MULTIMEDIA QOS IN CONSUMER TERMINALS

Reinder J. Bril, Christian Hentschel, Elisabeth F.M. Steffens, Maria Gabrani, G. (Sjir) van Loo and Jean H.A. Gelissen

Philips Research Laboratories Eindhoven (PRLE) Prof. Holstlaan 4, 5656 AA Eindhoven, The Netherlands

( Reinder.Bri1, Christian.Hentsche1, Liesbeth.Steffens, MariaGabrani, Sjir.van.Loo, Jean.Gelissen) @philips.com

Abstract - Over the past years, there has been a considerable amount of research in the field of QoS support for (distributed) multimedia systems, i.e. multimedia processing in, for example, a (networked) workstation environment. QoS for multimedia systems is about media processing in software, using dynamically scalable functions, and trading resources for quality. Unlike QoS for mainstream multimedia systems, QoS support for high volume electronics (HVE) consumer terminals (CTs), such as digital TV sets, digitally improved analog TV sets and STBs (set-top boxes), hardly received any attention in the literature. This paper considers multimedia QoS for consumer terminals, with focus on the high-quality video domain.

INTRODUCTION

Consumer terminals ( C T s ) are gradually evolving from straightforward terminals of a video broadcast network (TV sets) and a communication network (telephone) to interactive multimedia terminals, and beyond that, to elements in an in-home network, or even an ambient intelligence [ 11 environment.

Subjects addressed are multimedia QoS in high volume electronics (HVE) CTs, such as digital TV sets, digitally improved analog TV sets and STBs (set-top boxes). The basic media in CTs are high-quality audio and video. If the basic media processing functions are scalable, other media processing functions can be added at little or no extra cost. Scaling audio is less important than scaling video, for two reasons: combined with lower quality audio (e.g. mono), video is perceived at lower quality, and high-quality audio (e.g. multi-channel) consumes just a fraction of the resources compared to high-quality video. The challenge of multimedia QoS for CTs is in finding a QoS approach that can primarily be applied to high-quality video, and also supports other media, such as 3D graphics. The main focus of this paper will therefore be on QoS for high-quality video.

This paper has the following structure. The first section describes the evolution of CTs to consumer multimedia terminals (CMTs). The second section compares high-quality video processing in CTs to mainstream multimedia processing in a workstation environment. QoS parameters, such as frame rate for video

0-7803-7145-3/01/$10.00 02001 IEEE 332

applications in a workstation environment, are shown to be very application- domain specific. In our approach to QoS for CMTs, the QoS Resource Manager (QoS RM) therefore deals with system aspects that are semantically neutral. This section also shows that high-quality video has more stringent requirements than other media processing functions, justifying our focus on high-quality video. Scalable video processing is the topic of the third second section. The topic is addressed by presenting a general model for scalable algorithms and a video processing model based on the notion of jobs. QoS for video processing is the topic of the fourth section. This section presents our approach to QoS for CMTs, showing that our QoS approach has much in common with QoS for mainstream multimedia systems [2] despite the fact that the characteristics of both domains are very different.

EVOLUTION OF CONSUMER TERMINALS

In the last four decades, mid-range and high-end TV sets have evolved from stand-alone, analog hardware systems to digital systems containing megabytes of embedded software. Digitisation of television started with, and to a large extent still builds on, the digital enhancement of analog signals (e.g. 100 Hz TV). More recently, the emergence of Digital Video Broadcasting (DVB) gave rise to fully digital TV sets, and digital set-top boxes (STB) that can be combined with existing analog sets. The digitisation of television has clear technical benefits, such as better image and audio quality, more channels, and multi-window TV at moderate additional cost. Most of the digital signal processing is still performed in hardware, however. The embedded software in present-day TV sets is mainly in the area of control and services. Control tasks determine which hardware devices will be on the signal processing path, and what their settings will be, based on characteristics of the set and the incoming signal, and, of course, on user inputs. Typical software-based services include teletext (TXT), on-screen display (OSD), and menus. When combined with other devices, such as a telephone, game console, or a video recorder (VR), a TV set may provide additional services, such as TV- commerce and personalised teletext (with a telephone), games (with a game- console), and delayed viewing (with a VR).

