedEncoding time: 10-25 minutes per frameDecoding time: 0.2 seconds per frame

[Matsuda et al., 2005]


1.8

1.9

2.0

2.1

2.2

MRP

BMF

TMW

Glicba

wls

CALICa

JPEG-LS

CALICh

JPEG 20

00

Com

pres

sion

Rat

io

SUT Codec Comparison (gray-scale)SUT Codec Comparison (gray-scale)

http://itohws03.ee.noda.sut.ac.jp/~matsuda/mrp/

Average of 13 gray-scale (8-bit) imagesWeighted by image size (#pixels)CALICa: CALIC with arithmetic codingCALICh: CALIC with Huffman coding

9.5%

1.6%

Note: Conditions are different from previous comparison


OutlineOutline



– Using reversible transforms [Takamura, ICASSP2005]

– Using DCT residual packing– FYI: Lossless video codecs comparison

Summary


Lossless Coding of H.264 FRExtLossless Coding of H.264 FRExt

H.264 FRExt skips the transform and quantization

Inter/Intraprediction

4x4/8x8transform

Quanti-zation

Zigzag scan

Entropycoding

Entropy remains the same

thruFREｘｔ

entropy reduction

entropy reduction

entropy reduction

Video


Enhancement of FRExtEnhancement of FRExt

Instead of no-operation (FRExt), we apply reversible transform to reduce entropyWe compared four transforms

Inter/Intraprediction

4x4/8x8transform

Quanti-zation

Zigzagscan

Entropycoding

Reversibletransform

Aiming at entropy reduction

thruFREｘｔ

Video


Reversible Transforms TestedReversible Transforms Tested

4x4 Transform of H.264 (reversible if not quantized)Piecewise Scaled TransformReversible DCTLinear Prediction– Non-adaptive version– Adaptive version


4x4 Reversible Transform

H.264 4x4 Transform4x4 Reversible Transform

H.264 4x4 Transform

Obviously reversible (if not quantized)Signal correlation is reduced…However, the signal space is expanded– 2.66 bit/pixel increment– Corrupts coding performance

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−=

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

112121112111

1121

122111112112

1111

PONMLKJIHGFEDCBA

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−=

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

112121112111

1121

122111112112

1111

PONMLKJIHGFEDCBA

Transform Transform coefficientscoefficients

Intra/Inter

prediction residual



Piecewise Scaled Transform4x4 Reversible Transform

Piecewise Scaled Transform

Decompose H.264’s 4x4 transform into equivalent four 2x2 transforms,and compensate the expansion at each stage

1 11 –1

1 11 –1

1 11 –1

2 11 –2

1 11 –1

1 11 –1

1 11 –1

2 11 –2

2

2

2

5

2 Divide by 2 and round5 Divide by 5 and round

H.264 transform(decomposed)

Piecewise scaled transformOutput is still reversible!



Reversible DCT4x4 Reversible Transform

Reversible DCT

⎣ ⎦5.0)( where)/)((

)2/)(()/)(()4/)((

21303

32102

21301

32100

+=+−−=

+−−=−+−=+++=

xxRxxCxxRX

xxxxRXCxxRxxX

xxxxRX

This is parallel to– H.264’s 4x4 transform, when C=2– Parallel to mathematical DCT, when

(recommended by the paper)We tried , 3 and 3.521+=C

21+=C

[Komatsu, Sezaki, 1996]



Linear Prediction (non-adaptive)


Linear Prediction (non-adaptive)

Horizontal: Apply following transform to four rows

x0 x1 x2 x3

αΗx0 αΗx1 αΗx2

x0 R(x1 -αΗ x0) R(x2 -αΗ x1) R(x3 -αΗ x2)

Vertical: transform the columns in the same manner, using αVαH = αV = 0.7 works good for I framesαH = αV = 0.4 works good for P and B frames

Input

Output

pred. pred. pred.



