`
`[BEE TRANSACTIONS ON COMMUNICATTONS, VOLI COM-29. NO. 12, DECEMBER A1981
`
`'
`
`8
`
`t,
`
`12“
`
`.
`
`.
`
`Statistical Performance Analysis of an Interframe Encodch
`for BroadcasthTelevision Signals
`‘
`'
`"
`y
`TOSHIO KOGA. YUKIHIKO lIJlMA. KAZUMOTO ilNUMA, mu TATSUO lSHlGURO
`, I;
`g,
`.
`KM .
`.”
`.
`‘\ k‘
`Albumen—This paper describes an objective evaluation for coding
`performance of an interfrnmélencoder (NETEC~22H). Also described
`is the coding performance improvement bylaw adaptive_bit sharing
`multiplexer (AllS—MUX) in which transmission bit rate is dynamically
`allocated to several channels.
`'
`7
`Measurements made for actual broadcast TV programs over a time
`of 36 h show that an SNR or higher than so an unweighted is obtnined
`by thlslcodlng equipment for 99 percent of the time for broadcast TV
`programs It “retransmission hit rate of 36 Mbits/s and for” percent
`of the time at 20 Mhits/s. The residual 1 percent at 30 Mbits/s or 7
`percent at 20 Mhits/s is transmitted with a slightly lower SNR. The
`picture quality difference between the 20 and 30 Mbitfs transmission
`is about 6 dB in SNfiflontbe average.
`It is also shown that a three-channel ABS-MUX {20 Mbits/s per
`channel on the average) reduces probability of coarse quantization by
`a factor of 5-10 compared with the fixed bit rate lransmlsfiéon at 20
`Mbits/s.
`
`PMCAPL02442661
`
`,
`
`f
`
`N
`
`‘
`
`I. INTRODUCTION
`
`ARIETIES of television coding algorithms have been de-
`Vvised and developed [l ] , [2] . lntérframe coding has been
`expected to be most promising {BL A: first, it was applied to
`1 MHz video telephone for face-to-face communication [4} . in
`\due course, the appliczftion was extended to 4 MHZ video tele-
`conferencing. Interframe coding has paved the way to visual
`communication with full motion video [5] {10} .
`Until now, emphasis has been placed on achieving a high
`compression ratio for video teleconferencing application On
`the contrary. as far as broadcast TV signal transmission is con-
`cerned,
`it is more important to transmit a high quality signal
`than to achieve a high compression ratio. Some examples are
`mentioned in the literature of such digital television encoders
`for broadcast TV program transmission use at 16-30 Mbits/s
`[11 J—[13].
`The interframc encoder (NETEC~22H) described here is an
`improved version of that described in [l l I , in which an adap-
`tive interframe/intraframe prediction is used to improve coding
`performance for pictures with substantial motion [14]. Ac-
`cording to subjective 'evaluation, encoded picture quality is
`excellent for most pictures at a coding bit rate of 20-30 Mbits/s.
`However, if a very busy picture with active motion is supplied,
`sornc picture quality degradations are observed, although they
`are encountered with a very small probability in actual tele-
`
`Manuscript rnncivcd March 13, 198]; roviscd August 5. 1981. This
`paper was prcxcnted in part in the National ‘l'elcoommunications Con—
`ference,
`fielllnh', TX, llecembcr 1976, and the International Conference
`rm Communications, Boston, MA, June 1979.
`The. authors are Willi Nippon l‘lL‘JSlVll“ Company,
`l,t(l.. Kawasaki,
`mrmo
`
`the interframe coding performance
`vision programs. Thus,
`greatly depends upon picture content.
`I
`I
`in order to obtain an objective measure of the interframe
`coding performance, knowledge of the statistics of TV signals
`is necessary.
`in this paper, probability distributions of the
`amount of information encoded by NETEC-QZH [is]
`are
`measured for actual broadcast TV signals for many hours.
`Also, statistics of the encoder parameters, which‘are adap-
`tively controlled according to the buffer memory occupancy
`or equivalently the rate of source signal information, are meas-
`ured. These results are exploited in calculating SNR probabili-
`ties of encoded broadcast TV signals. Furthermore, effective
`utilization of transmission bit rate and picture, duality im— *
`provernent can be achieved by means ofthe proposgd adaptive
`bit sharing multiplexer {ABS-MUX) for multiple sirnultaneous
`channel transmission {16} , using the advantage ofinstantaneous
`differences among multiple channels which are statistically
`similar. The effect of ABS-lille inalso measured for actual
`broadcast TV signals by using physically realized hardware
`systems.