Recent developments aim at new kinds of interactive services that drastically change the traditional role of a TV set. It is expected that a significant part of the interactive multimedia services that have evolved on the Internet will be transacted over the television in the near future [5 ] . In addition, the TV set (or STB) is expected to become a multimedia platform, which will serve as a gateway to the Internet [8], and may be integrated with other consumer devices within a home network. New architectures for STBs [15] and digital TV sets [25] explicitly address support for interactive services. As an example, the ITEA' project EUROPA' [7] aims at defining an STB reference architecture based on the DVB- MHP (Multimedia Home Platform) model, with extended functionality to enable

' ITEA is an acronym of Information Technology for European Advancement. * EUROPA is an acronym of End User Resident Open Platform Architecture.

333

next-generation services, such as secure online shopping and banking, based on multi-modal interaction (MPEG-4, MPEG-21), and secure transactions (cryptography). The project closely co-operates with standardisation bodies (MPEG, DVB), and is expected to contribute to the conception of a joint (European) STB architecture that will strengthen the position of European CE and IT-industry. The new interactive services require that STBs and TV sets, or with a general term CMTs, become open and flexible, not only in the area of control and services, but also in the area of media processing. Since openness and flexibility are typical characteristics of software based systems, significant parts of the media processing are expected to move from hardware to software. On the one hand, software media processing allows media functions to become dynamically scalable, i.e. dynamically adapt their quality to the availability of resources. With these scalable media functions, the new architectures can provide quality-of- service (QoS), i.e. trade quality against functionality within the resource restrictions of a given platform. On the other hand, software media processing is more expensive than equivalent processing on dedicated hardware, and more error- prone. Over time, robustness has become a basic requirement for CTs. No one accepts a TV to stall with the message “please reboot the system”. To meet the new challenges (openness and flexibility) and the existing ones (cost-effectiveness and robustness), future architectures must provide explicitly designed dynamic behaviour, explicit management of the available resources (including QoS), and stability under stress and fault conditions.

In the future, the consumer electronics (CE) industry may gradually move in the direction of ambient intelligence. That concept foresees a world in which the CE devices are integrated into the background of people’s environment (walls, clothing), and in which it is possible for any person to have access to any source of information, entertainment and communication (ICE) at any place and any time [l]. The CE devices that play a crucial role in the ambient intelligence environment will be based on product platform architectures, like the ones that are currently being developed for CMTs.

HIGH-QUALITY VIDEO PROCESSING IN CONSUMER TERMINALS

High-quality video processing in CTs has a number of distinctive characteristics when compared to mainstream multimedia processing in, for example, a (networked) workstation environment [ 191. In the first subsection, we compare QoS parameters for different types of media processing. These QoS parameters are shown to be very application domain specific. Moreover, high- quality video has very stringent timing requirements compared to other media processing functions. The specific requirements for consumer terminals are the topic of the second subsection.

334

QoS parameters for media processing The mesh, texture, and screen resolution are used as QoS parameters for 3D

computational graceful degradation in [lo], while maintaining a fixed frame rate. Frequently used QoS parameters for video applications in a workstation environment are screen resolution, frame rate, image size, colour depth, bit rate and compression quality [13], [18], [22]. Spatial (resolution) and temporal (bit rate and frame rate) scalabilities are exploited in great detail in the field of image compression (MPEG*, H.26*; see for example, [18]).

In CMTs with high-quality video requirements, these parameters are not generally applicable. High-quality video has a fixed field/frame-rate, no tolerance for jitter (i.e. frame-rate fluctuations), and low tolerance for frame skips, i.e. very stricgent timing requirements. Moreover, the resolution of a TV screen is fit to its standard (e.g. PAL, NTSC, ATSC, VGA), and the image (or window) size is either fixed (e.g. main window or Pip window) or determined by the user. Finally, receivers in a broadcast environment, and that is what CMTs currently are, 'do not have the option to negotiate compression quality and bit-rate, although that may change in the future for CMTs in an in-home digital network.