Linear Prediction (adaptive)


Linear Prediction (adaptive)

Optimal coefficient α is image dependentIn energy concentration sense, autocorrelation factor gives an optimal α. (But not known to the decoder unless explicitly transmitted)We estimate α from neighboring decoded blocks

ρH , ρV : horizontal and vertical autocorrelation factors in a block

ρHρV

codedblock

ρHρV

ρHρV

ρHρV

codedblock

ρHρV

ρHρV

If horizontally correlated If vertically correlated

weightedaverage

codedblock

If not correlated

default αΗ, αVweightedaverage

Highestweight

Highestweight

αΗ, αVαΗ, αV


⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

1000010000100001

FREｘｔ

Transform ComparisonTransform Comparison

H.264Transform ⎥

⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−

122111112112

1111

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−

10/15/15/110/12/12/12/12/112/12/11

1111

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−−−

−−

CC

CC

/111/12/12/12/12/11/1/114/14/14/14/1

⎥⎥⎥⎥

⎦

⎤

⎢⎢⎢⎢

⎣

⎡

−−

−

1000100010001

αα

α

PiecewiseScaled

ReversibleDCT

LinearPrediction

Determinant

4040

11

11

1+1/1+1/CC22

11

Norm of each row

( )10,2,10,2

( )1,1,1,1

( )10/10,1,2/10,2

( )CCCC /22,1,/22,2/1 22 ++

( )222 1,1,1,1 ααα +++

The smaller, the better

More constant, the better

Non-orthogonal

Transform


Simulation ConditionsSimulation Conditions

4 QCIF sequences and 5 CIF sequencesI frame performance– Encoded 100 frames (all intra)– Compared with Motion JPEG 2000 (kakadu codec)

P frame performance– Encoded 100 frames in IPPP… structure– Coding performance is obtained from 99 P frames

B frame performance– Encoded 298 frames in IBBPBBP… structure– Coding performance is obtained from 198 B frames


Results (average of 9 sequences)Results (average of 9 sequences)

1

1.5

2

2.5

3

3.5

I frames P frames B frames

Com

pres

sion

Rat

io

H.264 trans.w/o quant.PiecewisescaledReversible DCTC=1+ 2Reversible DCTC=3Reversible DCTC=3.5FRExt

Linear pred.non-adaptiveLinear pred.adaptiveJPEG 2000

21+=C

Almost no compression

FRExt is not good at intra


Summary for Reversible Transform CodingSummary for Reversible Transform Coding

Compared FRExt with four different reversible transforms– Non-orthogonal transform (linear prediction)

yields better compression than other orthogonal transforms

– Adaptive linear prediction works the best–12% better than FRExt (intra)–3% better than FRExt (inter)–2% worse than JPEG 2000

NoteOur method deserves yet another ~1% improvement [Sano et al, 2006]


OutlineOutline



– Using reversible transforms– Using DCT residual packing [Takamura, ICIP2005]


Summary


MotivationMotivation

H.264/AVC — promising lossy video coding standard– About twice better compression than other standards– More and more H.264-compliant bit streams will be

created and distributedWhat if H.264 bit stream + additional bits = lossless?– Useful for many applications– Good interoperability with existing H.264 systems– Two-layered scalability

Existing technologies– Conventional pixel-wise residual coding: inefficient– H.264 High 4:4:4 Profile (FRExt): single-layered,

incompatibility

H.264Bit stream

EnhancementBit stream

lossless


255

y

0

255

x

2D mappingA

510

B

0

-255

255

Transform

Transform and QuantizationTransform and Quantization

x y

An image(2D array of pixels)

