Integrated Intelligent Research (IIR) International Journal of Computing Algorithm

Volume: 01 Issue: 01 June 2012 Page No.7-13

ISSN: 2278-2397

Image and Video Compression Techniques in

Image Processesing an Overview
(Audio, Video, Image Compression Techniques)
S.Sukumaran1, T.Velumani2
Associate Professor, Department of Computer science, Erode Arts & Science College, Erode, Tamilnadu, India
1

2
Research Scholar, Department Of Computer Science, Kongu Arts & Science College, Erode, Tamilnadu, India
E-mail: [email protected], [email protected]

Abstract: At the present time, we stand upon the threshold of meant that almost all jobs around the house had to be done
a Revolution in the means available to us for the widespread manually, and no communication meant, well, no
dissemination of information in visual form through the communication. Telephone systems were in their infancy,
rapidly increasing use of international standards for image
there were no broadcast or television services, and this left
and video compression. Yet, such standards, as observed by
the casual user, are only the tip of the coding iceberg. newspapers as the only widespread information provider (and
Something like half a century of scientific and technological only one way at that). Person-to-person contact was carried
development has contributed to a vast body of knowledge out either face-to-face or by letter. Probably the greatest
concerning techniques for coding still and moving pictures, influence (although there are many contenders) on changes
and the present article presents a survey of developments in social attitudes has been our growing .
which have taken place since the first (predictive) coding
algorithms were implemented in the 1950s. Initially, we Ability to engage in almost instant broadcast and person-to-
briefly review the characteristics of the human eye which person communication over ever greater distances (in some
influence how we approach the design of coding algorithms; cases it might even be argued that there now exists more
then we examine the still picture techniques of major communication provision than we need or is, strictly speaking,
interest predictive and transform coding, vector quantization, good for us). Undoubtedly, this change has been fostered by
and sub band and wavelet multi resolution approaches.
Recognizing that other forms of algorithm have also been of the widespread supplanting of analogue by digital technology
interest during this period, we next consider such techniques over the past 3 decades or so (although, paradoxically, the
as quad tree decomposition and segmentation before looking final link in the chain, the telephone line, radio link, or
at the problems arising from the presence of motion and its whatever, may still well be analogue in nature), for this has
compensation in the coding of video signals. In the next allowed us to do three things much more easily than before: (a)
section, various approaches to the coding of image carry out signal processing operations very rapidly; (b) build
sequences are reviewed, and we concentrate upon the now
very complex large scale systems; and, most important, (c)
universally used hybrid motion compensated transform
algorithm before examining more advanced techniques such store data easily. Where would telecommunications
as model and object based coding. Of course, the key to technology be, for example, if it were still as difficult to
widespread acceptance of any technological development is store information as it was, say, 50 years ago? So where
the establishment of standards, and all major proposals PEG, does image coding fit into all this? Throughout history,
MPEG-I, II, and IV, H.261, and H.263, are considered with pictures have always had a high profile role to play in
emphasis on the way in which the coding algorithm is communication. In the days when the majority of people
implemented rather than on protocol and syntax
could not read whatever written words were available to
considerations. Finally, comments are offered in respect of
the future viability of coding standards, of less well them anyway, images allowed an immediacy of impact and
researched algorithms, and the overall position of image and directness of contact achievable in no other way. Later,
video compression techniques in the rapidly developing field representation of moving pictures, as provided by the film
of visual information provision. and, subsequently, television, enhanced this capability to an
enormous degree. It is impossible to appreciate the impact of
K e y w o r d s : MPEG, JPEG, Quadtree, ltiresolution, H.261 moving color picture presentation (something we take for
granted) on anyone who has not grown up with the idea. In the
I. INTRODUCTION case of television, however, it was quickly realized that
communicating a video image was vastly more expensive in
As this century draws to a close, it is virtually impossible to terms of necessary channel capacity than was speech or
imagine what life must have been like for the average man in music transmission; and even from early days, methods of
the street 100 years ago. No cars meant either cycling to work reducing this requirement were sought. Given the constraints
or going by rail or on foot, no aircraft left international operating at the time, the development of interlace, for
journeys to be undertaken by sea, no domestic appliances
7
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

