Efficient encoding of significance maps in wavelet based image compression
- 格式:pdf
- 大小:43.57 KB
- 文档页数:4
EFFICIENT ENCODING OF THE SIGNIFICANCE MAPS IN W A VELET BASED IMAGECOMPRESSIONLevent¨Oktem,Rus¸en¨Oktem,Karen Eguiazarian,Jaakko AstolaSignal Processing Laboratory,Tampere University of Technology,P.O.B.553,FIN-33101Tampere,Finlandloktem,rusen,karen,jta@cs.tut.fiABSTRACTWe propose a method for the efficient encoding of the signif-icance maps in wavelet-based image compression.Adopting the significance map encoding part of the morphology-based scheme in[1],we introduce three new features:exploita-tion of the knowledge of coefficient magnitudes in the parent band;reordering of the residual part of the significance map; and the employment of hierarchical enumerative coding,a new entropy coding method,instead of arithmetic coding. Experimental results show that utilization of these features bring a consistent improvement.The average improvement is around7.6%.1.INTRODUCTIONWavelet transform decomposes an image into multiresolu-tion bands.At each stage of decomposition,approximation band is decomposed into four subbands:approximation band at the next scale,horizontal details band,vertical details band, and diagonal details band.The original image can be consid-ered as the approximation band at thefirst scale.The re-sulting coefficients are classified as a set of approximation coefficients at thefinal scale,and hierarchical sets of hori-zontal,vertical,and diagonal detail coefficients.For typical images,most of the energy is compacted into the approxi-mation band,which is the smallest band in size.Most of the detail coefficients have small magnitudes and they be-come zero after the quantization step of a lossy compression algorithm.A considerable amount of compression can be achieved by separate encoding of the magnitudes and the lo-cations of nonzero detail coefficients,if the locations can be encoded efficiently.In[2]and later in[1],it was shown that locations of nonzero detail coefficients can be predicted from the locations of nonzero coefficients at a coarser scale.The locations of nonzero detail coefficients can be repre-sented as a binary image,named a significance map.More formally,let a coefficient be named as significant with re-spect to a threshold if its absolute magnitude is greater than that threshold.A significance map of a two dimen-sional data is a binary image with the same size as. The th pixel of is if is signifi-cant with respect to,and is,otherwise.A significance map can be generated for each detail band of wavelet trans-form coefficients with respect to some threshold.Letbe a band of wavelet coefficients where is the scale anddenote horizontal,detail,vertical bands,re-spectively.Shapiro has formed a parent-child dependency between the elements of and in[2]such that is the parent of four children.Then,he observed that if a parent is significant with respect to a threshold,children descending from that parent are more likely to be significant.Servetto et al.observed that not only the children descending from a significant parent,but also the neighbor of those children are more likely to be significant[1].Hence,they proposed to predict the significance map of a detail band from the signif-icance map of the detail band at a coarser scale,by using the M orphological R epresentation of W avelet D ata.MRWD al-gorithm has been generalized to encode the locations of sig-nificant wavelet packet transform coefficients in[3].In[4], MRWD algorithm has been improved by employing signif-icance links across clusters(or connected components)in-stead of direct exploitment of pixelwise parent-children re-lationship.In this paper,we take another direction for im-proving the significance map encoding part of the MRWD algorithm.Our proposal introduces three new features: Exploitation of the knowledge of coefficient magni-tudes in the parent band;Reordering of the residual part of the significance mapaccording to the distance to the previously visited points.Employment of hierarchical enumerative coding as theentropy coder.In section2,we review a simplified version of MRWD, and introduce our algorithm.In section3,we present exper-imental setup and the results.Conclusions follow as the last section.2.ALGORITHM DESCRIPTION2.1.Preliminary Decomposition of the Significance Map Our preliminary decomposition of the significance map is adopted and slightly simplified from the MRWD method in [1].Before going into the description of the decomposition algorithm,it might be useful to specify the recursive subrou-tine of region growing.This subroutine is employed at two different places in the algorithm.Region growing around a point(i,j)has the following form:grow()For all in the neighborhood of which havenot been visited,–if is not significant,output a,–if is significant,output a,grow().where the neighborhood of is defined as the set of pixels within the rectangular box with upper left corner coordinates and lower right corner coordinates.The detail coefficients of the highest scale are the roots of the parent-children tree,hence they do not have any par-ent.Significance maps of the detail coefficients in the highest scale are encoded as follows:Start scanning the coefficients in the raster scanningorder.For each significant coefficient encountered andhas not been visited,–Encode the location of,–grow()Hence,the scanning order is arranged so that the neigh-borhood of a