(x,y)・・・

・・・

x (and y) : 8-bit integer

45 deg. rotated & magnified

⎥⎦

⎤⎢⎣

⎡⎥⎦

⎤⎢⎣

⎡−

=⎥⎦

⎤⎢⎣

⎡yx

vu

1111

Quantizationstep size

AB

original

Quantization bin8-b

it bord

er


Recovering Original from Q-Bin Recovering Original from Q-Bin

Inverse transform

Pixel-wise residual

White grid:Impossible signal

xdif

ydif2

1

0

-1

-2-2 -1 0 1 2

20 choices

No conversion(simplified FGS)B

A1 2 3 4 5

5

4

3

2

1

25 choicesNo conversion

(general FGS)

A3.9…

4.1…

B

A lot of choices

1 2 34 5

6 7 89 10

11 12

12 Choices(optimal)

To specify original( )…


H.264’s 4x4 TransformH.264’s 4x4 Transform

Above matrices are all integer-valuedTo have all 16 coefficient values (A to P) = To have exact 4x4 residual values = To have original image!

••••••••••••••••

PONMLKJIHGFEDCBA

–12–211–1–11

–2–1121111

–11–212–1–11

–2–1111121

=

Transformedcoefficients

Key Points:

Intra / Interprediction residuals


Enhancement CodingEnhancement Coding

Scan all the integer grids ( , , ) within 16-dimensional quantization binΩ16=(A|A=Amin…Amax,B|B=Bmin…Bmax,…,P|P=Pmin…Pmax)using 16-nested for-loop to obtain:cases - total number of (possible candidates)index - the serial number of (original image)

Encode index using log2(cases) bits

However…Loop count (=volume of Ω16) becomes 800,000,000,000,000,000,000 when QP=12

It takes 24 million YEARS if each loop is done in 1 nsec, only for one 4x4 block!

1 2 34 5

6 7 89 10

11 12

Amin Amax

Bmax

Bmin

cases=12, index=5


H.264 Basis FunctionsH.264 Basis Functions

A B C D

E F G H

K LP

11111111111111111111111111111111

−−22–111122−−22–111122−−22–111122−−22–111122

11–11–111111–11–111111–11–111111–11–1111

–1122–2211–1122–2211–1122–2211–1122–2211

JIM N O

–22–22–22–22–11–11–11–11

1111111122222222

4422–22–442211–11–22

–22–111122–44–222244

–222222–22–111111–11

11–11–111122–22–2222

22–4444–2211--2222–11

–1122–2211–2244–4422

–1122–221111−−2222–1111−−2222–11

–1122–2211

11–11–1111–111111–11–111111–11

11–11–1111

–22–1111222211–11–222211–11–22

–22–111122

11111111–11–11–11–11–11–11–11–11

11111111

–11–11–11–1122222222

–22–22–22–2211111111

2211–11–22–44–222244

4422–22–44–22–111122

–111111–1122–22–2222

–222222–2211–11–1111

11–2222–11–2244–4422

22–4444–22–1122−−2211


Nifty Properties of H.264 CoefficientsNifty Properties of H.264 Coefficients

For arbitrary input, A+C is always even– One bit reduction of C, if A is known (can be ‘pack’ed

half)– So are A+I, B+J, E+G, M+O etc.

A+C+I+K is always a multiple of four.

A C1111111111111111

1–1–111–1–111–1–111–1–11

2002200220022002

+ =

4004000000004004

A C1111111111111111

1–1–111–1–111–1–111–1–11

+ KI1–1–11

–111–1–111–1

1–1–11

1111–1–1–1–1–1–1–1–1

1111

+ + =


More Nifty PropertiesMore Nifty Properties

B+(C>>1)+(A>>1): Always even– If A and C are known, B can be discretized

with interval of 2.– Same for F+(E>>1)+(G>>1), E+(A>>1)+(I>>1),

G+(C>>1)+(K>>1), etc.