example, has to be seen as an elegant practical solution to the image will be around 512 3 512 in extent, quantized to 8-bit
problem (despite the drawbacks it presents today for digital (256 levels) amplitude resolution and either monochrome or
processing of conventional video material). Again, early color. In the latter case, it is usual to move from a three color
studies showed that the video signal could advantageously plane (RGB) representation to a television style YUV or YIQ
be split into different frequency bands and these sent formulation. It is found that the additional load on the coding
separately and yet needing, overall, less capacity than the system due to the color information is relatively small—the
original signal. With the development of digital services, a luminance signal carries all of the detail resolution
growing problem soon appeared in that it was no longer a information and the U and V or I and Q data can be
matter of dealing with a few highly specialized application subsampled and coded using the same algorithm as the
areas, but rather a wide ranging spectrum encompassing luminance term, but only needing something like 20% of
person-to-person (videophone) communication at one end the rate. Figure 1 shows the first frame of the sequence
through videoconference and conventional television to high “CLAIRE”: 256 3 256 in extent with 8-bit amplitude
definition television at the other. Of course, supply and resolution.
demand are inextricably linked in this field, as elsewhere.
Given the capability of sending images and video efficiently, A . Predictive Coding.
more people want to make use of the facility, and at the same
time, they generate new application areas, and so on. Wide It is usually the case that useful images comprise
spread use over many different fields of application. Also recognizable objects— cars, buildings, people, etc.—
depends upon the development of and agreement on, standards, portrayed against similarly recognizable areas of reasonable
which is the point at which we stand at the moment. In a extent.
broad sense, we can now assume that the first phase of the
development of image coding has been completed.
Established, well regarded techniques are in place in
several standards for still picture, video communication,
broadcast use, etc., and there is some uncertainty about
where we should be going next. There are also other well
researched and efficient techniques which did not make it
into the standards, but which nevertheless may still be of use
in particular areas of application.

On the other hand, it is hard to escape the realization that, Figure 1. Predictive coding.
notwithstanding past successes, some fresh ideas are now
necessary to impart a new direction to coding development and Only for regions of high detail or texture are this unlikely to
also free us from some of the problems which still bedevil be true and these tend to occur over only a small part of the
what we have come to call conventional approaches. Again, image or video sequence. This being so, we can make use of
the enormous bandwidth offered to us by optic fiber may the high degree of correlation existing within the picture to
remove the need for image coding algorithms to be used in reduce the number of bits necessary for reproduction, and the
fixed service applications, leaving mobile systems as the algorithm which does this in the most obvious fashion is
major user of such techniques in an effort to conserve finite, predictive coding. Deriving from work on general signal
and thus precious, spectrum space. There are thus many prediction in the 1940s, this algorithm was the first image
factors, not all of a purely technical nature, which will coding technique to be extensively researched starting in the
determine the future course of image coding research. In this 1950s, and in its basic form it still finds place (albeit in the
article, we initially review the status of well researched time dimension) in the standards algorithms of today. It is
techniques; both those now incorporated into standards and very simple and capable of reasonable rate reduction with
others of similar standing, and then consider how the excellent quality when applied to still pictures (O’Neal. 1996;
standards work of the last 10 years has employed such Musmann, 1979). From the values of previously scanned
algorithms for the processing of a wide variety of image data. picture elements at both transmitter and receiver, predictions
Thereafter, we consider the possibilities for new directions in are made of the next sample which the coder is to process.
image coding. This prediction is subtracted from the actual value of the
next sample and the error signal (difference) quantized and
II. BASIC CODING TECHNIQUES transmitted, to be added to the prediction made at the decoder
to form the reconstructed signal. The procedure is shown in
In this and the next section, basic algorithms for coding still Figure 2. For a good prediction the error will be small, but
images (intraframe coding) are described. Typically, the just as important is the fact that its probability distribution is
8
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