significant coefficient is scannedfirst,and this neighborhood is grown until no further significant coefficient is found in the neighborhood.After region growing stops, raster scanning continues in the region which has not been scanned before,until the next significant coefficient is found. When the scanning of the band is completed,the bits output by grow are collected together and entropy coded.The points not visited by grow are known to be zero.The rest of the detail bands are scanned in two passes. In thefirst pass,a prediction of significance map is obtained from the significance map of the coarser scale .In,all children descending from a in,and a predefined neighborhood of thosechildren are marked as.The region marked as in the prediction map is scanned in the raster scanning order.The region growing algorithm explained above is used for each significant coefficient.A binary sequence is gen-erated for those pred icted locations indicating whether the prediction is true or false.In the second pass,the regions which are not visited in thefirst pass are scanned and a bi-nary sequence is generated for those res idual locations in-dicating whether there is any significant coefficient or not. The simplification of the decomposition scheme of MRWD takes place in this second pass:in MRWD,region growing is applied for scanning the remaining coefficients,whereas we apply just raster scan.In our experiments,we could not ob-serve a consistent improvement brought by region growing in the second pass that would justify its use.2.2.Statistical Behaviour of pred and resThe main purpose of decomposing the significance map into two1-D sequences,pred and res,is compacting as many of the significant points as possible into pred,while keeping the size of pred as small as possible.Ideally,pred would consist of all ones,and res would consist of all zeros.Typically, pred contains much more nonzero points than res,has higher entropy,and its length is much smaller.The statistics of both sequences can vary significantly from image to image.Even for the same image,the statistics can be very different for two different bit rates.This dictates the use of an adaptive entropy coder.Another aspect of the statistical behaviour of these se-quences is nonstationarity.While both sequences can be con-sidered to be more-or-less locally stationary,there is often a considerable overall nonstationarity.So,the employed en-tropy coding method should be fast in adapting to thefluctu-ating statistics.Considering these statistical properties of pred and res, hierarchical enumerative coding(HEC)[5]can be expected to compress them efficiently,because HEC was shown to be more efficient than context-based arithmetic coding for com-pressing locally stationary binary signals which have a sig-nificant overall nonstationarity.2.3.Reordering Within pred and resOnce the sequences pred and res are modelled as locally stationary signals,it becomes meaningful to apply some re-ordering to each sequence for increasing the local stationar-ity.For each of the sequences,we aim that those points hav-ing higher probability of being significant will be scanned earlier in the sequence.Different reordering strategies are used for pred and for res to achieve this aim.2.3.1.Exploiting the Parent Magnitude Information Since the magnitudes are encoded separately for each band, the decoder can know the parent magnitudes at the time of decoding the significance map of the child band.This infor-mation can be exploited as follows.Following the assumption that the children of a large-magnitude parent are more likely to be significant than the children of a small-magnitude parent,the children of signifi-cant parents can be scanned according to the descending or-der of parent magnitudes.This change in scanning order is equivalent to a reordering of the pred sequence of the child band.2.3.2.Reordering of the Residual SequenceFor res,we propose a reordering according to the distance from closest points in pred sequence.The residual points are classified into two sets:those who are closer to pred points than a threshold distance(e.g.3units),and the rest.This classification can be implemented easily by a binary morpho-logical dilation operation on the binary map of already vis-ited points.The reordered res sequence is obtained by raster scanning the points in thefirst setfirst,and then raster scan-ning the points in the second set.This reordering is based on the assumption that those residual points who are closer to pred points are more likely to be significant.putational Complexity of ReorderingThe reordering of the pred sequence does not really require an explicit sorting,and can be implemented by a two-pass scanning of the parent band.Thus,the total computational cost of reordering pred and res sequences is about two extra raster scans and a binary dilation.This causes a rather little increase in the total computational complexity of the overall algorithm.2.4.Entropy CodingIn MRWD,pred and res sequences of each band are encoded independently,using context-based arithmetic coding[6].We employ hierarchical enumerative coding[5]instead of arith-metic coding for coding the pred and res sequences.We do not propose a method for coding the coefficient magnitudes, but we assume that the coefficient magnitudes of the parent band have been decoded at the time of decoding the signifi-cance map of the child band.3.EXPERIMENTAL SETUP AND RESULTS Experiments