Although the setting of many parameters are imposed by the environment rather than adaptable by the system for optimisation purposes, they do determine the amount of processing required for a particular video output quality. As an example, the window size determines the embedded resizing techniques [27] that can be applied without loss of visual output quality.

Hence, alternative parameters have to be used for high-quality video. These parameters are typically video algorithm specific and may vary per algorithm. Optional parameters for high-quality video are number of filter coefficients (e.g. 0, 8, 32, ...), number of fields used (e.g. 1, 2, 3, ...), reference objects (e.g., points, lines, . . .), and type of processing (e.g. linear, non-linear). Examples of algorithms exploiting such parameters may be found in [9] (scalable sharpness-enhancement), [20] and [26] (scalable MPEG-2 video decoding by computing 061y a few IDCT coefficients).

System requirements Consumer products are heavily resource constrained, with a high pressure on

silicon cost and power consumption. In order to be able to compete with dedicated hardware solutions, the available resources will have to be used very cost- effectively, while preserving typical qualities of HVE CTs, such as robustness, and meeting stringent timing requirements imposed by high-quality digital audio and video processing.

In HVE CMTs, software media processing is done using dedicated media processors, such as TriMediaTM Technologies Inc.'s family of very long instruction word (VLIW) processors; see [23]. Compared to dedicated hardware solutions, these media processors are expensive, both in cost and power consumption.

,

335

Therefore, cost-effectiveness is a major issue in HVE CMTs. Cost-effectiveness requires a high average resource utilisation.

Current HVE CTs provide robust behaviour, and users expect the same robustness when media processing is performed in software and these terminals become more open. For the time being, users do not have similar expectations of multimedia applications on desktops and Internet appliances (and it is also not uncommon that these applications exhibit non-robust behaviour). Note that the notion of resource budgets (or reservations [17]) is a proven concept to provide robustness between applications.

High-quality video has a field -rate of 50 - 120 Hz, no tolerance for jitter, and low tolerance for frame skips, properties that are characteristic of the hard real-time domain. In contrast, mainstream multimedia applications are characterised by low frame rates (with a maximum of 30 Hz) and high jitter tolerance, and in addition accept frequent frame skips, properties that are characteristic of the soft real-time domain. It is conceivable, however, that future users will expect guaranteed timing behaviour from multimedia applications on desktops and Internet appliances as well (see, for example, [21]).

SCALABLE VIDEO PROCESSING

A CMT may accept input from different types of input sources, such as satellite, cable, storage device, Internet and Ethernet. The video input can be digital or analog. A CMT may have a number of video outputs: a display, a storage device (such as VR, DVD+RW, or a hard disk),.and an IEEE 1394 or Internet link. The outputs on a display may be sub-divided into two (dynamically changing) groups based on user-focus [3]. User focus induces a relative importance on outputs.

Between these inputs and outputs, a number of processing paths may exist, containing joins and forks in complex situations. Each processing path typically consists of a number of functional processing parts, termed jobs, e.g. channel decoding, picture enhancement, and rendering (for a display) or encoding (for a link). Jobs inherit the relative importance of the output with the highest relative importance to which they contribute. Whereas the functional description of a job is general (e.g. enhancement), there may be a number of specific algorithms (processing variants) within a single job.

The remainder of this section presents aspects of scalable media processing. The first subsection presents the basic structure of a scalable algorithm (SA) for media processing. Scalable video algorithms (SVAs) are a special class of SAS, as exemplified above by their specific scalability parameters. The second subsection considers jobs in more detail.

336

Scalable algorithms An SA basically consists of an algorithm for media processing and a quality

control block [9 ] ; see Figure 1. The algorithm can be split in a number of specific functions, some of which are scalable. The quality of the output depends on the

FUNCTION 3

, i

I; \- I-- 1

Figure 1 : Detailed diagram of a scalable algorithm.