very well behaved compared with that of the original picture, The system has the advantage for certain applications that all
and of Laplacian or even Gamma distribution form—very the processing power is required at the transmitter, the
highly peaked around zero with relatively few large values, receiver/decoder being trivially simple— one transmitted
which latter will be a consequence of the edge detail in the index word per block is all that is needed to access the
picture. Coding this signal with maybe three bits per element codebook (look-up table) entry for display, maybe with some
and using no uniform quantization and variable word length simple scaling operation. As a simple example, if we have
coding allows good quality to be achieved for images 1024 represen- tative entries, a 10-bit index is needed. If
containing television type detail (Musmann et al., 1985). Of this identifies a 4 3 4 block, then the data rate is about 2/3
course many refinements are possible; both predictor and bit/element. In practice, some further sophistication is needed
quantized may be made adaptive to account for localized in the algorithm to cope with the artifacts which would be
detail (needing transmission of overhead information to the produced by such a basic scheme.
receiver), possibly through the use of a least mean squares
updating algorithm (Alexander and Rajala,1984) or, more There are many methods of generating the codebook, one
recently, through the use of higher order statistics (Tekalp et tried and tested example of which was reported in 1980
al., 1990). Alternatively, neural approaches may be used to (Linde et al., 1980). Given a first try reproduction codebook,
optimize predictor structure (Dony and Haykin, 1995). Setting all vectors from a suitable training sequence are allocated the
the basic element by-element algorithm in a block context in closest entry according to some distance measure (mean
this way and maybe using arithmetic coding for the output square, or mean absolute, energy of the error vector, for
signal allows rates down to around one bit per element to be example).
achieved.
Optimization proceeds by determining the new best code
word for each of the partitions of training vectors so produced
B. Vector Quantization.
and then iterating the process. An initial codebook may be
Transform coding makes use of the interelement correlation produced by first finding the one optimum code word for
within a picture to concentrate coding capacity on the the whole of the training sequence, splitting it into two close
dominant low frequency terms produced in the corresponding but different vectors, and then proceeding as above. This
low ‘frequency’ block representation. Another technique codebook generation process is intensive both in time and
which makes use of similarities within the data, albeit at a computation, as is the other basic operation needed for coding
higher level, is vector quantization (Gray, 1984). In scalar an input vector: full search of the codebook for the nearest
quantization, data amplitudes are reconstructed as the nearest reproduction vector to determine the appropriate index to be
predetermined value to that actually occurring within any transmitted. Most research on vector quantization since its
given (one-dimensional) space between two adjacent decision introduction for image coding in the early 1980s has
levels. Given the usual level of interelement correlation concentrated on these two problems, and a multiplicity of
within pictures, it is evident that we could jointly quantize methods is now available for their (partial) solution:
pairs of neighboring picture elements as points in a two- applying the splitting process preferentially to those nodes
dimensional decision space and, apart from quantization in giving greatest decrease in distortion for the smallest increase
ac- curacies, gain efficiency by so doing; pairs of similar overall in the number of codebook entries, and maybe using
values are much more likely to occur than those which are multiple splits as well. Separating out block mean value and
widely different (these latter representing edges). Vector standard deviation (corresponding to activity level) for
quantization extends this idea to a larger region (a usual separate transmission can also be helpful (Murakami et al.,
approach is to take a 4 3 4 picture block, considered as a 1982) (Figure 4), as can classification of codebooks
vector of length 16) to produce results which rival transform according to the presence of strong directional detail
coding in terms of the quality of reproduction at rates around (Gersho and Ramamurthi, 1982). Neural optimization
0.5 bit/element, particularly if adaptive schemes are used techniques can also be employed (Dony and Haykin, 1995;
[Panchanathan and Goldberg, 1991). In vector quantization, Lee and Petersen, 1990).
use is made of the fact that many picture blocks will be
Likewise, fast search methods have been intensively
very similar (background, interiors of large objects, etc.) in
researched, with all manner of partial, tree, and approximate
terms of luminance, color, and so on, or maybe contain
searches contributing to the speed up of the process. It is also
strong detail of the same orientation. Such blocks will all
possible to use a regular lattice structure for the codebook
be displayed as the same representative block chosen from a
(Chen, 1984). This has the advantage that no actual
codebook of typical blocks (vectors) via some appropriate
codebook need be stored and processing is very rapid—
distance measure.
especially beneficial in video applications (see later). Over
the years, vector quantization has evolved into an efficient