B–J+E–G: Multiple of 42.5(A+C)+2B+D: Multiple of 102.5(A+C+I+K)+2(E+G)+M+O: Multiple of 206.25(A+C+I+K)+5(B+E+G+J)+4F+2.5(D+L+M+O) +2(H+N)+P: Multiple of 100

00010000100001000010

0–4–40–4–404–40440440

000000000000

200020

000000000000000100


Enhanced Enhancement CodingEnhanced Enhancement Coding

For each block1. Entropy-code the coefficient A using

log2(Amax−Amin+1) bits2. ‘Pack’ each of seven coefficients and

entropy-code in the same manner3. For-loop the rest eight coefficients (Ω8)

to find out index and cases4. Entropy-code index using log2(cases) bitsAt 2 and 3, numerical properties are exploited

•••M•K•I•••EDCB

PON•L•J•HGF•••••

•••••••••••••••A

pack

index, casesLoop count=400

(2·1018 times speed-up!)


Quantization and BoundsQuantization and Bounds

BoundsAmin, AmaxBmin, Bmax

….Pmin, Pmax

estimationLevelslevelA … levelP

quantizationCoefficientsA ... P

levelA = (A*α + β) >> γ

x = levelA << γy = x + (1 << γ) – 1

Amin = (x – β + α – 1) / αAmax = (y – β) / α

ThenAmin <= A <= Amax

The bounds can be calculated/shared by both encoder and decoder.


Coding Simulation with JM-BLCoding Simulation with JM-BL

Implemented on JM8.3 (H.264 reference codec)Two mode selection strategies regarding the base layer (BL) and enhancement layer (EL):1. OverallOpt: chooses a mode that minimizes total

(BL+ EL) bit amount2. BaseOpt: chooses a mode that RD-optimizes the

BL, no care for the EL

Compared the performance with:– Motion JPEG 2000 – FRExt-based lossless coding– Pixelwise residual entropy

BLEL

optimize

optimize


Rate Trade-OffRate Trade-Off

Better

Silent, IPPQP(P)8

12

16

20

24

28

4 5 6 7 8

Rat

e (%

of u

ncom

pres

sed)

OverallOpt BaseOpt

EL

BL

Total (EL+BL)

Pixel-wise residual entropy


Performance ComparisonPerformance Comparison

container news silent stefan

rate

(% o

f unc

ompr

esse

d)

Better

0

10

20

30

40

50

60

70Motion JPEG2000OverallOpt (EL/BL)BaseOpt (EL/BL)FRExt

Intra

IPP IBP IBBP IPP IBP IBBPIPP IBP IBBP

IPPIBP IBBP

Inter Inter Inter

InterIntra Intra

Intra


Coding Simulation with JSVM-BLCoding Simulation with JSVM-BL

Implemented on JSVM 4 (H.264 scalable extension reference codec)14 test sequences (QCIF/CIF/4CIF, 15/30/60Hz)BL structures: III…, IPP..., IBP... and IBBBP...Other settings– FastSearch: used– ME search range: 96– Reference frame #: 1– 8x8 transform: disabled

Compared with:– Ordinary coding of Motion JPEG2000 (lossless)– Residual coding of Motion JPEG2000 (lossless) for EL– Residual entropy


Bit Rate EvaluationBit Rate Evaluation

JSVMencoder

Originalvideo

JSVMdecoder

+

Entropymeasurement

Motion JPEG 2000Lossless Coder

Residualpacker

– pixel-wise residual video

Arithmeticcoder

“Proposed EL”

“Proposed BL”(III, IPP, etc.)

“EL Entropy” “MJ2K EL”

Motion JPEG 2000Lossless Coder

“MJ2K (total)”


Bit Rate Breakout (CITY 4CIF, BL=III/IPP)Bit Rate Breakout (CITY 4CIF, BL=III/IPP)

0

10

20

30

40

50

60

3 5 7 9 11 13QP

Compression ratio (%)

Total (III) Base (III) Enhance (III)

Total (IPP) Base (IPP) Enhance (IPP)MJ2K

EL

BL

EL+BL

Motion JPEG2000

rate

(% o

f unc

ompr

esse

d)


EL Bit Rate (BL=III…)EL Bit Rate (BL=III…)