From these combinations, only a few provide acceptable quality levels for the SA (see Figure 2). The optimal quality-resource combinations are connected by the curve with maximum quality at lowest resources. The quality control block contains this information and the appropriate settings for the functions.

- Resources

Figure 2: Best choices of quality-resource combinations for functions or the entire scalable algorithm.

Jobs A unique combination of algorithms within a job is termed a job mode. The job

mode is selected dynamically, and a change of job mode, e.g. due to a channel change or an exchange of the contents of the main window and a Pip window, is termed a job-mode change (JMC). A JMC may lead to a change in the specific functionality of the job, and the number and order of its algorithms. For each job mode, a number of operational sets are defined. Each operational set determines specific processing for each algorithm depending on characteristics like window- size (determining the applicable embedded resizing techniques) and user-focus. An

337

operational set is selected dynamically, and a change of operational set is termed operational-set change (OSC).

Just like a number of functions constitutes an SA (see Figure I ) , a unique combination of algorithms constitutes a job mode. Similar to SAS, jobs can be scaled, giving rise to a set of discrete quality levels associated with each job mode. Jobs provide (estimated) resource requirements for each quality level. In order to allow system optimization by QoS RM, the value of the result of a job is expressed, in semantically neutral terms, as a job utility. Each operational set of a job mode consists of a number of quality levels, and an associated quality mapping, which determines resource requirements and a job utility for each quality level in the operational set.

QOS FOR VIDEO PROCESSING

Our approach to QoS for CMTs is based on co-operation between application domain experts and system software specialists. In this approach, a QoS Resource Manager (QoS RM) deals with system aspects that are semantically neutral. Application domain specific aspects are dealt with by the media processing applications.

In [4] and [6], the dynamic behaviour of HVE CMTs is described from two complementary perspectives. The former paper presents the system perspective, based on notions of steady state and state change, and describes the implications for the system in general and QoS RM in particular. The implications for the media processing applications are described in the latter paper. This section joins both perspectives. In a first subsection, time scales of dynamic load are presented. In a second subsection, a multi-level adaptive control structure is described, addressing the dynamic behaviour at these time scales. Seamless switching, one of the main issues for JMCs during video processing, is the topic of the final subsection.

Time scales of dynamic load In the high-quality video domain, the load of a system varies dynamically on

multiple time scales. User initiated changes, such as the exchange of the contents of a main window and a PIP window and switching to another channel, take place at a time-scale of minutes (typically with bursts of occurrences). Similarly, changes initiated by the service provider, such as the interruption of a movie by a commercial, take place at a time scale of minutes. Data dependent changes in the average load of applications take place at a time scale of seconds, e.g. scene changes in a movie. Finally, many media processing functions, such as MPEG encoding and decoding, and Natural Motion [14], have a load that show large data dependent variations over time. These data dependent load variations take place at a time scale of tens of milliseconds. In summary, there are variations around a quasi-fixed average load, and variations that involve a change in the average load.

338

Multi-level adaptive control In order to address dynamic behaviour at different time scales, a multi-level

adaptive control structure [ 161 has been conceived, corresponding with different time-horizons [24]. Multi-level adaptive control is realised by means of a co- operative approach between media applications and QoS RM. This control is founded on a layer providing the functionality of a resource kernel (similar to [2 1 I), guaranteeing and enforcing resource budgets and supporting an admission test. For video processing, resource budgets are associated with jobs.

Robustness A resource kernel provides robustness between applications, resolving the temporal interference between applications, which is in particular a major threat for open systems. The issue of cost-effectiveness of HVE CMTs gives rise to an additional robustness problem within applications. Cost-effective media processing requires a high average resource utilisation. This requirement is in conflict with the hard real-time requirements of high-quality video that are traditionally met by a worst-case resource allocation. Since the traditional real- time approach is not affordable for CMTs, we are forced to opt for an average-case resource allocation. Given the dynamic load, applications will therefore be faced with occasional (or transient) and structural overloads. The resulting robustness problems are to be resolved by the applications themselves. Stated in other words, applications have to get by with their budget. Co-operative solutions to robustness problems caused by overloads are addressed in detail in [16] and [3].

