Escolar Documentos
Profissional Documentos
Cultura Documentos
org/ijc
PerceptuallyOptimizedCodingofColor
ImagesforBandLimitedInformation
Networks
EvgenyGershikov
DepartmentofElectricalEngineering,OrtBraudeAcademicCollegeof Engineering,Karmiel,Israel
andDepartmentofElectricalEngineering,TechnionIIT,Haifa,Israel
eugeny@tx.technion.ac.il
Abstract
The Mean Square Error (MSE) or the Peak Signal to Noise
Ratio (PSNR) are common distortion measures used to
assess image quality. Nevertheless, they are usually chosen
duetotheirsimplicityandnottheirperformanceastheyare
notalwayssuitablecomparedtothehumanobserver.Inthis
work we present a RateDistortion approach tocolorimage
compressionbasedonsubbandtransformsusingperceptual
optimization of the compression quality. This approach is
based on minimization of the Weighted Mean Square Error
(WMSE)oftheencodedimage,whichbettercorrespondsto
thequalityassessmentofthehumaneye.TheWMSEcanbe
measuredintheYCbCrcolorspace,forwhichvisualweights
are relatively easily derived. Based on the new approach,
new optimized compression algorithms are introduced
using the Discrete Cosine Transform and the Discrete
Wavelet Transform. We compare the new algorithms to
presently available algorithms such as JPEG and JPEG2000.
Our conclusion is that the new WMSE optimization
approach outperforms presently available compression
systemswhenahumanobserverisconsidered.
Keywords
ColorImageCompression;WeightedMeanSquareError;Discrete
Cosine Transform; Discrete Wavelet Transform; Perceptual Rate
DistortionModel;OptimalColorComponentTransform;Optimal
RateAllocation
Introduction
Many color image coding algorithms are based on
subband transforms for the compression process. The
complexity of such algorithms varies from systems
based on elementary block transforms like the DCT
(DiscreteCosineTransform)[14]used,forexample,in
JPEG [21] to more complicated algorithms based on
the Lapped Biorthogonal Transform (LBT), the
Discrete Wavelet Transform (DWT), wavelet packets
and filter banks, such as EZW (Embedded Zerotree
Wavelet)[19], JPEG2000 [13][15], JPEG XR [2][16] or
ObjectiveRateDistortionTheoryofSubband
TransformCoders
Considerageneralsubbandtransformcoderforcolor
images. Typically, the image samples are first pre
processed, then subband transformed and quantized
and finally postprocessed losslessly. A detailed
descriptionofthesestagesisgivenbelow.
49
www.seipub.org/ijc
1) Preprocessing
HereaCCT(ColorComponentsTransform)isapplied
totheRGBcolorcomponentsoftheimage.Wedenote
the RGB components in vector form as
(4)
(1)
ThegoalofusingtheCCTtransformisusuallytode
correlate the highly correlated RGB components
[7][10][15][23].TheCCTtransformisoftenfollowedby
level shifting as for example is the case in JPEG2000
[13] so that the sample range of values becomes
symmetricaroundzero.
(5)
2) SubbandTransformingandQuantizing
Asubbandtransform,suchastheDCTortheDWTis
applied to each color component. The subband
coefficients of each color component are then
quantized. An independent uniform scalar quantizer
foreachsubbandisused.
3) Postprocessing
Thequantizedcoefficientsareencodedlosslessly.The
goal is to reduce the number of bits required for the
coefficients without loss of information. Techniques
such as runlength encoding, zero trees, delta
modulation and entropy coding are used here. This
stagehastobeadaptedtothesubbandtransformused.
j 1 j j
i2Gb bi2 ( MM T )1 ii
1
i
ln
kk
a
3
3 ( MM T )1 k2GM kAct j 1 j j
kk
k 1
(2)
R
3
(6)
(3)
Here istheMSEofsubband
,
is its variance, is its energy gain [20] and is the
rate allocated to it. Also is its sample rate, i.e., the
relativepartofthenumberofcoefficientsinitfromthe
total number of samples in the signal. is a constant
equalto
.