`‘
`’
`ll. CODlNG ALGORITHM
`
`The encoder/decoder block diagram is shown in Fig. l. A
`composite NTSC color TV signal
`is normally sampled at
`i0.76 MHZ and digitized into an 8 bit 'PCM signal, and then
`compressed to reduced bit rate data of 2-3 bits/picture ele-
`ment (bit/pol) on the average through digital signal processing.
`First,
`the preprocessor makes compensation for the phase in-
`version of the color subcarrier between two successive frames.
`The adaptive interframc predictive encoder removes redundancy
`from the signal by adaptive prediction and the Variable word-
`length coder compresses prediction error
`information with
`variable wordlength codes. The compressed data stream, which
`is generated irregularly, is smoothed out by the buffer memory
`and sent to the transmission line. At the decoder, the inverse
`
`processing is made to reproduce the NTSC color TV signal.
`The variable wordlength decoder expands the compressed
`data supplied from the butter memory. The expanded data
`is decoded through the adaptive interframe predictive decoder
`to yield the phase-compensated signal. The postprocessor pro-
`duces the composite NTSC color TV signal to be D/A con
`verted.
`Thus, the codes is designed so dull the cmnpositc video sig-
`is directly encoded and the waveform of the input signal
`iull
`is preserved ext-opt tor qunntlmtion cited,
`The encoder/dormer operates under two different sampling
`immunities: 10‘er Mllr. {filial X f”, f4: horizontal sync tre-
`
`()U‘M)0778(8l/lZOO-léltiélllllo "/5 «1) WM Hal-ll
`
`PMC3683168
`
`PMC Exhibit 2035
`Apple v. PMC
`IPR2016-00755
`Page 1
`
`
`
`KOGA efala PERFORMANCE ANALYSIS OF INTERFRAME ENCODER
`
`Buffer Memory Mcupdncy lEMO)
`1
`
`NTSC
`Color TV
`Signal
`AUDIO
`
`i
`
`Pm -
`processor
`
`l
`
`Adoptive
`Variable
`inter from
`Word-length
`Pmdrcrm
`Encodcr
`‘
`Coder
`
`n
`
`Post-
`processor
`
`Reproduced
`NTSC
`Color TV
`'
`Sic nal
`AUDIO
`M
`520‘
`
`\
`
`Decoder
`
`Fig.
`
`l.
`
`)1
`
`.
`i
`
`Adoptm 1
`Varian '
`Frost ctwu
`Decoder
`Deccan
`innrtrqme H Warm",
`'
`[,1
`NETE’C-ZZH encoder/decoder block diagramv BMO (‘) is
`\
`supplied to ABS-MUX.
`
`'
`
`. CLOCK
`
`DUX
`
`Bifilav
`» Memory
`
`mission
`Lint
`
`l
`l
`
`,,
`
`J
`
`Since the subcarrier phase of the NTSC color TV signal
`alternates by 1800 at a sampling point, frame by frame, sub-
`carrier phase compensation is necessary before taking frame~
`to-frame differences.
`Basically, the input composite signal is,first separated into
`luminance and chrominance componentsjand then the phase
`of the chrominance component signal isfinverted every other
`frame to produce a phasenompensate‘d signal. The phase-
`compensated signals Y», and Cm are encoded through the
`interframe coder as follows. Shown in Fig.2 is the preprocessor,
`consisting of an orthogonal transformer for T~mode and a
`comb filter for S-mode. In the orthogonal transformer (OTF),
`a pair of lines L2,, and L2,"? are used to yield luminance
`Y," and phase-compensated chrominance Cm components by
`the following second-order orthogonal transformation.
`
`PMCAPL02442662
`
`quency) in T-mode and 7.16 MHz (456 X fH) in S-mode. The
`T-mode is a normal operation mode. The S-mode is a sub-
`Nyquist frequency operation which is only applied for ex.
`treme cases where a large amount of information is generated
`and buffer fill occurs or is likely to occur. Therefore, T—mode
`operation is mainly described in What follows.
`
`A. Subcam'er Phase Compensation
`
`Y;
`
`c
`
`__r
`
`1
`
`‘2 (—1)"~(—1)”
`
`1
`
`[Lzm ]
`
`'Lmr
`
`\
`
`where n is a frame number and the operation (71)" corre-
`sponds to the phase inversion to be made every other Frame.
`At the decoder, the inverse transformation of (l) is made
`to reproduce the composite signal L2", and L2”,+ i.
`
`” L2".
`L“...
`
`~_
`
`(—1)"
`I
`14—1)"
`
`Y...
`Cm '
`
`(2)
`
`It should be noted that the phase compensation represented
`by (l ) and (2) produces no distortion, because it is a reversible
`process.
