Traffic Shaping, Bandwidth Allocation, and
`Quality Assessment for
`M PEG Video Distribution over
`Broadband Networks
`
`Steven Gringeri, Khaled Shuaib, Roman Egorov, Arianne lewis,
`Bhumip Khasnabish, and Bert Basch
`GTE laboratories Incorporated
`
`Abstract
`This article provides an overview of residential video delivery systems and presents
`the applications, benefits, and challenges of using VBR MPEG video encoding in
`broadband video distribution networks. The network resources required to transmit
`stored variable-rate MPEG can be reduced by properly analyzing and smoothing
`the video stream before transmission. A scheduling technique is presented which
`selects a traffic contract for a pre-encoded MPEG video stream with the criteria of
`minimizing network resources and maintaining video quality. Several effective
`bandwidth metrics are discussed and used to model the potential savings in net-
`work resources for the shaped streams.
`
`elecommunications providers have started to offer
`broadband services such as broadcast video, high-speed
`Internet access, and videoconferencing. Broadband net-
`work architectures such as fiber to the curb (FTTC)
`coupled with digital subscriber line (xDSL) modem technolo-
`gies [l, 21 are extending the fiber plant closer to the end user
`and facilitating the migration of narrowband networks to
`broadband networks. These architectures use asynchronous
`transfer mode (ATM) technology as the underlying transport
`protocol. ATM uses small fixed-size (53 bytes) cells to allow
`multiplexing of various services such as voice, video, and data
`with guaranteed cell rate, cell loss, and cell delay variation
`parameters. These capabilities make ATM well suited for
`real-time applications such as multimedia applications.
`The Moving Picture Experts Group (MPEG) standards [3,
`41 have been widely adopted for providing digital video ser-
`vices. The MPEG standards do not prescribe the encoding
`process; instead, they specify the data input format for the
`decoder (the syntax) as well as detailed specifications for
`interpreting this data (the decoding semantics). An encoder
`must follow a set of steps known as the encoding process to
`compress video data. This process is not standardized, and
`may vary from application to application depending on a par-
`ticular application’s requirements and complexity limitations.
`This allows the encoding process to be optimized for an appli-
`
`cation’s bandwidth and quality requirements. Today, commer-
`cially available encoders support a wide range of transmission
`and storage applications using both constant bit rate (CBR)
`and variable bit rate (VBR) encoding.
`Transmission of real-time video streams is resource-inten-
`sive even when the video is compressed using sophisticated
`algorithms like MPEG. As digital video systems are deployed,
`efficient utilization of storage and network resources will be
`needed to reduce costs and increase revenues. Variable-rate
`MPEG encoding has the potential to save bandwidth and
`increase channel capacity while reducing system cost and serv-
`er storage requirements [5-71. Compressed digital video such
`as MPEG is inherently variable-rate since its complexity and
`motion content affect the encoding bit rate required to main-
`tain picture quality. ATM networks have the flexibility to sup-
`port both constant- and variable-rate services, and can provide
`statistical gains when using variable-rate MPEG. However,
`uncontrolled burstiness will lead to inefficient use of network
`resources by occasionally requiring excessively high process-
`ing, storage (buffering), and transmission capacity from the
`network. These requirements motivated the early imple-
`menters and providers of digital video services to use a CBR
`channel for delivering real-time video to customers. Digital
`video encoder manufacturers, accordingly, implemented addi-
`tional rate-control and buffering in the encoder to generate a
`
`94
`
`0890-8044/98/$10.00 0 1998 IEEE
`
`IEEE Network NovemberDecember 1998
`
`1
`
`DISH 1028
`
`

`

`v
`
`<
`
`CBR stream for transmission and distribution
`applications. The fullness of the rate-control
`buffer dynamically controls the quantization res-
`olution so that the number of bits generated per
`picture satisfies the bit rate constraints of the
`video stream. These rate adaptation methods not
`only lead to variable video quality, but also may
`result in poor utilization of bandwidth since the
`rate must be selected to accommodate the most
`complex scenes, and the encoder may need to
`use stuffing bits to maintain the CBR.