50
InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc
(11)
setofnonzero(oractive)ratesinthecolorcomponent
,i.e.,
(7)
Also
andusing(9)wecanexpress
(14)
where is the row of
in column form. In a
similar fashion, the diagonal elements of
can be
recognized as the MSE distortions of the , ,
color components, given by (2) and slightly rewritten
tobecome:
where
(15)
denote the
(16)
once(15)issubstitutedfor .Nowif,forthesakeof
convenience, we denote the YCbCr color components
at each pixel as a vector
, then the
WMSEofthe colorcomponent
is:
(17)
Ascanbeseen,thisexpressionincorporatestheenergy
gainsofthesubbands aswellastheirsamplerates
. Also the visual weights
are part of the
expression,providingvaryingsignificancetodifferent
subbands of the same color component as well as
between color components. Defining the total WMSE
astheaverageWMSEoftheYCbCrcolorcomponents,
weget:
WMSE
(10)
SincetheSBTcoefficientsarelossyencoded,errorsare
introducedbetweenthereconstructedcoefficients
(9)
rateandvarianceofsubband ofcolorcomponent ,
respectively. Note that we continue here with the
consistent notation of a tilde for the variables related
to the C1C2C3 color space. Assuming that the
quantization errors of the three color components in
eachsubbandintheC1C2C3domainareuncorrelated,
becomesadiagonalmatrixand(14)becomes
as:
(13)
(8)
(12)
by
1 3
WMSE ( xYCbCr )i
3 i 1
1 3 B 1
bGb wbi d bi
3 i 1 b0
(18)
the expression
51
www.seipub.org/ijc
is the same
coefficient of at index and
coefficientafterquantizationandreconstruction.Now
3 _ 1
2
1 3 B1
WMSE bGbwbi M k2 bk eaRbk
3 i1 b0
k 1
ik
2
_ 1
1
bGb eaRbk wbi M .
3 b0
k 1
i 1
ik
B1
2
2
k bk
(19)
for
Tosimplify(19)wedenote:
(24)
(20)
sothattheWMSEexpressionbecomes:
where
. Assuming zero
correlation of the quantization errors of the different
color components in each subband and between
(21)
Clearly,ifthevisualweights areallequalto1,the
WMSE expression of (21) should become the
expression for the MSE in the YCbCr domain. This
expression is given exactly by (4) with the difference
that there is to be replaced by in our case. From
thecomparisonofequations(21)and(4)weconclude
that
in that case, which means
accordingto(20)that
BasicOptimizationUsingtheWMSEModel
(22)
AfterderivingtheWMSEexpression,thenaturalnext
step is to use it to find the optimal rates and optimal
CCT in the WMSE sense. First we wish to minimize
the WMSE of (21), subject to the rate constraint
, resulting in the following
Lagrangian( istheLagrangemultiplier):
Astraightforwardcheckprovesthatthisisindeedthe
case.
DecorrelationoftheQuantizationErrors
In the derivation of the WMSE expression of (21) we
haveassumedthatthequantizationerrorsofthe , ,
colorcomponentsareuncorrelatedineachsubband.
It is of interest to note that the assumption in the
derivationoftheMSEexpressionof(4)wasthelackof
correlation of the quantization errors in the image
domain [2], i.e. that
and
have zero
correlation for
,
. Note that
(25)
whichisminimizedbytheoptimalratesgivenby:
(26)
Here
(27)
Notethatnoconstraintsfornonnegativityoftherates
are used here, which means that high rates are
assumed. As for the optimal CCT matrix : it can be
foundbyminimizingthetargetfunction
,thatis
actuallythedenominatorofthe
in(26)aftersome
straightforwardmanipulations:
(23)
52
follows
, hence
(28)
InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc
and i areasin(8).Asfor
YRGB E YbRGB
b
(30)
accordingto:
(31)
Here
is defined similarly to the
definitions in the beginning of the section of . Also
denotesthe rowofthe matrixinvectorform.