Control within media applications The lowest layer of adaptive control, on a time scale of tens of milliseconds, therefore reside within the media applications. An example of a video application providing adaptive control is the MPEG-2 decoder described in [ l 11. By using an MPEG-2 decoding complexity estimation model, the required computational load can be predicted, and the computation is subsequently scaled such that it will not exceed its resource budget. A similar example may be found in [lo], describing adaptive control for 3D decoding and rendering.

Control within QoS RM Subsequent layers of adaptive control reside in QoS RM. QoS RM is partitioned in two main components, a resource manager (RM) and a quality manager (QM). RM works on a time-scale up to hundreds of milliseconds, and provides the next layer of control. RM monitors the resource usage of applications, and based on these measurements adapts the budgets to their optimal values, and informs QM about these adaptations. Occasionally, RM is not able to accommodate the resource needs of the applications, and requests assistance from QM.

The highest layer of adaptive control resides in QM, which works on time-scales longer than hundreds of milliseconds. The challenge of QM is to select the quality levels at which the jobs are executed in such a way, that the overall system utility is maximised, and the resource requirements meet the resource availability. The overall system utility is determined by the job utilities of the running jobs, the dependencies (resulting from the processing paths) of the running jobs, and the

339

relative importance levels that are associated with those jobs. Although this model has much in common with the one in [12], the incorporation of dependencies between jobs is novel. Next to performing the global (centralised) optimisation of the system utility, the QM maintains the quality mappings from the running jobs (based on the dynamic resource needs provided by RM). Changes in the number of jobs, relative importance of the jobs, quality mappings of the jobs (due to JMCs and OSCs), and requests for assistance from RM require re-optimisations. Because rapidly changing quality levels are perceived as non-quality, quality levels must be adjusted sparingly. Note that this aspect is not covered in [12] either.

Seamless switching A JMC may lead to a change in the number and order of the algorithms of the

job. Main issues of a JMC are that switching must be seamless, and that intermixing of old and new-mode data upon reconfigurations of the connections of algorithms must be prevented [6]. The media applications are responsible for providing seamless switching and preventing data intermixing. QoS RM must ensure that the applications get the necessary resources to do so, as illustrated by two examples below.

A switch to another channel may involve JMCs, requiring a re-optimisation by QoS RM, giving rise to new quality levels of running jobs. In such a case, quality level reductions of jobs are performed first, followed by resource budget reductions of those jobs. As a next step, the resource budgets of jobs that will receive a higher quality level are increased, followed by the increase of the quality level. The jobs are responsible for providing smooth transitions from the old to the new quality level.

When an additional application is requested, QoS RM first performs an admission test. When the test succeeds, it determines the optimal quality levels at which the new set of jobs (i.e. both old and new jobs) will run. As a next step, the quality levels of the old jobs are reduced, followed by a reduction of their resource budgets. When the necessary resources for the new jobs are freed, the new jobs are started.

User induced dynamic changes, such as switching to another channel, will typically give the system sufficient time to offer seamless switching. One of the challenging issues is how to handle dynamic load changes that require a fast reaction time, such as structural overloads due to changes in the incoming media data. For stability reasons, the system should be sufficiently certain that i t actually concerns a structural rather than a transient overload. Such a detection time may be long compared to the reaction time needed by the media processing applications to prevent output quality degradation. A specific solution in the context of applications with and without user focus that addresses this challenge has been described in [3].

340

CONCLUDING REMARKS

In this paper, we presented our co-operative approach for multimedia QoS in CMTs. A first prototype of an STB showing the feasibility of our approach for high-quality digital audio and video has recently been demonstrated. The validation of our approach in combination with other media processing, such as 3D graphics [lo], on a platform also supporting DVB-MHP will be performed within the context of the ITEAlEuropa project.