9
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

coding technique which may either be used on its own or as a responsible for this process. Such latter algorithms can be
postprocessing step to code the output of a previous algorithm much more efficiently matched to the individual sub band
—arrays of transform coefficients, for example. characteristics than to the image as a whole, and this allows
good results in the 0.5 bit/element region to be achieved
C. Sub band and Wavelet Coding. (Lookabaugh and Perkins, 1990). Usually, the band split will
be more extensive than the simple 2 3. 2 split described
One of the earliest approaches suggested for the reduction
above. The reapplication of this step to the outputs of the
in bandwidth or channel capacity for the transmission of
initial split will result in a 16-band (4 3 4) structure.
image data was frequency division, in which the total
Alternatively, it may be better to split lower frequency bands
bandwidth is split, at the simplest level, into low and high
more finely still and leave the highest horizontal and vertical
frequency segments. The low frequency band has reduced
frequency band unsplit—it will rarely contain any detail of real
resolution and so needs far fewer bits for transmission; the
significance. It might be noted here that there are close
upper frequency band will generally have few significant
connections between sub band and transform coding.
components and can likewise be easily coded. Over the past 10
Sequential multilevel filtering using a simple high/low split of
years or so, this Ii 5 input signal; Io 5 output signal; LA and
the kind mentioned above can be shown to produce the same
HA 5 low and high pass basic analysis filters; LS and HS 5 result as a transform operator; indeed, transform coding may
low and high pass synthesis filters; K(s) 5 sub sampling at be considered to be a form of sub band coding in which each
the coder; K(i) 5 interpolation at the decoder. sub band contains only one coefficient.

Lately there has been much interest in the analysis of

signals whose properties change with time. For this purpose,
the Fourier transform is not suitable, since its baseline (in
theory at least) is infinite, and the necessary window
function added for practical implementation always
introduces undesirable artifacts into the analysis. Out of this
interest has come intensive research into wave lets (Chui,
1992). Such functions have restricted support and can allow
Figure 2. Normalized vector quantization. flexible and efficient time (or space)/frequency processing,
with good resolution at low frequencies via long windows, and
good spatial resolution at high frequencies obtained by using
short, wide bandwidth ones. As far as image coding is
concerned, wavelet processing is not unlike sub band coding
in basic outline, save that it is usually carried out in a
multiresolution context.

Here, an initial2 3 2 split and subsample operation is

carried out as previously described, following which only
the lowest frequency sub band is resplit. At each level, this
results in one low resolution image and three so called detail
Figure3. Basic filter structure for subband coding.
images. We can thus generate a hierarchy of images at
Idea has developed into a powerful and flexible image various levels of resolution which will allow image
reconstruction for a wide variety of applications from the one
coding technique: sub band coding (Woods, 1991). Here, the
image spectrum is split basically into two components as data stream. The smallest and lowest-resolution image might
be used as the initial output of a database search, for
above, but now both horizontally and vertically, which are
then sub sampled by the same factor to give four two example. If higher resolution is needed, this signal is up
sampled (interpolated) and the three detail signals at the next
dimensional frequency bands containing combinations such
as low horizontal and low vertical frequencies, low level added, and so on, until the final full resolution version
is obtained. As usual, predictive coding or vector
horizontal and high vertical frequencies, and so on, the total
number of samples being unchanged. The basic one quantization can be used to code the various individual bands
involved, and the scheme may be significantly improved
dimensional scheme is shown in Figure 5. As with
predictive and transform coding, no data reduction is by the inclusion of tree structured algorithms for tracking
significant coefficients through the various wavelet levels
achieved by the first part of the algorithm, and it is the
efficient coding of the various sub bands using predictive (Schapiro, 1993; Said and Pearlman, 1996). Wavelet coding
represents a sophisticated version of multi resolution
coding, or vector quantization, for example, which is
10
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