`VBR encoding can achieve improved coding
`efficiency by better matching the encoding rate
`to the video complexity. Variable-rate encoding
`is currently used in storage applications such as
`digital versatile discs (DVDs) to achieve signifi-
`cant storage savings. Using VBR MPEG, it
`should be possible to achieve substantial savings
`in networking resources while maintaining video
`quality, if the burstiness of the video can be con-
`trolled using some prespecified constraints. In
`addition, it is possible to offer a more constant
`video quality than can be provided using con-
`stant-rate encoding. Savings in storage resources
`such as server capacity allow more movies or
`video seauences t o be stored using the same
`capacity. Similarly, savings in networking resources imply that
`the same amount of switching, buffering, and transmission
`capacity can be used to deliver more video content to the cus-
`tomer while maintaining the application-level quality of ser-
`vice (QoS) requirements. Either source coding or output
`shaping or a combination of the two can be used to adapt
`MPEG video streams for transmission over VBR ATM chan-
`nels. To achieve such savings, it is imperative that we control
`the burstiness and bitrate variability of a single stream or a set
`of multiplexed streams in order to make the stream(s) adhere
`to a traffic contract while maintaining the desired video quali-
`ty. The bandwidth saving for transmission applications may be
`smaller than for storage applications due to additional restric-
`tions the network places on the video’s burstiness.
`This article addresses issues in encoding and distributing
`variable-rate MPEG video over broadband networks. The pri-
`mary objective is to realize bandwidth savings while maintaining
`video quality. The issues related to shaping MPEG video to
`control its burstiness and produce streams that are more suit-
`able for transmission over switched ATM networks are dis-
`cussed. An overview of residential video delivery systems is also
`included to provide the architectural information about these
`systems relevant to implementing variable-rate MPEG. The
`relative quality of the various CBR and VBR video streams is
`compared using both subjective viewing and quantitative mea-
`surements. The goal of the comparison is to ensure that video
`quality is being maintained for variable-rate encoding. An
`analysis was conducted to determine the required ATM net-
`work resources and traffic contracts for various video cate-
`gories (action movie, sports, talking head, etc.) when MPEG
`video is transported over an ATM network using real-time
`VBR service. Potential savings are analyzed for single streams
`and groups of streams to determine whether variable rate
`MPEG streams have traffic contracts consistent with the capa-
`bilities of today’s ATM switches. The bandwidth savings are
`quantified using several effective bandwidth metrics [8-111.
`Our results to date are based primarily on shaping existing
`streams before they enter the network to reduce their bursti-
`ness and thus their utilization of network resources. This tech-
`nique is well suited to MPEG streams that have already been
`encoded, although encoding variable-rate streams specifically
`
`ONU: Optical network unit
`Figure 1 . The architecture of hybridfiber coax networks.
`
`for transmission applications could further reduce the required
`network resources. These rate control techniques for encoding
`“ATM-friendly” streams have been discussed extensively in
`the literature [5, 6, 12-16] but to date are not implemented in
`real-time transmission systems. Our results indicate that using
`variable-rate MPEG in transmission applications such as resi-
`dential video distribution systems has the potential to decrease
`bandwidth requirements and improve channel capacity.
`
`An Overview of Residential Video
`Distribution Systems
`Traditional video service providers deliver analog video to res-
`idential customers using cable TV networks or direct broad-
`cast satellitc (DBS) technology, mainly for entertainment
`purposes. Digital representation of video opens up the possi-
`bility of computerized processing of multimedia information,
`allowing consumers to archive, index, and retrieve programs
`in a content-based manner. Digitally encoded information can
`be made more resilient to degradation during transmission,
`distribution, and storage, giving the consumer a better picture
`and sound quality. Digital transmission of audio-visual infor-
`mation also enables multimedia communication over the same
`network that supports voice and data communication using
`xDSL technologies [ l , 2, 171. In the next two paragraphs,
`related issues from the point of view of network and transport
`architectures are discussed. Details of the various architec-
`tures can be found in [2]. We then briefly describe the domi-
`nant
`transmission
`protocols
`followed by
`some
`recommendations on the usage of the various architectures
`and protocols for video delivery.