0

5

10

15

20

25

30

3 5 7 9 11 13QPi

Bit-rate ratio (%)

MJ2K

Entropy

Proposed

rate

(% o

f unc

ompr

esse

d)


EL Bit Rate (BL=IPP…)EL Bit Rate (BL=IPP…)

0

5

10

15

20

25

30

3 5 7 9 11 13QPp

Bit-rate ratio (%)

MJ2K

Entropy

Proposed

rate

(% o

f unc

ompr

esse

d)


Performance ComparisonPerformance Comparison

Intra BL compression ratio is 3 points worse than Motion JPEG2000Inter BL is 7 points better than Motion JPEG2000

30

35

40

45

50

55

60

65

70

75

Bus(

CIF

@30

Hz)

(QC

IF@

15H

z)M

obile

(CIF

@30

Hz)

(QC

IF@

15H

z)C

ity(4

CIF@

60Hz

)

(CIF

@30

Hz)

Cre

w(4C

IF@

60H

z)

(CIF

@30

Hz)

Fore

man

(CIF

@30

Hz)

(QC

IF@

15H

z)H

arbo

ur(4

CIF

@60

Hz)

(CIF

@30

Hz)

Socc

er(4

CIF

@60

Hz)

(QC

IF@

30H

z)

Aver

age

MJ2K III IPP IBP IBBBPC

om

press

ion r

atio

(%)

rate

(% o

f unc

ompr

esse

d)


BL:EL Bit Rate RatioBL:EL Bit Rate Ratio

~1/3 bits are spent for EL, regardless of BL GOP structure and sequence

III base/enh bit rate ratio

0%10%20%30%40%50%60%70%80%90%

100%

BUS

CIF

BUS

QC

IF

MO

BILE

CIF

MO

BILE

CIF

CIT

Y 4C

IF

CIT

Y C

IF

CR

EW 4

CIF

CR

EW C

IF

FOR

EMAN

CIF

FOR

EMAN

QC

IF

HAR

BOU

R 4

CIF

HAR

BOU

R C

IF

SOC

CER

4C

IF

SOC

CER

CIF

EnhancementBase

IBBBBP base/enh bit rate ratio

0%10%20%30%40%50%60%70%80%90%

100%

BUS

CIF

BUS

QC

IF

MO

BILE

CIF

MO

BILE

CIF

CIT

Y 4C

IF

CIT

Y C

IF

CR

EW 4

CIF

CR

EW C

IF

FOR

EMAN

CIF

FOR

EMAN

QC

IF

HAR

BOU

R 4

CIF

HAR

BOU

R C

IF

SOC

CER

4C

IF

SOC

CER

CIF

EnhancementBase

Intra (III) Inter (IBBBP)

Average EL ratio=29% Average EL ratio=32%


Summary of Residual PackingSummary of Residual Packing

Scalable lossless coding having H.264-compliant BL– Drastic speed-up without losing efficiency– 20% less bits than conventional pixel-wise residual

coding– Intra BL coding (III)

– ~55% compression (compared to original)– ~3 points worse than Motion JPEG2000– 1-3% worse than Motion JPEG2000– 14% less bits than FRExt

– Inter BL coding (IPP, IBP, IBBBP)– ~44% compression (compared to original)– 5~7 points better than Motion JPEG2000– 14% more bits than FRExt

BL can be safely RD-Optimized (i.e., BaseOpt works good)Optimal EL amount: ~1/3 of total bit rate


OutlineOutline





Summary


VBS — A Gray-Scale Lossless Video CodecVBS — A Gray-Scale Lossless Video Codec

Video extension of MRP (see page 10)Variable block-sized motion compensation– Quad-tree based, 32x32 to 8x8 sizes– Integer-pel accuracy