decomposition whereby one algorithm may have a variety of workers continued for the next decade or so, but problems
image qualities at the output. It is appropriate to mention with reliable determination of changing areas and choice of a
here the original impetus for the idea, which has been with suitable initial estimate for the recursion meant that
us for some 20 years now. In this realization, lower alternative schemes came into prominence and, as will be seen,
resolution images were produced from higher level ones by were eventually incorporated into standard algorithms. The
Gaussian filtering and sub sampling and then expanded technique most widely used at present is based upon the
(interpolated) again to be used as a prediction for the upper taking of a block of elements in the present frame and
level image. The prediction error then corresponds to the simply searching for a similar block in the previous frame
detail signal and, since it has a Laplacian probability which minimizes some function of the frame-to-frame
distribution, and, moreover, appears as a set of image signal difference over the area—mean square or mean absolute
levels, one above the other, the idea of the Laplacian pyramid error (Jain and Jain, 1981). Prior to search, it can be
emerged (Burt and Adelson, 1983). Multi resolution advantageous to test the initial error against a threshold; if it is
decomposition is of great significance, given the variety and small enough motion compensation is not needed, anyway.
scope of present day digital image services, and the wavelet The relative locations of blocks in present and previous
approach is more elegant and flexible than many other frames are then characterized as a motion vector which must
methods (using the DCT, for example) which have been be transmitted to the decoder. This simple correlation like
proposed for this purpose. technique is computationally intensive (search over a 67
element displacement in both x and y directions requires the
III. MOTION error to be calculated at over 200 locations); and
although it is now possible at real time rates and is indeed
If there is one thing above all that the prevalence of the preferred approach, as if fully searched it guarantees to
television throughout the world as a provider of information find a true minimum in the error function, the literature
and entertainment demonstrates, it is the overwhelming contains a long history of reduced search approaches— using a
preference of the human observer for moving images. The search route covering only a selection of locations, minimizing
case of still pictures, there has long been a concomitant x and y error terms sequentially, and so on (Kappagantula
interest in ways of processing these at as low a rate, given and Rao, 1985). An added advantage is that, having all
quality of reproduction constraints, as possible. It is only with possible error values, displacement can be determined via
the comparatively recent development of large scale, on chip interpolation to accuracy better than a single element. A
storage, however, combined with the ready availability of refinement which can aid fast determination of the
ultra high speed hardware, which it has become practicable to location corresponding to the minimum frame-to-frame error
implement such schemes. One way, of course, is simply to is hierarchical block matching (Wang and Clarke, 1990).
process image sequences on a frame by frame basis, i.e., Here, an image pyramid consisting of a sequence of planes,
without regard for any interrelation between them. Just as it each formed by averaging over a small region of the previous
is logical, though, to consider objects rather than arbitrary one, is used top to bottom: A rapid search at the lowest
square blocks in the case of still images, so too, these objects resolution level forms the initial estimate for the next, and so
are not only present, but also move within image sequences, on. This approach is also useful in dealing with large frame-
and so the estimation of motion and its compensation have to-frame differences.
assumed increasing importance in image coding over the past
20 years or so. Motion estimation and compensation form part of the
standard present day approach to low rate video transmission,
Early work on motion estimation is represented by but their applicability is wider that simply optimizing frame-
algorithms involving both the space and the frequency to-frame prediction. They can also be used for interpolation,
domain (Limb and Murphy, 1975; Haskell, 1974). In the when maybe every other frame in a sequence is dropped to
former, the ratio of frame-to- frame differences, over the achieve a minimum bit rate (Thoma and Bierling, 1989). In
moving area, to the sum of element to element differences in this situation, simple static interpolation produces
the present frame was used to give an object speed measure, unacceptable motion artifacts when used to reconstruct the
whereas in the latter the Fourier shift theorem can give an missing frames, and motion compensation enables the
indication of motion via its phase shift term. Work started in movement of the object(s) to be accounted for in this operation
earnest, however, in the late 1970s with the development (note that in this case, vectors representing true motion are
of a recursive steepest descent algorithm which minimized necessary, not simply those which indicate a minimum in
the inter frame difference signal, in an algorithm which could the frame difference signal). Areas covered up and
also be modified to account for problems with changing uncovered by the moving object are accounted for by forward
illumination (Netravali and Robbins, 1979; Stuller et al., and backward extrapolation, respectively. Another application
1980). Intensive development of this algorithm by various of motion compensation is in noise filtering, where, if the
11
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