`Most of the traditional residential video distribution sys-
`tems usually deliver one-way analog video using either full
`coaxial cable (CATV) or hybrid fiber-coax (HFC) networks
`[l, 21. A typical HFC network is shown in Fig. 1. These net-
`works offer only limited capability for real-time interaction
`with the service provider’s facilities for content selection. For
`wireless delivery of video, cost-effective small dish antennae
`for receiving DBS signals have recently appeared on the mar-
`ket. DBS networks use geostationary satellites which orbit the
`
`IEEE Network NovemberDecember 1998
`
`95
`
`2
`
`

`

`Direct broadcast satellite
`( O W
`
`Ku-band 11-1 8
`
`Local multipoint distribu-
`tion system (LMDS)
`
`27.5-29.5
`
`2.1 5-2.70
`
`Multichannel multipoint
`distribution system
`(MMDS)
`W Table 1 . video distribution using wireless transmission technologies.
`
`500 downstream (1 1.7-
`12.2 GHz) 500 upstream
`(14.0-14.5 GHZ)
`850 downstream (27.5-
`28.35 GHz) 150 upstream
`(29.2-29.35 GHz)
`198 downstream
`4 upstream
`
`150 digital channels
`
`42 analog channels
`
`1 0 0 0 ~ of km radius,
`using -50 cm. home dish
`antenna
`-5 km radius, using -30
`cm. home dish antenna
`
`I
`
`33 analog channels
`(1 20 or more digital
`Channels)
`
`-50 km radius, using a
`-60 cm home dish
`antenna
`
`Earth 23,500 mi above the surface. Although they were origi-
`nally intended for delivery of analoddigital video to customer
`premises, it is possible to utilize these systems for interactive
`delivery of digital video and Internet information. Figure 2
`presents a high-level description of such architecture. To sup-
`port interactivity, this architecture uses the public switched
`telephone network (PSTN) as the return path from the cus-
`tomer premises. All of the above systems are based on shared
`media architectures. These architectures exhibit a number of
`problems, including maintaining secure and reliable operation
`as well as providing customized services for individual users.
`In addition, these architectures cannot easily support the nec-
`essary infrastructure for delivering integrated telephone, data,
`and video services.
`Traditional CATV networks are designed to offer low-cost
`unidirectional bandwidth from 550 MHz to 1 GHz to provide
`analog TV services over coaxial cable. Tree-and-branch-type
`architectures with cascaded unidirectional amplifiers to
`improve signal strength and reliability are commonly used in
`such networks. Existing CATV networks are being upgraded
`to an HFC architecture, as shown in Fig. 1. This architecture
`uses an all-optical backbone network to interconnect the
`head-ends. From the head-end, fiber links run to regional dis-
`tribution centers which provide opto-electric conversion.
`Coaxial cable is used in the local feeder loop from the distri-
`bution center to the home. Traditional analog broadcast ser-
`vices occupy the spectrum between 50 MHz and 550 MHz for
`about 80 channels, where each channel uses 6 MHz of band-
`width. The 6 MHz analog channel can carry digital informa-
`
`Figure 2. Video delivery using wireless distribution hubs and DBS.
`
`tion using different modulation techniques. For example, with
`256 quadrature amplitude modulation (256-QAM) or 16 vesti-
`gial sideband (16-VSB) modulation, a 6 MHz channel could
`deliver approximately 38 Mb/s data (the remaining bits are
`used for error correction and control) to the home [I, 21. If a
`CBR of 6 Mb/s is used for digital video, a maximum of six
`video streams can be accommodated in one analog channel.
`Recent research [5, 91 in the area of bandwidth allocation and
`transmission of video shows how VBR encoding of video can
`increase the channel multiplexing capability while maintaining
`the video quality desired by the application.
`Another option for video distribution networks is to use
`wireless transmission to the home. Two possible techniques
`for offering high-quality digital video distribution and fast
`Internet access using low-power, high-frequency radio signals
`over short to medium distance are the local multipoint distri-
`bution system (LMDS) and multichannel MDS (MMDS), as
`described in Table 1 and shown in Fig. 2. Investments in these
`systems are incremental in the sense that they can grow as
`new customers are added to the system.