Thusthetargetfunction
canberewrittenas:
3
B 1
f ( M ) ( mk T b mk )Gb bk ,
_
k 1 b 0
bk
(32)
_ 1
wbi M .
i 1
ik
3
DCTBasedCompressionAlgorithm
Since the DCT is a subband transform, the Rate
Distortion theory of Section The Perceptual RD
Model can be applied to it. To find the DCT visual
weights we use the HVS CSF (Contrast Sensitivity
Function)curvesfortheYCbCrcolorspacethatcanbe
found, for example, in [20]. To convert the cpd (cycle
per degree) units of thesegraphs to spatialfrequency
unitsfortheDCT,theequationsproposedin[22]can
be
used.
We
consider,
for
example,
512 512
images
displayed
as
onadisplaywithdotpitchof0.25mm.
The viewingdistance isassumed to befour times the
imageheight[12],i.e.,inthisexample50cm.Similarly
we can consider
images displayed as
onabigscreenataviewingdistance
of100cm.
OptimalRateswithDownSampling
When considering potential downsampling of some
ofthecolorcomponents,therateconstraintbecomes(5)
and the Lagrangian that incorporates this constraint,
as well as constraints for the nonnegativity of the
subbandrates,is:
2
1 B 1 3
L {Rbi }, M , ,{bi } bGb bk k2e aRbk bk
3 b0 k 1
(33)
3 B 1
3 B 1
i b Rbi R bi Rbi ,
i 1 b0
i 1 b0
Rbi
Thestagesoftheproposedalgorithmareasfollows:
j j
1.
j 1
i2Gb bi2 bk
1
i
ln
k k
a
3
Act 2
Act
3 GM k k k j 1 j j
k 1
(35)
andcanalsobeexpressedusingthe matrixandthe
subband b covariance matrix in the RGB image
domain:
T
,
b E YbRGB YRGB
YbRGB YRGB
b
b
,itisgivenby:
(29)
FindtheoptimalCCT byminimizing(32).
(34)
53
www.seipub.org/ijc
5. QuantizetheDCTcoefficientsusingauniform
scalar quantizer in each subband. The (optimal)
quantization steps are found using an iterative
algorithm[5].
where
for each color component is calculated
in the DWT domain in the YCbCr color space
accordingtothevisualweightssuggestedin[20].Then
theaveragePSPNRonthe3colorcomponentsistaken.
Based on our experience and results, this is a good
measureofsubjectiveimagequality.
(36)
TABLE1Perceptuallybasedresults(PSPNR)for(fromlefttoright):TheDCTbasedWMSEoptimizedalgorithmintheYCbCrdomain;Thesame
algorithmwithoptimalCCT;TheMSEoptimizedalgorithm;JPEG.Thecompressionrateforeachimageisshownintherightcolumn.
54
WMSEAlg.inthe
optimaldomain
40.6
MSEAlg.
JPEG
Rate[bpp]
Lena
WMSEAlg.inthe
YCbCrdomain
39.4
38.9
37.6
0.76
Peppers
39.6
39.6
38.1
36.6
0.81
Baboon
42.0
42.5
39.2
36.1
1.76
Cat
41.3
43.1
41.3
39.9
1.30
Landscape
42.5
42.5
40.5
38.0
1.85
House
39.8
40.3
39.2
38.1
0.54
JellyBeans
38.5
38.6
38.3
37.5
0.47
Fruits
41.0
42.3
40.4
38.9
0.71
Sails
41.0
42.9
39.7
37.6
1.84
Image
InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc
Monarch
39.8
40.2
38.7
37.5
1.03
Goldhill
42.9
43.4
41.9
40.6
2.17
Mean
40.7
41.5
39.7
38.0
TABLE2SameasTable1,butforPSNRinsteadofPSNR.NotethatoptimizationofPSPNR,asinducedbythehumanobserver,doesnot
necessarilymeanoptimizationofthearbitrarilyusedPSNR(seetext).
PSNR
Image
WMSEAlg.inthe
optimaldomain
30.5
MSEAlg.