motion estimate is good, strong low pass filtering may be still or moving? How can we perform segmentation even
applied along the object path to reduce noise (Dubois, with a tiny fraction of the ease with which the eye does it?
1992).Naturally, the assumption that square block translation Oddly enough, what was once considered to be the major
represents true object motion is only approximate, and stumbling block, processing speed (or the lack of it) seems to
experiments have been done in an attempt to allow more have disappeared from the equation. No longer do we have to
refined tracking of object rotation, scale change, and general accept compromises in algorithm design because the hardware
motion throughout a video sequence (Wu and Kittler, 1990). cannot cope, and the days of an overnight job at the
There is also continuing work on effective frequency computer center simply to transform code a single image frame
domain algorithms for motion compensation (Young and now seem like just a bad dream.
Kingsbury, 1993). The area in general is one where we are yet
some way from the goal of being able to track object So where do we go from here? For fixed services, there seem to
motion reliably through a sequence of frames and employ the be two distinct opinions. One says that given the enormous
knowledge so gained in reducing yet further the data rate for bandwidth of optical fiber, the provision of new image services
video transmission. can be handled with only moderate levels of compression, or
even none at all (this view ignores the fact of life that, given a
IV. GENERAL COMMENTS AND CONCLUSIONS new source of transmission capacity, human ingenuity will soon
find ways not only of filling it up, but of finding reasons for
No one needs reminding nowadays that as we approach the requiring still more). The alternative argues that even so, there
millennium, we are living in a time of phenomenal change in will always be a place for satellite and conventional broadcast
our ability to access information. Undoubtedly, this has been distribution, in which case compression algorithms will play a
due to the development of the computer and large scale high vital part in providing choice within an ever increasing volume
speed integrated circuits, and these have contributed to major of program material. Whichever of these holds sway eventually,
advances in communication via speech and the written word. we can be safe in the knowledge that it is difficult to attach a
Overwhelmingly, however, it has been the image that has been fiber optic link to a moving vehicle, and finite radio channel
in the forefront of this revolution as even a cursory space together with an ever increasing demand for video
examination of our image processing capabilities half a communication in this context (public service applications, for
century ago and now will reveal. Oddly enough, we may be example) is a guarantee that image and video compression
in the position we are now just because pictures, especially techniques will become and remain common in this area. Again,
in moving form, contain so much information that they the standardization activity of the past 10 years has to be seen by
present an interesting and relevant challenge to the any criterion as a monumental achievement in drawing together
technology existing at any particular point in time as far as scientific, technological, and economic and commercial
data reduction developments are concerned. considerations. Yet, in one sense, all arguments about standards
and rationalization may not matter at all—the diversity of
From the early days of predictive coding, via extremely
applications and techniques for image and video coding may
powerful transform coding algorithms and the development of
mean that I as a service provider can arrange for you as a
variable word length coding and motion compensation, all of
consumer to download my decoding software prior to
which have given us the standards we have today,
transmission of the data, and we can all use whatever
necessary transmission rates have fallen from megabytes to
algorithm is dictated as economic by the actual application.
kilobytes per second; in addition, a pleth ora of alternative
Therefore, it may well be that there is room for everyone after
techniques has grown up, all of which, even if not
all. The Impact of enhanced algorithms can be justified with high
standardized or, in some cases, very successful, have taught us
speed derivate hardware algorithms. Thus Hardware architectures
more about the relation between the human observer and
are the essential base to build the high efficient algorithm.
the visual representation of scenes. And priorities? In 1985,
it was suggested that conventional algorithms had reached
their limit and that we should be coding something other than
REFERENCES
[1] N. Ahmed, T. Natarajan, and K.R. Rao, Discrete cosine transform,
square image blocks. There- fore, here we are 13 years later IEEE Trans Compute C-23, (1974), 90 –93.
with all of our standards still based on just that approach, but [2] T. Akiyama, T. Takahashi, and K. Takahashi, Adaptive three-
with quite a bit more compression and flexibility to show, dimensional transform coding for moving pictures, Proc Picture
whereas in parallel, object and model based approaches Coding Symp, Cam- bridge, MA, 26 –28 March, 1990, Paper 8.2.
[3] S.T. Alexander and S.A. Rajala, Optimal gain derivation for the
creep painfully toward more generic and robust application. LMS algorithm using a visual fidelity criterion, IEEE Trans Acoustic
Shall we now say that the old methods have really reached Speech Signal Process ASSP-32 (1984), 434 – 437.
the end of the road? We can always argue, of course, that [4] A.D. Barnsley and A.D. Sloan, A better way to compress images,
BYTE, January (1988) 215–223.
more research is needed, and so it is, especially in bringing
[5] P.J. Burt and E.H. Adelson, The Laplacian pyramid as a compact
the HVS into the system how do we really perceive images, image code, IEEE Trans Commun COM-31 (1983), 532–540.