were performed on seven standard test images, using three different picture qualities for each image.Picture Table1:Number of bits spent by each method for encoding the pred sequences.AC column stands for arithmetic coding with no reordering.AC+R columns are for reordering plus arithmetic coding;HEC for hierarchical enumerative coding with no reordering,and HEC+R is for hierarchical enumera-tive coding plus reordering.Image AC AC+R HEC HEC+R QS=402083207819561913 Lenna QS=204267417839633846 QS=107897746471837033QS=404630456641064053 Barbara QS=208790831675347416 QS=1014115130751217111990QS=405460537155065408 Bridge QS=2013960135021394613765 QS=1023028221822272822529QS=403603357234883427 Lake QS=207485725372997155 QS=1015150146901494114782QS=403466343733903314 Couple QS=207816747274867300 QS=1013649128251293012685QS=402835283425432478 Plane QS=205188503748614666 QS=108984854383608097QS=402244224421422025 Peppers QS=204107404439283711 QS=107927750474937233 qualities were parameterized by the quantizer step size QS for the quantizaton of the wavelet coefficients.Four different methods were compared:Decomposition of the significance maps into pred andres as in the preliminary decomposition scheme de-scribed in Section2.1,and encoding pred and res bycontext-based arithmetic coding(AC);Decomposition of the significance maps into pred andres by magnitude and distance preferences,as describedin Section2.3,and encoding pred and res by context-based arithmetic coding(AC+R);Decomposition of the significance maps into pred andres as in Section2.1,and encoding pred and res byhierarchical enumerative coding(HEC);Decomposition of the significance maps into pred andres as in Section2.3,and encoding pred and res byhierarchical enumerative coding(HEC+R).Table2:Number of bits spent by each method for encoding the res sequences.Image AC AC+R HEC HEC+RQS=40347328338313 Lenna QS=20527503521497QS=101205118012201197QS=40947948909887 Barbara QS=201458150013571360QS=101252120812491186QS=401912184219581871 Bridge QS=202533241926012462QS=101428142414711417QS=40652620654620 Lake QS=201251118212711198QS=103684356637293588QS=40776728762711 Couple QS=201108104711121042QS=101611155616541564QS=40372347371317 Plane QS=20606548622531QS=10793731817724QS=40246228244222 Peppers QS=20513486526490QS=103296326433823334 For arithmetic coding,a free software by Moffat et al. based on[7]and[8]was used with adaptivefixed-context bit-based model option.Hierarchy parameters of HEC were selected as,,,and for .With these parameters,encoding and decoding can be performed using64-bit integer additions and comparisons.The number of bits spent by each method for encoding the pred and res sequences are tabulated in Table1and Table 2,respectively.Let the overall performancefigure of each tested method be defined as the average number of bits spent by the method for encoding pred and res sequences normal-ized to the average number of bits yielded by AC.The overall performancefigures of the tested methods are as follows: AC:100AC+R:96.8HEC:95.1HEC+R:92.4The joint improvement brought by the proposed modifi-cations is7.6%on the average,and the percentage improve-ments do not vary so much with respect to the employed quantizer step size.4.CONCLUSIONSWe have proposed several modifications to a state-of-the-art method for encoding significance maps of wavelet coeffi-cients.The modifications are based on the assumption that the sequences pred and res are locally stationary.Consid-erable and consistent improvement brought by these modifi-cations verify this assumption in some sense.The reorder-ing of the sequences worksfine with the arithmetic coder as well,although this reordering reduces the overall stationarity of the sequences.This is possibly due to the probing of the time-variations of the statistics by the16-bit context.Although we do not provide results for the number of bits spent for encoding the coefficient magnitudes,it is worth mentioning that the suggested reordering also improves the compressibility of the coefficient magnitudes.In other words, when the nonzero coefficient magnitudes are scanned accord-ing to the suggested reordering,entropy coding(e.g.arith-metic coding)of the coefficient magnitudes yields less bits.5.REFERENCES[1]S.D.Servetto,K.Ramchandran,M.T.Orchard,“Im-age Coding Based on a Morphological Representation of Wavelet Data”,IEEE Transactions on Image Pro-cessing,vol.8,no.9,pp.1161-1174,1999.[2]J.M.Shapiro,“Embedded Image Coding Using Ze-roTrees of Wavelet Coefficients”,IEEE Transactions on Signal Processing,vol.41,no.12,pp.3445-3462, 1993.[3]R.¨Oktem and K.Egiazarian,“Morphology Based Im-age Compression with R-D Optimized Wavelet Packet Transform”,Proc.of Norsig’98,June8-11,Denmark.[4]B.Chai,J.Vass,and X.Zhuang,“Significance-LinkedConnected Component Analysis for Wavelet Image Coding”,IEEE Transactions on Image Processing,vol.8,no.6,pp.774-784,1999.[5]L.¨Oktem and J.Astola,“Hierarchical EnumerativeCoding of Locally Stationary Binary Data”,Electronics Letters,vol.35,no.17,pp.1428-1429,August1999.[6]J.Rissanen and ngdon,“Universal modelingand coding,”IEEE Transactions on Information The-ory,vol.IT-27,pp.12-23,January1981.[7]A.Moffat,R.Neal and I.H.Witten,“Arithmetic cod-ing revisited,”Proceedings of Data Compression Con-ference,Snowbird,Utah,pp.202-211,March1995. [8]P.M.Fenwick,“A new data structure for cumulativeprobability tables,”Software-Practice and Experience, vol.24,no.3,pp.327-336,1994.。