JPEG
Rate[bpp]
Lena
WMSEAlg.inthe
YCbCrdomain
30.0
30.7
29.7
0.76
Peppers
30.0
30.1
30.5
29.3
0.81
Baboon
30.0
29.0
30.5
26.5
1.76
Cat
30.0
29.6
31.3
29.5
1.30
Landscape
30.0
30.1
30.3
25.9
1.85
House
30.0
30.2
30.3
29.5
0.54
JellyBeans
30.0
30.3
30.6
29.7
0.47
Fruits
30.0
29.8
30.6
30.6
0.71
Sails
30.0
29.7
30.6
28.9
1.84
Monarch
30.0
29.6
30.6
29.4
1.03
Goldhill
30.0
30.2
31.7
29.2
2.17
Mean
30.0
29.9
30.7
28.9
55
www.seipub.org/ijc
FIG.1COMPRESSIONRESULTSFORLENAAT0.72BPP.ORIGINALIMAGE(TOPLEFT);IMAGECOMPRESSEDBYTHEWMSE
OPTIMIZEDALGORITHM(TOPRIGHT,PSPNR=40.4DB);IMAGECOMPRESSEDBYJPEG(BOTTOMLEFT,PSPNR=37.7DB);IMAGE
COMPRESSEDBYTHEMSEOPTIMIZEDALGORITHM(BOTTOMRIGHT,PSPNR=39.3DB).ASEXPECTED,THEWMSEALGORITHM
OUTPERFORMSTHEOTHERMETHODS,ESPECIALLYINTHEMARKEDAREAS.
FIG.2COMPRESSIONRESULTSFORTHEBABOON(ZOOMEDIN)AT0.88BPP.ORIGINALIMAGE(TOPLEFT);IMAGECOMPRESSEDBY
THEWMSEOPTIMIZEDALGORITHM(TOPRIGHT,PSPNR=36.9DB);IMAGECOMPRESSEDBYJPEG(BOTTOMLEFT,PSPNR=33.6DB);
IMAGECOMPRESSEDBYTHEMSEOPTIMIZEDALGORITHM(BOTTOMRIGHT,PSPNR=35.6DB).HEREAGAIN,THEWMSE
ALGORITHMOUTPERFORMSTHEOTHERMETHODS.
56
InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc
DWTBasedCompressionAlgorithm
When the DWT is considered, there are quite a few
options for the wavelet filter bank to be used for the
decomposition. We have chosen the Daubechies 9/7
filter bank, but obviously other choices can be
consideredaswell.Notiling[13]isused.Thechoiceof
the visual weights is according to [20]. The stages of
theproposedalgorithmare:
1.
FindtheoptimalCCT byminimizing(32).
Itisinterestingtocomparetheproposedalgorithmto
JPEG2000. We have considered the JPEG2000
implementationusingtheJasPersoftwarepackage[24]
andanotherversionoftheimplementationwithfixed
visual weighting at subband level using the CSF
weights of[20]. The visual results for the Lenaimage
canbeseeninFig.3.ThePSNRresultshereare29.5dB
for the proposed WMSE optimized algorithm, 28.6dB
for JPEG2000 (original JasPer implementation) and
28.5dB for JPEG2000 with CSF weights. We conclude
that the use of CSF weights, that affects the tier2
codingstageoftheJPEG2000algorithm,decreasesthe
PSNR,butslightlyimprovesthevisualperformance.
3. ApplytheDWTtreedecompositionuptothe
required depth of the tree (3, 4, 5 or higher
according toimagesize) to each colorcomponent
,
.
4. Calculate the optimal rates according to (34)
substituting there the used CCT matrix and the
variances, sample rates and energy gains of the
DWT subbands. The determination of the active
subbands is the same as for the DCTbased
algorithmoftheprevioussubsection.