12
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

[6] C.F. Chen and K.K. Pang, Hybrid coders with motion compensation, [33] H.G. Musmann, M. Hotter, and J. Ostermann. Object-oriented
Mul tidimen Syst Signal Process 3 (1992) 241–266. analysis- synthesis of moving images, Signal Process Image
[7] T.C. Chen, A lattice vector quantization using a geometric Commun 1 (1989),117–138.
decomposition, IEEE Trans Commun COM-38 (1984), 704 –714. [34] H.G. Mussman, P. Pirsch, and H.-J. Grallert, Advances in picture
[8] W.H. Chen and W.K. Pratt, Scene adaptive coder, IEEE Trans coding, IEEE Proc 73 (1985), 523–548.
CommunCOM-32 (1984), 225–232. [35] A.N. Netravali and J.D. Robbins, Motion-compensated television
[9] W.H. Chen and C.H. Smith, Adaptive coding of monochrome and coding: Part 1, Bell System Techn 58 (1979), 631– 670.
color images, IEEE Trans Commun COM-25 (1977), 1285–1292. [36] J.B. O’Neal, Jr., Predictive quantizing systems (differential pulse code
[10] C.K. Chui, Wavelets: A tutorial in theory and applications, Academic mod- ulation) for the transmission of television signals, Bell Syst Techn
Press, San Diego, 1992. 45 (1966),
[11] R.J. Clarke, Transform coding of images, Academic Press, San Diego, [37] S. Panchanathan and M. Goldberg, Min-max algorithm for image
1985. adaptive vector quantization, IEE Proc Commun Speech Vision 138
[12] R.D. Dony and S. Haykin, Neural network approaches to image (1991), 53– 60.
compres- sion, IEEE Proc 83 (1995), 288 –303. [38] F. Pereira and R. Koenen. Very low bit rate audio-visual applications,
[13] E. Dubois, Motion-compensated filtering of time-varying images, Signal Process Image Commun 9, 1996, 55–77.
Multidi- mentional System Signal Process 3 (1992), 211–239. [39] K.R. Rao and J.J. Hwang, Techniques and standards for image, video
[14] S. Ericsson, Fixed and adaptive predictors for hybrid and audio coding, Prentice Hall, Upper Saddle River, New Jersey,
predictive/transform coding, IEEE Trans Comm COM-33 (1985), 1996.
1291–1302. [40] Said and W.A. Pearlman, A new, fast, and efficient image codec based
[15] R. Forchheimer and T. Kronander, Image coding—from waveforms on set partitioning in hierarchical trees, IEEE Trans Circuits Syst Video
to animation, IEEE Trans Acoust Speech Signal Process ASSP-37 Technol 6 (1996), 243–250.
(1989),2008 –2023. [41] H. Samet, The quadtree and related hierarchical structures, Comput
[16] Gersho and B. Ramamurthy, Image coding using vector surveys 16 (1984), 187–260.
quantization, ICASSP Proc, 1982, pp. 428 – 431. [42] D.G. Sampson and M. Ghanbari, Interframe coding of images using
[17] Girod, Psycho visual aspects of image communication, Signal Process lattice vector quantization, Proc 4th IEE Int Conf Image Process,
28 (1992), 239 –251.B. Girod, N. Faerber, and E. Steinbach, Standards Maastricht, Netherlands, 7–9 April, 1992, pp. 131–134.
based video communications at very low bit rates, EUSIPCO Proc, [43] R.A. Schaphorst, Status of H.324 —the videoconference standard
1996, pp. 427– 430. for the public switched telephone network and mobile radio, Opt
[18] R.M. Gray, Vector quantization, IEEE ASSP Mag, April (1984), pp. 4 Eng 35 (1996),109 –112.
–29. A. Habibi, Hybrid coding of pictorial data, IEEE Trans Comm [44] J.M. Schapiro, Embedded image coding using zerotrees of wavelet
COM-22(1974), 614 – 624. coeffi- cients, IEEE Trans Signal Process SP-41 (1993), 3445–3462.
[19] B.G. Haskell, Frame-to-frame coding of television pictures using two- [45] T. Sikora, The MPEG-4 video standard verification model, IEEE Trans
dimen- sional Fourier transforms, IEEE Trans Inf Theory IT-20 (1974), Ckts Syst Video Technol 7 (1997), 19 –31.
119 –120. [46] M. Soryani and R.J. Clarke, Segmented coding of digital image
[20] J. Huguet and L. Torres, Vector quantization in image sequence sequences, Proc IEE I Commun Speech Vision 139 (1992) 212–218.
