ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

FLC (NTNU) VLC

Error resilience capabilities Error resilience capabilities ((cont’dcont’d))

R=0.5 bit/pixel, ber=0.001

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

• Re-synch marker at packet boundaries• Ability to locate errors in a packet

Magic String(3 bytes)

“JP2”

Header Length(2 bytes: msb first)

HL

Global Header

( HL bytes)

Packet

Head BodyRes

ync

(2 b

ytes

) Packet

Head BodyRes

ync

(2 b

ytes

)

optional

Error resilience capabilities Error resilience capabilities ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

• Re-synch marker atblock boundaries

• Locate errors in ablock

Segment Marker

Segment Marker

Segment Marker

Segment Marker

Bitplane

Bitplane

Bitplane

Bitplane

error

Code block

Error resilience capabilities Error resilience capabilities ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Reversible color transformation:Reversible color transformation:making making lossless colourlossless colour coding possible coding possible

GBVr

GRUr

BGRYr

−=−=

++

=4

*2

GVrB

GUrR

VrUrYrG

+=+=

+−= )

4(

All components must have identical subsamplingparameters and same depth

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Visual Frequency WeightingVisual Frequency Weighting• Allows system designers to take advantage of

visual perception

• Fixed Visual Weighting (FVW) &Progressive Visual Coding (PVC)

• FVW: CSF are chosen according to the finalviewing condition

• Implementation: Q steps in subband I aremodified based on the CSF

• PVC: Visual weights are changes during theembedded process

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Visual frequency weighting: FVWVisual frequency weighting: FVW• modify transform coefficients by multiplying

by the CSF weight (decoder has to know)

• modify Q step sizes (decoder needs not know)

• modify the embedded coding order– the distortion weights fed to the R-D optimization are

altered

– this controls the relative significance of includingdifferent number of bitplanes from the embeddedbitstream of each code-block

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Visual frequency weighting: PVCVisual frequency weighting: PVC

• Visual weights have to be changed duringthe embedded process– difficult to change the coefficient values of Q steps– performance of entropy coder might degrade due to

changing statistics of the binary representation

• Solution– change on the fly the order in which code blocks sub

biplanes should appear in the embedded bitstreambased on the visual weights

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Weighting set 0: r(0), with w(0) = { w0

(0) , w1(0) , . . . , wn

(0)} ;Weighting set 1: r(1)

, with w(1) = { w0(1) , w1

(1) , . . . , wn(1)} ;

.

.

.Weighting set m: r(m), with w(m) = { w0

(m) , w1(m) , . . . , wn

(m)}

•Distortion metric is changed progressively based on the visual weights during bitstream formation

•Bitsream formation is driven by postprocessing R-D optimization

•Progressive visual weights control the embeddingorder of code-block sub-bitplanes on the fly

Visual frequency weighting: PVCVisual frequency weighting: PVC

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

ExampleExample0.25 bpp, TCQ(RMSE:8.81), Visual TCQ (13.53)

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Line based transformsLine based transforms

• A way for low memory implementation of thewavelet transform

• Same wavelet coefficients as full framewavelet transform

• Same encoding results as the standard VM

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Line based transforms Line based transforms ((cont’dcont’d))

VH

VL

HH

LH

HL

LL

Input image data

High pass Horizontal

High pass Horizontal

Low pass Horizontal

Low pass Horizontal

Line buffering for vertical decomposition

Low pass Vertical

High pass Vertical

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Line Buffer

Line Buffer

Line Buffer

0

1

2

3

4

Input Image

Encode

Encode

Encode

Encode

Encode

Line Buffer

Filtering Elements

Line Buffer

Line based transforms Line based transforms ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Compressed image manipulationCompressed image manipulation

• Allows for– rotations of 90, 180, 270 degrees– vertical flipping (horizontal axis symmetry)– horizontal flipping (vertical axis symmetry)– all possible combinations of abovein the wavelet domain

by rearranging the quantized subbandcoefficients (no modification of the coeffs.)

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Compressed imageCompressed imagemanipulation: advantagesmanipulation: advantages

• More efficient in terms of– memory– complexity requirements

• No additional distortion is introduceddue to inverse / forward transformation

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Bit stream

entropydecoding

geometricmanipulation

entropydecoding

Transcoder

Bit stream

JPEG2000 Encoder

JPEG2000 Encoder

Update of OCB h, OCBv and T bits

originalimage

Reconstructed image Bit stream

entropydecoding

geometricmanipulation

entropydecoding

Transcoder

Bit stream

JPEG2000 Encoder

JPEG2000 Encoder

Update of OCB h, OCBv and T bits

originalimage

Reconstructed image

OCBh=0, OCBv=0, T=0

Compressed image manipulation Compressed image manipulation ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Transcoding: vertical flipping on subbands

Compressed image manipulation Compressed image manipulation ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Transcoding: 90 degrees rotation on subbands

Compressed image manipulation Compressed image manipulation ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

• The only information needed at thedecoder are the filtering conventionorders (OCBh/OCHv) for horizontal andvertical direction

• The decoder must know if the linesand/or columns have been flipped

Compressed image manipulation Compressed image manipulation ((cont’dcont’d))

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

PostprocessingPostprocessing

• Typical artifacts in Wavelet coding areringing effects

• Postprocesing can reduce these artifacts

• Postprocesing filter based on robust M-estimator

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

PostprocessingPostprocessing ((cont’dcont’d.).)

Examples of filtering windowsExamples of filtering windows

∑=

−=N

ijix

xxxj 1

)(minargˆ ρ

))(1log()(

}1,min{)(

21,||)(

||)|(|2

||)(

221

22

2

2

γ

γγ

ργρ

γρ

γγγγγ

ρ

xxLorenzian

xxLTruncated

xxL

xx

xxxHuber

+=

=

≤≤=

>−+≤

=

P(x) characterizes the behavior of the estimator,or the smoothing capability

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

ConclusionsConclusions

• Advanced still image coding standard

• Better than current baseline JPEG

• Includes many interesting functionalities

• Intended to become the key standard for

still image coding in the next millennium

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

More informationMore information• JPEG

– http://www.jpeg.org

• EUROSTILL– http://ltswww.epfl.ch/~eurostill

• SPEAR– http://spear.jpeg.org/

• JJ2000– JavaTM JPEG2000 development– http://jj2000.epfl.ch

ICIP’99, Oct. 24-28, Kobe, JapanCharilaos Christopoulos

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

AcknowledgementsAcknowledgements• Mr. Joel Askelöf, Ericsson• Dr. Eiji Atsumi, Mitsubishi, Japan• Martin Boliek, Ricoh• Dr. Christos Chrysafis, HP Labs• Prof. Touradj Ebrahimi, EPFL• Prof. Nariman Farvardin, Univ. Maryland• Prof. Faouzi Kossentini, UBC• Mathias Larsson, Ericson• Dr. Daniel Lee, HP Labs• Dr. Eric Majani, CRF• Prof. Michael Marcellin, Univ. of Arizona• Prof. Andrew Perkis, NTNU• Dr. Majid Rabbani, Kodak• Dr. David Taubman, HP Labs & Univ. New South Wales

**

* In alphabetical order* In alphabetical order

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

Thank you for your attention!Thank you for your attention!

ERICSSON RESEARCH

Media Lab

UNIVERSITY OF PATRAS

Electronics Laboratory

C.A.C.A.ChristopoulosChristopouloscharilaos.christopoulos@era.ericsson.se

A.N.A.N.SkodrasSkodrasskodras@cti.gr