`In the comb filter for S mode, ahrominance (C) and lumiv
`nance (V) components are separated using a ll] delay circuit
`and a bandpass filter. The plume oi the chmmlnnncr‘ signal C
`
`l
`5
`.0
`3 sn-
`To Adoptiv-
`Intorkum
`Prvdictive
`Encoder
`
`7-mo-
`i_—’—¥
`r onncmr
`ll'ransfwmhcn Ms
`‘
`
`l‘
`
`;
`‘
`
`.
`From
`Ali seaming
`
`
`
`silage.
`Preprocessor for 511 bearrier phase compensation.
`
`Fig. 2,
`
`is inverted etery other frame and added to the low-pass filtered,
`Y to yield a phase compensated color TV signal. Strictly speak-
`ing, this process gives rise to slight degradation in color fidelity
`because it is irreversible.
`
`B. AdaptiverPredic-tion Interfmme Coding
`
`1} Prediction Function: A block diagram of an adaptive in-
`terframe predictive encoder is shown in Fig. 3. The coder has
`two predictors. One is an interframe predictor, P1(z), and the
`other is an intraframe predictor, P1(z).
`Using the Z-transform representation
`
`191(2) =z' F,
`
`for TandS—mode
`
`P (2) =
`2
`
`2—3,
`z‘2
`
`for T-mode
`for S-mode
`
`where 2—3 means three sample delay and z‘F one frame delay.
`A more nearly optimum prediction function could be deter-
`mined [IJ , [l7] , but the simple function P2(‘Z) above is used
`for the sake of hardware simplicity. The simplification results
`in an increase in the amount of information by only 5 percent.
`The adaptive prediction is made by exclusively choosing one
`of the two predictors.
`The choice of the prediction functions is made according to
`the quantized prediction error amplitude. lfthe prediction er-
`ror is smaller than a threshold level TH, the presont switching
`signal is continued in order to hold the same prediction. When
`the quantized prediction error amplitude exceeds TH, the pre-
`diction swiltihlng signnl is inverted in polarity by liX-(JK ate,
`
`{PMC3683169
`
`PMC Exhibit 2035
`Apple v. PMC
`IPR2016-00755
`Page 2
`
`
`
`.
`
`‘
`
`r
`
`'
`
`\ Adaptive lntari’mme
`
`,
`
`.,
`
`‘Hle/ZSE
`
`
`
`
`
`PredictionErrorEntropy
`
`orO
`
`lnmrmailonRan N
`
`S
`
`
`
`
`
`PMCAPL02442663
`
`‘ and the prediction is switched to the other. Thus, the choice
`is made on _a sample-by-sample basis. For still’parts of the pic-
`tures; the 'interframe prediction error is almost zero. Con-
`versely,_it
`is larger than intraframe predictionerror for the
`moving parts. Qn‘accuunt of these prediction error properties,
`the adaptiveprediction algorithm gives nearly optimum predic-
`tion, since the interframe Coding is used for still parts of pi;-
`turesand the intraframe coding is used for moving parts. it is
`notnecessary to trapsmit the prediction switching information
`because it
`is included in the magnitude of prediction error
`which is transmitted. The optimum value of TH depends upon
`picture contents, but a value around 10/256 is nearly optimum
`for most pictures.
`‘
`This adaptive algorithm provides better coding performance
`than the former NETEC-ZZH algorithm [1 1] based upon the
`third previous sample difference of the frame difference tech-
`nique. Computer simulation results are shown in Fig. 4, where '
`b a picture is panned at a speed of 0-11 pels/frarne. Fig. 4 shows
`that the coding performance of this algorithm is high compared
`with that of interframe,
`intraframe, and the third prévious
`sample difference of the frame difference coding techniques.
`Particularly, the improvement by the adaptive predictions is
`prominent for pictures panned at high speeds. As the speed
`becomes higher,
`the probability of interframe predation
`being selected tends to be 0.5. A theoretical analysis based upon
`a simplé‘signel model gives a result which agrees with the simu-
`lation study [18].
`s
`is
`2) Nonlinear Function (NL); A nonlinear function MI.