FIG.3COMPRESSIONRESULTSFORLENAAT0.52BPP.ORIGINALIMAGE(TOPLEFT);IMAGECOMPRESSEDBYTHEDWTBASED
WMSEOPTIMIZEDALGORITHM(TOPRIGHT,PSPNR=19.7DB);IMAGECOMPRESSEDBYJPEG2000(BOTTOMLEFT,PSPNR=19.1DB);
IMAGECOMPRESSEDBYJPEG2000WITHCSFWEIGHTS(BOTTOMRIGHT,PSPNR=19.2DB).
ALSOHERE,THEWMSEALGORITHMISSUPERIORTOTHEREST.
57
www.seipub.org/ijc
andmustachetexturesintheCatimagearelost.Inthe
case of the Peppers image, the compression result of
JPEG2000 is less pleasing to the eye due to the color
artifacts introduced. Fig. 6 further demonstrates the
loss of spatial details in the case of JPEG2000
compression of the Sails image, the blurring of the
contoursintheMonarchimageandbotheffectsinthe
Goldhill image (see the top marked area for the
blurred contour effect and, for example, the bottom
left marked area for the loss of spatial details).
Furthermore, color artifacts are introduced by
JPEG2000 in the Goldhill image as indicated, for
instance,inthemarkedareainthecenteroftheimage.
FIG.4LANDSCAPE,HOUSEANDJELLYBEANSIMAGESFROMLEFTTORIGHT:ORIGINAL,COMPRESSEDBYTHEWMSE
ALGORITHM(WMSEALG.)ANDCOMPRESSEDBYJPEG2000.
PSPNRFORTHELANDSCAPEIMAGE:17.1DB(WMSEALG.)AND15.7DB(JPEG2000).
PSNR:28.7DB(WMSEALG.)AND25.3DB(JPEG2000)AT0.97BPP.
PSPNRFORTHEHOUSEIMAGE:19.4DB(WMSEALG.)AND19.0DB(JPEG2000).
PSNR:31.2DB(WMSEALG.)AND33.1DB(JPEG2000)AT0.68BPP.
PSPNRFORTHEJELLYBEANSIMAGE:18.8DB(WMSEALG.)AND18.2DB(JPEG2000).
PSNR:32.3DB(WMSEALG.)AND32.1DB(JPEG2000)AT0.48BPP.
INTHEONLYCASEWHERETHEPSNROFJPEG2000ISHIGHERTHANTHENEWALGORITHM(HOUSE),THEPSPNRRESULT
SUPPORTSTHEFACTTHATVISUALLYTHENEWALGORITHMPROVIDESSUPERIORRESULTS.
58
InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc
FIG.5FRUITS,CATANDPEPPERSIMAGESFROMLEFTTORIGHT:ORIGINAL,COMPRESSEDBYTHEWMSEALGORITHM(WMSE
ALG.)ANDCOMPRESSEDBYJPEG2000.
PSPNRFORTHEFRUITSIMAGE:22.2DB(WMSEALG.)AND21.1DB(JPEG2000).
PSNR:30.0DB(WMSEALG.)AND29.0DB(JPEG2000)AT1.34BPP.
PSPNRFORTHECATIMAGE:17.0DB(WMSEALG.)AND16.2DB(JPEG2000).
PSNR:28.9DB(WMSEALG.)AND26.9DB(JPEG2000)AT0.63BPP.
PSPNRFORTHEPEPPERSIMAGE:20.3DB(WMSEALG.)AND19.3DB(JPEG2000).
PSNR:30.8DB(WMSEALG.)AND30.7DB(JPEG2000)AT0.86BPP.
ASCANBESEEN,PSNRANDPSPNRRESULTSARESUPERIORFORTHENEWALGORITHMCOMPAREDTOJPEG2000.ITISALSO
OBSERVEDVISUALLYEXAMPLESAREINDICATEDINTHEMARKEDAREAS.
59
www.seipub.org/ijc
FIG.6SAILS(ZOOMEDIN),MONARCH(ZOOMEDIN)ANDGOLDHILLIMAGESFROMLEFTTORIGHT:ORIGINAL,COMPRESSED
BYTHEWMSEALGORITHM(WMSEALG.)ANDCOMPRESSEDBYJPEG2000.