`Digital multimedia services require significantly more band-
`width than traditional voice services. One way to achieve this is
`to use xDSL modem technologies [l, 21 which enable high-bit-
`rate data transmission using the existing copper plant (sub-
`scriber loop). It is also possible to use this technique for digital
`video delivery services. Asymmetric DSL (ADSL) can support
`a bit rate of 6.312 Mb/s (the maximum over 3000 ft is about 9
`Mb/s [18]) from the central office (CO) to the user, and
`around 640 kb/s from the user to the CO over a distance of less
`than 12,000 ft. Very-high-speed DSL (VDSL)
`can deliver data at OC-1 rate (51.84 Mb/s) to the
`home using twisted pair telephone lines over a
`distance of less than 1000 ft, and the data rate
`could be 25.92 Mb/s over a distance of 3000 ft.
`Table 2 shows the maximum ADSL, rate-adap-
`tive DSL (RDSL), and VDSL data rates under
`ideal line conditions. Data rates depend on a
`number of factors, including the length of the
`copper line, its wire gauge, the presence of
`bridged taps, and cross-coupled interference [ 171.
`Broadband networking architectures such as
`fiber to the node (FTTN) or FTTC extend fiber
`from the CO all the way to the home o r to
`some intermediate point such as the node or
`curb. FTTC or FTTN networks coupled with
`xDSL for the “last mile” to the customer
`premises can extend the capabilities of broad-
`band networks (fiber plants) closer to the end
`user without requiring the initial investments of
`fiber to the home (FTTH), and thereby can
`facilitate the migration of narrowband networks
`to broadband networks [ l , 21. For delivering
`
`96
`
`IEEE Nchvork November/Deccmbcr 1998
`
`3
`
`

`

`W Table 2. Theoretical data rates for ADSL, RDSL, and D S L .
`
`switched digital video, local exchange
`carriers are currently investigating
`FTTC or FTTN architectures (Fig. 3),
`which can be extended to FTTH (Fig.
`4) networks in the future. FTTH is
`the ultimate wireline architecture for
`broadband services to the home; the
`raw bandwidth could be OC-3 (155
`Mbls) or higher.
`As video distribution systems
`change from analog to digital, or new
`digital systems are deployed, efficient
`utilization of storage and network
`resources will be needed to improve
`the capacity and cost effectiveness of
`these systems. One way to achieve
`this would be for video broadcast sys-
`tems to support variable-rate MPEG
`encoding and transmission in addition
`to the traditional CBR. Issues related
`to the encoding and transmission of
`MPEG video at constant and variable
`bit rates are discussed next.
`
`I Figure 3. Video delivery overATMlVDSL using FTTNIFTTC networks.
`
`Constant and Variable Bit Rate MPEG Video
`Encoding
`Through a combination of spatial and temporal compression
`techniques, analog video can be significantly compressed while
`preserving image quality using the MPEG [3, 4, 191 compres-
`sion standards. The spatial compression techniques are the
`same as those used by JPEG and include discrete cosine
`
`ONU: Optical network unit
`Figure 4. The architecture of FTTH networks.
`
`transform (DCT), quantization, and entropy coding. The tem-
`poral compression techniques rely on block-based motion
`compensation to reduce the temporal redundancy. In MPEG,
`three main picture types are defined: intra-coded (I-picture),
`predictively coded (P-picture), and bidirectionally coded (B-
`picture). I-pictures are self-contained and are coded without
`reference to other pictures. They are used as reference frames
`during the encoding of other picture types and are coded with
`only moderate compression. P-pictures are coded more effi-
`ciently using motion-compensated prediction from a past I- or
`P-picture. B-pictures provide the highest degree of compres-
`sion, and require both past and future reference pictures for
`motion compensation. Several frames are grouped together in
`a pattern to form a group of pictures (GOP).