coding, EUSIPCO Proc, 1990, pp. 1079 –1082. [47] P. Strobach. Quadtree-structured recursive plane decomposition
[21] A.E. Jacquin, Image coding based on a fractal theory of iterated coding of images, IEEE Trans Signal Process SP-39 (1991), 1380 –
contractive image transformation, IEEE Trans Image Proc 1 (1992), 1397.
11–30. [48] J.A. Stuller, A.N. Netravali, and J.D. Robbins, Interframe television
[22] J.R. Jain and A.K. Jain, Displacement measurement and its coding using gain and displacement compensation, Bell Syst Techn J
application to interframe image coding, IEEE Trans Commun COM- 58 (1980),1227–1240.
29 (1981), 1799 –1808. [49] A.M. Tekalp, M.K. Ozkan, and A.T. Erdem, Image modeling using
[23] S. Kappagantula and K.R. Rao, Motion compensated interframe higher- order statistics with application to predictive image coding,
image prediction, IEEE Trans Commun COM-33 (1985) 1011–1015. ICASSP Proc,1990, pp. 1893–1896.
[24] M. Kunt, A. Economopoulos, and M. Kocher, Second generation [50] R. Thoma and M. Bierling, Motion compensation interpolation
image coding techniques, IEEE Proc 73 (1985), 549 –574. considering covered and uncovered background, Signal Process
[25] T.C. Lee and A.M. Peterson, Adaptive vector quantization using a Image Commun 1 (1989),191–212.Q. Wang and R.J. Clarke, Motion
self- development neural network, IEEE J Select Areas Commun 8 compensated sequence coding using image pyramids, Electron Lett 26
(1990), 1458 –1471. (1990), 575–576.
[26] H. Li, A. Lundmark, and R. Forchheimer, Image sequence coding at [51] H. Watanabe and Y. Suzuki 64 kbit/s video coding algorithm using
very low bitrates: A review, IEEE Trans Image Proc IP3 (1994), 589 – adaptive gain/shape vector quantization, Signal Process Image
609. Commun 1 (1989),87–102.
[27] J.O. Limb and J.A. Murphy. Measuring the speed of moving objects [52] P.H. Westerink, J. Biemond, and G. Muller, Subband coding of image
from television signals, IEEE Trans Commun COM-23 (1975), 474 – sequences at low bit rates, Signal Process Image Commun 2 (1990),
478. 441– 448.
[28] Y. Linde, A. Buzo, and R.M. Gray, An algorithm for vector quantizer [53] T. Wiegand, M. Lightstone, D. Mukherjee, T.G. Campbell, and S.K.
design, IEEE Trans Commun COM-28 (1980), 84 –95. Mitra, Rate-distortion optimized mode selection for very low bit rate
[29] T.D. Lookabaugh and M.G. Perkins, Application of the Princen- video coding and the emerging H.263 standard, IEEE Trans Ckts Syst
Bradley filter bank to speech and image compression, IEEE Trans Video Technol 6 (1996), 182–190.
Acoust Speech Signal Process ASSP-38 (1990), 128 –136. [54] J.W. Woods (Editor), Subband image coding, Kluwer, Dordrecht,
[30] H.S. Malvar, Signal processing with lapped transforms, Artech House, 1991.
London, 1992. [55] S. Wu and J. Kittler, A differential method of simultaneous
[31] T. Murakami, K. Asai, and E. Yamazaki, Vector quantizer of video estimation of rotation, change of scale and translation, Signal Process
signals, Electron Lett 18 (1982) 1005–1006. Image Commun 2 (1990), 69 – 80.
[32] H.G. Musmann, “Predictive image coding,” Image transmission [56] S. Yao and R.J. Clarke, Motion-compensated wavelet coding of
techniques (Advances in electronics and electron physics, Suppl 12), colour images using adaptive vector quantization, Proc Conf Image
W.K. Pratt (Edi- tor), Academic Press, New York, 1979, pp. 73–112.

13
Integrated Intelligent Research (IIR) International Journal of Computing Algorithm
Volume: 01 Issue: 01 June 2012 Page No.7-13
ISSN: 2278-2397

Process, Theory and Applications, San Remo, Italy, 14 –16 June, 1993,
pp. 99 –102.
[57] R.W. Young and N.G. Kingsbury, Frequency domain motion
estimation using a complex lapped transform, IEEE Trans Image Proc
2 (1993), 2–17.
[58] L.H. Zetterberg, S. Ericsson, and H. Brusewitz, Interframe DPCM
with adaptive quantization and entropy coding, IEEE Trans Commun
COM-30 (1982), 1888 –1899.

14