ACKNOWLEDGEMENTS

The work described in this document is performed by a team of people from Philips Research Laboratories Eindhoven (PRLE) and Philips Research Briarcliff (PRB) in close co-operation with the real-time systems group of the Technical University of Madrid (ditAJPM). This work is partially funded by ITEAEuropa [7]. We would like to thank our partners (IMEC, KU Leuven and Thomson Multimedia) in the advanced content activity of ITEAEuropa for the fruitful co- operation, and in particular Gauthier Lafruit and Wolfgang van Raemdonck (IMEC) for stimulating discussions and inspiration. Finally, we thank Dietwig J.C. Lowet and Evert-Jan D. Pol for their feedback on a previous version of this document.

REFERENCES

[ 11 E. Aarts, R. Harwig, and M. Schuurmans, Ambient Intelligence, To appear in: After Cyberspace: When Computing Becomes our World (Ed. P.J. Denning).

[2] C. Aurrecoecha, A. Campbell and L. Hauw, A Survey of QoS Architectures, ACMBpringer-Verlag Multimedia Systems Journal, 6(3): 138 - 151, May 1998.

[3] R.J. Bril and E.F.M. Steffens, User focus in consumer terminals and conditionally guaranteed budgets, In: Proc. 9th International Workshop on Quality of Service (IWQoS), Lecture Notes in Computer Science (LNCS) 2092 (Ed.: Lars Wolf, David Hutchison and Ralf Steinmetz), Springer- Verlag, pp. 107 - 120, June 2001.

[4] R.J. Bril, E.F.M. Steffens, G.C. van Loo, M. Gabrani, and C. Hentschel, Dynamic behavior of consumer multimedia terminals: System aspects, To appear in: Proc. IEEE International Conference on Multimedia and Expo (ICME), Tokyo, Japan, August 22-25,2001.

[5] J . Brunheroto, R. Chewrnock, P. Dettori, X. Dong, J. Paraszczak, F. Schaffa, and D. Seidman, Issues in Data Embedding and Synchronization for Digital Television, In: Proc. IEEE International Conference on Multimedia and Expo (ICME), Vol. 3, pp. 1233 - 1236, July 2000.

341

161 M. Gabrani, C. Hentschel, L. Steffens, and R.J. Bril, Dynamic behavior of consumer multimedia terminals: Video processing aspects, To appear in: Proc. IEEE International Conference on Multimedia and Expo (ICME), Tokyo, Japan, August 22-25,2001.

171 J.H.A. Gelissen, The ITEA project EUROPA. a software platform for digital CE appliances, To appear in: Proc. IEEE International Conference on Multimedia and Expo (ICME), Tokyo, Japan, August 22-25,2001.

[8] C. Gran and A. Scheller, From proven ofice technologies to the intelligent multimedia home, In: Proc. IEEE International Conference on Multimedia and EXPO (ICME), Vol. 3, pp. 1225 - 1228, July 2000.

[9] C. Hentschel, R. Braspenning, and M. Gabrani, Scalable algorithms for Media Processing, To appear in: Proc. IEEE International Conference on Image Processing (ICIP), Thessaloniki, Greece, October 7-10,2001.

[ 101 G. Lafruit, L. Nachtergale, K. Denolf, and J. Bormans, 3 0 Computational Graceful Degradation, In: Proc. IEEE International Symposium on Circuits and Systems (ISCAS), Vol. 3, pp. 547 - 550, May 2000.

[I 1]T. Lan, Y. Chen, and Z. Zhong, MPEG2 decoding complexity prediction and dynamic complexity control in a TriMedia processor, To appear in: Proc. IEEE Workshop on Multimedia Signal Processing (MMSP), Cannes, France, October 3-5, 2001.

[12]C. Lee, J. Lehoczky, R. Rajkumar and D. Siewiorek, A scalable solution to the multi-resource QoS problem, In: Proc. 20Lh IEEE Real-Time Systems Symposium (RTSS), pp. 315 - 326, December 1999.