`Digital video is bursty in nature [20], and the burstiness
`depends on the frequency of changes in the background and
`the movement of objects in the foreground. Without rate con-
`trol the output bitstream of a video encoder will be VBR
`since it depends on the complexity of the scene, degree of
`motion, and frequency of scene changes. Digital video encoder
`manufacturers accordingly implemented additional rate con-
`trol in the encoder to generate a CBR stream for transmission
`and distribution applications. The rate control algorithm of
`the MPEG-2 encoder dictates the quantization resolution,
`which directly controls the number of bits used to encode a
`picture. By using an output buffer to smooth the bit rate vari-
`ation, the encoder can maintain a CBR at its output some-
`what independent of the number of bits needed to code
`individual pictures. The rate control algorithm uses the full-
`ness of the buffer as a limit on how large particular frames
`can be without causing overflow (i.e., it uses source rate con-
`
`IEEE Network NovemberiDecember 1998
`
`97
`
`4
`
`

`

`trol to limit the number of bits entering the output buffer).
`This is achicvcd by using coarse quantization to gcnerate a
`smaller number of bits pcr picture when the buffer is almost
`full, or selecting a fincr quantization when the buffer is almost
`empty. These rate adaptation methods not only lead to vari-
`able video quality, but occasionally may cause poor utilization
`of bandwidth because the encoder may necd to use stuffing
`bits to maintain the CBR.
`VBR encoding of MPEG video can be achieved using sin-
`gle-pass (real-time) or multipass techniques. In both eases the
`available options for producing a VBR video stream include
`source rate control in the MPEG-2 encoder, shaping the out-
`put bitstream of the encoder, and rate control which uses
`feedback from the network. Source rate control, output shap-
`ing, or a combination of both will be required to limit the
`burstiness of a video stream. Network-feedback-based rate
`control will be needed if the network has to provide a guaran-
`teed QoS that requires some knowledge of the traffic sources’
`behavior. The video quality depends not only on the rate con-
`trol and smoothing method selected, but also on the video
`characteristics and category (sports, news, movie) of video
`being encoded. The challenge is to find the optimal rate shap-
`ing and smoothing strategy that maintains acceptable video
`quality without putting an excessive burden on the network.
`MPEG-2 video encoding offers a rich array of possible
`methods for source rate control. A fundamental parameter is
`the quantizer scale since it controls the instantaneous bit rate
`used at the macroblock level. The quantizer scale is uscd to
`control the resolution of the DCT coefficients. The quantizer
`scale can be set globally at the picture layer, and the desired
`rate can be achieved by varying the scale for different types of
`pictures. The quantizer scale can also be adjusted at the slice
`or macroblock layer to provide finer granularity in rate con-
`trol. The number and length of the D C T coefficients in a
`macroblock will depend on the complexity of the video coded
`and the value of the quantization scale selected. Low-frequen-
`cy coefficients represent the background of a picture, while
`high-frcqucncy coefficients represent fine picture detail. Dur-
`ing high-motion sequences the changes between frames can
`be large, and therefore more coefficients are required to code
`the difference. The picture layer controls the display of video
`frames, and it is possible to control the bit rate variability of
`video by setting different thresholds for different frame types
`in the encoder’s rate control buffer.
`Other methods exist for varying the bit rate during the
`encoding process. These methods do not provide the same
`instantaneous control available by changing the quantizer
`scale, but instead provide the ability to make large adjust-
`ments in the ratc with the potential of drastically affecting
`video quality. The most important technique involves adjust-
`ing the frame rate to reduce the bit rate requirements during
`hard-to-encode scenes. This is accomplished by using the
`repeat field flags provided in the MPEG syntax [3]. For exam-
`ple, if a scene change has been detected and a large number
`of bits is required to code that next frame, the previous frame
`can be repeated to lower the bit rate and make room for the
`large frame in the buffer. Although this technique is easy to
`implement, the repeat frame can easily be detected in careful
`viewing of the video: therefore, this technique should only be
`used when less drastic techniques based on modifying quantiz-
`er scale are not sufficient. The video rate can also be adjusted
`by modifying the video resolution (number of pixels per pic-
`ture) o r the chrominance sampling resolution [19] (e.g.,
`switching from 4:2:2 video format to 4:2:0), although in prac-
`tice these techniques are seldom used since they lead to large
`variations in video quality.