Context-based 3D-linear prediction– M=24 contexts (predictors)– P=20 pixels from current frame– Q=25 pixels from previous frame– All M(P+Q) coefficients are optimized for minimum rate, not for

minimum MSE, and downloaded for each frameEncoding time: 3-4 minutes per frameDecoding time: 0.02 seconds per frame

[Shiodera et al., 2006]


SUT Codec Comparison (gray-scale)SUT Codec Comparison (gray-scale)

Average of 6 gray-scale (8-bit) video sequences (all CIF, 25 frames)

1.8

2

2.2

2.4

2.6

2.8

3

VBS FRExt VBS-intra JPEG-LS(intra)

JPEG 2000(intra)

Com

pres

sion

Rat

io

14%

34% gain due to Motion Compensation

15%

[Shiodera et al., 2006]


MSU Codec Comparison (YUV 4:2:0)MSU Codec Comparison (YUV 4:2:0)

1.8

2

2.2

2.4

2.6

2.8

YULS

MSU

Snow

FFV1

x264

LOCO

Laga

rithHuff

YuvCore

PNGAlpa

ry

LZO

Com

pres

sion

Rat

io

Average of 9 YUV 4:2:0 (YV12) video sequencesAlso provides encoding time, RGB&YUV4:2:2 coding comparisonWeighted by image size (#pixels, not by #frames)

[MSU “Lossless Codecs Comparison”, Apr. 2007]

9.7%

FRExt

Note: Conditions are different from previous comparison


OutlineOutline





Summary


SummarySummary

Presented lossless coding technologies– H.264 variants: FRExt extension and Scalable

coding– Introduced MRP, VBS and other state-of-the-art

lossless codec comparisons (only quantitative comparison, since the conditions are not the same)

Lossless image compression– About 1/2 for gray image– About 1/3 for RGB24 image

Lossless video compression– About 1/2.9 for both gray and YUV4:2:0 video


Last But Not Least…Last But Not Least…

Drawbacks– High data rate– No rate control possible– Difficult to enclose users (re-encoding with another

codec is not a big deal)Merits– No need to locally-decode the signal– Can predict blocks/frames in parallel– No subjective viewing necessary– Independent of monitors, viewers and viewing

conditions


ReferencesReferences

MRPI. Matsuda et al: “A Lossless Coding Scheme Using Adaptive Predictors and Arithmetic

Code Optimized for Each Image,” Trans. IEICE (DII), Vol.J88-D-II, No.9, pp.1798-1807, Sep. 2005 (in Japanese)

The art of Lossless Data Compressionhttp://artest1.tripod.com/graphics25.htmlReversible DCTK. Komatsu and K. Sezaki: “Reversible transform coding of images,” IEICE Trans.

Fundamentals, vol. J79-A, no. 4, pp. 981–990, Apr. 1996 (in Japanese)H.264 + reversible transform (and its follower)S. Takamura and Y.Yashima: “H.264-based Lossless Video Coding Using Adaptive

Transforms,” vol.2, pp.301-304, ICASSP 2005, Philadelphia, Mar. 2005S. Sano et al: “H.264-based lossless video coding using correlation of adjacent block”,

FIT2006, J-023, 2006 (in Japanese)H.264 + residual packingS. Takamura and Y.Yashima: “Lossless Scalable Video Coding with H.264 Compliant

Base Layer,” ICIP 2005, vol.2, pp.II-754-757, Genova, Sep. 2005VBST. Shiodera et al.: “A Lossless Video Coding Scheme Using MC and 3D Prediction

Optimized for Each Frame,” Journal of the ITE, vol.60, no.7, Jul. 2006 (in Japanese)

MSU codecs comparison 2007http://www.compression.ru/video/codec_comparison/lossless_codecs_2007_en.html

Stanford EE398B April 19, 2007 Lossless VideoCoding · Better Compression. Takamura: ... – Fixed,...

Documents

Transcript of Stanford EE398B April 19, 2007 Lossless VideoCoding · Better Compression. Takamura: ... – Fixed,...