[ 131 B. Li and K. Nahrstedt, Dynamic Reconfiguration for Complex Multimedia Applications, In: Proc. International Conference on Multimedia Computing and Systems (ICMCS), Vol. 1, pp. 165 - 170, June 1999.

[14]P. Lippens, B. De Loore, G. de Haan, P. Eeckhout, H. Huijgen, A. f ining, B. McSweeney, M. Verstraelen, B. Pham, and J. Kettenis, A video signal processor for motion-compensated field-rate upconversion in consumer television, IEEE Journal of Solid-state Circuits, 31(11): 1762 - 1769, November 1996.

[15]F. Lonczewski, and R. Jaeger, An extensible Set-Top-Box Architecture for interactive and broadcast Services offering sophisticated User Guidance, In: Proc. IEEE International Conference on Multimedia and Expo (ICME), Vol. 3, pp. 1403 - 1406, July 2000.

[16]S. van Loo, L. Steffens, R. Derwig, Quality of Service Resource Management in Consumer Terminals, Philips Research Laboratories Eindhoven (PRLE), Doc. id. NL-MS 21166, available from httu:Nwww.rcsearch.~hilius.com/manuscrint/index.html, May 2001.

342

[17] C.W. Mercer, S . Savage, H. Tokuda, Processor Capability Reserves: Operuling Sysrem Support for Multimedia Applications, In: Proc. International Conference on Multimedia CqJmputing and Systems (ICMCS), pp. 90 - 99, May 1994.

[18] J.R. Morros, and F. MaryuCs, A proposal jor dependent optimization in scalable region-based coding systems, In: Proc. IEEE International Conference on Image Processing (ICIP), Vol. 4, pp. 295 - 299, October 1999.

[ I 9.1 K. Nahrstedt, H. Chu. S . Narayan, QoS-aware Resource Mcinageinentfor Disrvihuted Multimedia Applications, Journal on High-speed Networking, Special Issue on Multimedia Networking, IOS Press, Vol. 8. No.3-4, pp. 227 - 255. 1998.

[20] S. Peng, Comple.xi~ scalable video decoding viu IDCT datu pruning, Digest of Technical Papers IEEE International Conference on Consumer Electronics (ICCE), pp. 74 - 75, June 2001.

121 1 R. Rajkumar, K. Juvva, A. Molano, and S . Oikawa, Resource kernels: A resource-centric approach to real-time and multimedia system. In: Proc. SPIE/ACM Conference on Multimedia Computing and Networking. January 1998.

1221 B. Sabata. S. Chatterjee, and J . Sydir, Dynamic Aduptation of Video for Transmission under Resource Constraints, In: Proc. IEEE International Conference on Image Processing (ICIP), Vol. 3, pp. 22 - 26, 1998.

[23] F. Sijstermans and G. Slavenburg, Providing the processing power @ r consicnier multimedia, In: Digest of Technical Papers IEEE International Conference on Consumer Electronics (ICCE), pp. 156 - 157. 1997.

[24] Z. Uykan. “Hierarchical control and multimedia,” In: Mulrimedicc applications in industrial automation - Collected papers of the Spring 1997 posrgraduate seminar, (Eds. H. Hyotyniemi and H. Koiva), Helsinki University of Technology, Report 106, pp. 91 - 114, ISBN 951-22-3835, June 1997.

[25] P. Vuorimaa, A digital television architecture, In: Proc. IEEE International Conference on Multimedia and Expo (TCME), Vol. 3, pp. 141 1 - 1414, July 2000.

[26]H. Yanagihara, M. Sugano, A. Yoneyania, and Y . Nakajima, Scalable video decoder and its application IO multi-channel miilticast system. IEEE Transactions on Consumer Electronics (TCE), 46(3): 866 - 871, August 2000.

[271Z. Zhong and Y . Chen, Scaling in MPEG-2 Decoding Loop with Mixed Processing, Digest of Technical Papers IEEE International Conference on Consumer Electronics (ICCE), pp. 76 - 77, June 2001.

343