`VBR MPEG-2 video can be produced using both real-time
`
`and offline coding techniques. In the case of real-time VBR
`encoding, unless the coding and rate control parameters are
`selected very carefully, the quality of video may vary significant-
`ly over a single video session. The nature of real-time coding
`limits the bit rate optimization process since information about
`video complexity is limited to the current and past frames, and
`no information exists about future frames [20-221. Offline VBR
`encoding consists of several steps which include video charac-
`terization, rate profile generation, and, finally, video coding.
`For offline encoding the video characterization and rate profile
`generation steps are performed over the entire video sequence
`before the encoding process starts, thus optimizing the bit allo-
`cation over the entire sequence. This type of multipass encod-
`ing can be implemented in real-time systems by using a
`two-stage pipelined encoding process. In this Configuration
`the first stage of the pipeline characterizes the video, while
`the second stage encodes the video based on a rate profile
`which is generated from the first stage. It is possible to look
`ahead up to several seconds with this approach. This does not
`give the flexibility of true multipass encoding but can signifi-
`cantly improve both video quality and encoding efficiency
`since the pipeline allows up to several seconds of future frame
`statistics to be available before the current frame is encoded.
`Video characterization is the video preprocessing or pre-
`viewing phase. The objective is to determine the frequency of
`scene changes, the degree of motion of the foreground and
`background objects, and the coding complexity of the scenes
`and various objects in the video sequence. This pass helps
`decide the GOP structure to be used. It also allows the mea-
`surement of various parameters that can be used to generate
`a bit rate profile for a target video quality. It may be possible
`to use the information from the profile to determine the con-
`ditions for virtually open loop or unconstrained VBR video
`coding. During the rate profile generation pass, the upper and
`lower bounds on the amount of bits allocated to various types
`of frames within the GOP are varied. In each profile attempts
`are made to allocate a large number of bits while coding com-
`plex and high-motion segments of a scene, and fewer bits to
`relatively still scenes so that the target video quality can be
`maintained. Several such profiles may be generated during
`this phase. In the video coding pass, the target traffic contract
`parameters of the channel can be used to determine the GOP
`structure and to select one of the bit rate profiles generated in
`the second pass. The quantization scale factor can now be var-
`ied to maintain the required bit rate profilc while satisfying
`the long-term bit rate for the application. This may result in
`highly granular rate control at the subframe lcvel to satisfy the
`target quality and rate requirements.
`Variable-rate encoding is commonly used in storage applications
`such as DVDs. In this application, the bit utilization is optimized over
`the entire encoding session to produce the desired peak and average
`rates for the stream. The encoder needs to maintain a target video
`quality while not exceeding the specified peak and long-term
`average bit rates. The source rate control tries to optimize
`video quality for the DVD characteristics.
`The storage requirements for a movie clip encoded using
`VBR encoding are much less than those using CBR encoding.
`When elips of similar video quality are compared, the VBR
`encoding shows significant storage savings. For example, to code
`a 15-min video clip using CBR encoding at 5 Mb/s, 562 Mbytes
`of storage are required, whereas recording the samc clip using
`VBR encoding with a peak rate of 6 Mb/s and a mcan rate of
`3 Mb/s requires only 355 Mbytes of storage. This translates
`into almost 37 percent savings in disk space for storing clips of
`very similar video quality (quality assessment of CBR and
`VBR encoded video clips is presented in the next section).
`Additional savings can be achieved, depending on the selected
`
`98
`
`IEEE Network NovemberiDecember 1998
`
`5
`
`

`

`encoding parameters (peak, mean, and minimum rates) and
`the category and complexity of the video clip.