PSPNRFORTHESAILSIMAGE:19.2DB(WMSEALG.)AND18.0DB(JPEG2000).
PSNR:28.9DB(WMSEALG.)AND26.6DB(JPEG2000)AT0.70BPP.
PSPNRFORTHEMONARCHIMAGE:19.9DB(WMSEALG.)AND19.6DB(JPEG2000).
PSNR:29.0DB(WMSEALG.)AND28.8DB(JPEG2000)AT0.56BPP.
PSPNRFORTHEGOLDHILLIMAGE:17.6DB(WMSEALG.)AND16.6DB(JPEG2000).
PSNR:27.0DB(WMSEALG.)AND24.5DB(JPEG2000)AT0.59BPP.
ONCEAGAIN,THEPSNRANDPSPNRRESULTSARESUPERIORFORTHENEWALGORITHMCOMPAREDTOJPEG2000(SEE
EXAMPLESINDICATEDINTHEMARKEDAREAS).
stageusingaCCT,followedbyasubbandtransform,
quantizationstage,andlosslessentropyencoding.The
algorithms are optimized with regard to the color
component transform in the preprocessing stage of
the compression as well as the quantization tables
usedinthecodingstage,bothwithrespecttoWMSE.
TheproposedDCTbasedalgorithmoutperformsboth
JPEG and the corresponding MSE optimized
algorithm. The DWTbased algorithm, as expected,
achieveshighercompressionratiosforthesameimage
quality than DCTbased techniques. We demonstrate
in this work that even when a relatively basic
algorithm is used in the postprocessing stage
(introducedforEZW),superiorresultsareobtainedby
Summary
A perceptuallybased model for the RateDistortion
functionofcolorsubbandcodershasbeenintroduced.
ThenewmodelapproximatestheWMSEdistortionof
an image in a given color space, such as YCbCr. This
distortion is then minimized to achieve perceptual
optimizationofthecompression.Whentheweightsin
the WMSE calculation are taken based on the CSF
curves of the human visual system, better
correspondence to image quality assessment by the
humaneyeisachieved.
Based on the RateDistortion model, new algorithms
have been introduced consisting of a preprocessing
60
InternationalJournalofCommunications(IJC)Volume2Issue2,June2013www.seipub.org/ijc
ACKNOWLEDGMENT
REFERENCES
fidelitycriterionoftheencodingofimages.IEEETrans.
onInformationTheory20(Jul.1974):525536.
[1] Faugeras,O.D.DigitalColorImageProcessingWithin
theFrameworkofaHumanVisualModel.IEEETrans.
(1979):380393.
Processing37(Nov.1989):17431750.
and
JPEG
2000.",
Journal
of
Visual
Processing:ImageCommunication17(2002):348.
2012):835844.
1990.
ApproachtoOptimalColorImageCompression.Proc.
ofEUSIPCO,Florence,Italy,2006.
mode.Proc.ofICIP(Oct.2012):25012504.
[16] Richter,T.SpatialConstantQuantizationinJPEGXRis
Conference(March2010):7988.
BerlinHeidelberg,2007.
VisionComputing25(2007):637651.
bitallocationforoptimalsubbandimagecompression.
SignalProcessing:ImageCommunication22(2007):118.
BanksOptimizedforSubbandCodingofSpeech,IEEE
SignalCompression,Boston,MA:Kluwer,1992,ch.2.
(Nov.1989):16721679.
IEEEICIP(2002):II353II356.
SignalProcessing41(1993):33453462.
61
www.seipub.org/ijc
[20] Taubman,D.S.andMarcellin,M.W.JPEG2000:image
inR,G,BComponents.Trans.onCommunications32
KluwerAcademicPublishers,2002.
(Nov.1984):12011209.
[24] http://www.ece.uvic.ca/mdadams/jasper
xviiixxxiv.
[22] Wang, C. Y., Lee, S. M. and Chang, L. W. Designing
JPEG quantization tables based on human visual
system. Signal Processing: Image Communication 16
(2001):501506.
Recognition.
62