`For transmission applications the burstiness of a video
`stream must be bounded. One way to achieve this is to have
`the quantization levels determined by a rate control algo-
`rithm that takes into account the fullness of the output
`buffer as well as parameters that monitor compliance with
`an agreed upon traffic contract. When a number of VBR
`video streams are statistically multiplexed together, the band-
`width savings may depend on specific requirements. For some
`applications such as satellite broadcasting it may be feasible to
`maintain the grouping of video streams, and therefore it is
`only necessary to keep the aggregate bit rate constant. For
`other applications, such as switched digital video networks,
`the grouping of the channels may not be uniform throughout
`the network. In this case, the rate must be controlled for each
`stream and not for a group of streams.
`There are two commonly used approaches for statistically
`multiplexing MPEG streams to maintain a near-constant-rate
`aggregate stream. The first approach involves using a statistical
`multiplexer which provides feedback to the individual encoders.
`This feedback is used to adjust the rate of a particular channel
`based on available resources, and the complexity and priority
`of the particular channel being encoded. This method can
`produce a highly stable output bit rate, although the feedback
`process adds complexity to the system. This method also adds
`delay since video statistics must be collected before a final
`video rate can be selected. This is commonly accomplished by
`profiling the video and sending this profile information about
`future video complexity to the multiplexer so that it can be
`used to select the rate for a particular channel. For real-time
`encoding this implies that a two-stage pipeline encoding pro-
`cess will be used. The second approach is to blindly multiplex
`channels using temporal smoothing rather than feedback from
`the multiplexer. This approach does not guarantee as stable
`an aggregate bit rate, but does not require the complication of
`using feedback from the multiplexer. If a sufficient number of
`streams are multiplexed the overall rate will be nearly con-
`stant, and the results will be similar to the case with feedback.
`In the event of a bandwidth explosion, it is possible to drop
`frames to reduce the rate, although this method will produce
`noticeable artifacts. The feedback approach also reduces qual-
`ity during bandwidth explosions, but since the reductions can
`be spread over all of the channels being multiplexed, the visu-
`al effect is less noticeable.
`
`Quantitative Qualitv Assessment of Constant-
`and Variable-Rate tide0 Encoding
`When comparing the transmission and storage requirements
`for various rates of MPEG encoded clips, overall video quality
`is an important concern [22]. Traditionally, video quality has
`been measured by expert viewers or viewing panels. These
`measurements are subjective, and a quantitative metric would
`provide a more objective and repeatable measurement. The
`peak signal-to-noise ratio (PSNR) has been used as a quanti-
`tative metric to measure video quality. This metric captures
`the difference between the coded and original video, but does
`not provide any insight into how apparent the differences will
`be to a viewer. Objective measurement techniques based on
`the human visual system (HVS) are necessary because some
`perceptible video artifacts may not be apparent when measur-
`ing video quality with traditionally used metrics such as the
`PSNR. Several new metrics have been developed recently,
`including the just noticeable difference (JND) metric [23]
`which was developed by Sarnoff Laboratories and is imple-
`
`Figure 5. JND test setup.
`
`mented in a new video quality tester from Tektronix. The JND
`model is a method of predicting the perceptual ratings that
`human subjects will assign to a degraded (during compression,
`transmission, or decoding) color image sequence relative to its
`nondegraded counterpart. The model compares the reference
`image sequence with the degraded sequence and produces
`several difference estimates, including a single metric of per-
`ceptual differences between the sequences. These differences
`are quantified in units of the modeled human JND. Figure 5
`is a high-level illustration of the JND system.
`In the Tektronix equipment the JND algorithm is performed
`on 120 fields (2 s) of the test video sequence. The video field
`red, green, blue (RGB) images are converted to luminance and
`chrominance components. The luminance and chrominance
`images experience several measurement stages that model dif-
`ferent aspects of the HVS. To model the HVS’s loss in light
`sensitivity after a light-to-dark scene transition, the luminance
`image is normalized with a time-varying average. Contrast mea-
`surements are performed on the images, and certain impercep-
`tible video artifacts (e.g., distortions in busy image areas) are
`masked. The masked contrast measurements are compared
`with identical outputs from the reference video sequence to
`return the JND value for the field. A JND value of one indi-
`cates that the difference between the test and reference fields
`is at the human threshold of perceptibility, or that there is a
`75 percent probability that a human observer viewing the test
`and reference fields