`applied to the frame difference signal, taking a key role in im-
`proving the coding periormance. The number of significantly
`changed pels caused by signal noise and quantization noise is
`greatly decreased by applying ML. The nonlinear function has
`an input-to-output relation as shown in Fig. 5. The transfer
`gain is less than unity for small input amplituderand unity for
`large amplitude. When the transfertgain of the frame difference
`path is less than unity, the loop transfer function of the inter-
`frame coding has a recursive type low-pass characteristic'along
`the temporal axis. The low-pass filtering suppresses random
`noise in the input signal which would otherwise cause un~
`wanted changed picture Clements even for still parts of pic-
`tures. Recurring quantization noise is also suppressed through
`the temporal low~pass filtering
`The nonlinear characteristic as shown in Fig. 5 is useful be-
`cause small amplitude noise in the still parts of pictures is sup‘
`pressed through the temporal low-pass filtering. and large man
`
`IEEE TRANSACTIONS ON COMMUNICATIONS. VOL. CbM-i9. NO. 12, DECEMBER 1981
`
`lntuvtrum
`
`3(Mb/s)
`
`‘(rm/poi)
`
`
`
`
`
`.0‘“f”""‘“'”"m—“"“"'T"~’“‘m“"r"’"7“”
`
`l 0
`(eel/trams)
`Penn m; Spaad
`4. Comparison of four kinds of prediction methods.
`
`Output
`
`V
`
`Fig. 5,
`
`Input-to-output characteristics of nonlinear function NL.
`
`[ll/256}
`
`litude frame differences caused by'morions are not affected at
`all. The nonlinear function causes distortion in pictures with
`small brightness change from frame to frame. The effect is
`hardly seen in NLl and NL2, although it may be perceived in
`NL3.
`.
`é
`‘
`The influence of the nonlinear function on static pc'rfonn-
`ance measures such as differential gain (DC) and differential
`phase (DP) is small. The results obtained show that DC is 2
`percent and DP 10°. The signal-to-noise ratio measured for a
`15.734 kHz sinusoidal wave input
`is 55 dB weighted, which '
`meets the broadcast picture quality requirement.
`3) Quanrizing Characteristics: Since the amount of infor-
`mation generated by the interframe coder varies with picture
`contents, it should be controlled to prevent the buffer memory
`from overflowing and underflowing. The information genera-
`tion rate can be controlled by changing quantizing character'
`istics as well as other parts of the algorithm. in other words,
`the information rate can be controlled at the cost of picture
`quality. Generally, if the quantixing step size is doubled. the
`information mic is decreased by l hit/pol. In this encoder.
`eight quantizers, Qt) Q‘Y. are used to provide; a gradual control.
`
`PMC3683170
`
`PMC Exhibit 2035
`Apple v. PMC
`IPR2016-00755
`Page 3
`
`
`
`\
`
`r
`
`KOGA era-1.: PERFORMAldCE ANALYSIS OF INTERFRAME ENCODER
`,
`TABLE 1 _
`QUANTIZING CHARACTERISTICS OF Q0, Qt, an o Q2
`Quantizuv Guam Levels
`lx 5/258)
`WA010 2.0 30 4051) SQTO SC 90105,.
`.
`i275r
`or
`0LSBOQSSSBSIOSIéSMS16,75
`rersl
`02 ozssorsiosissrss..,,,
`.
`i2? sl
`
`
`
`and one of them is adaptively selected by feedback control
`usingbuffer memory occupancy (BMO) values.
`,3
`i The quantizing characteristics of the finest three quantizers
`Q0, Q1, and Q2 are shown in Table I. The quantizing step size
`of—the finest quantizer Q0 is 1.0/256 for small amplitude, pro-
`viding an 8 bit PCM equivalent quality for small amplitude
`change. Those of Ql and Q2 are 15/256 and 25/256. respec-
`> tively, The number of quantizing levelsis 6! in Q0. The maxi-
`mum output level is chosen to be l27.5/'25§,,~which is large
`enough to prevent slope overload.
`‘
`'
`SN‘R representation of picture qpality for these quantizing
`characteristigs is made by assuming that the SNR is given by
`the minimum quantization step size, although the quantizing
`charactetistics are nonuniform. When the peak-to-peak lumi-
`nance amplitude (MM) is set to be 142/256, the relation be».
`tween SNR and the quantizing step size (5/256) can be expres-
`sed by the following equation,
`
`sun e 20 log1 o (Vpp/Nms),
`
`= 53.8 — :0 log} 0 .5"
`
`(dB unweighted)
`
`where
`
`& NH“S = S'Ni’i
`
`According to the equation, quantizers Q0, Q1. and Q2 pro-
`vide the SNR values of about 54, 50, and 46 dB unweighted,
`respectively.
`i
`e
`
`C. Variable Wordlength Coding
`
`In the iriterframe predictive encoding, the prediction error
`is_quantized and coded into a code with 6 bits/sample. The
`codes are transformed into reduced bit rate data through the
`variable wordlength curler.
`A
`The function of the variable wordlength coder is block
`coding for the significant pal positions and variable length.
`coding for the significant pel amplitude. The variable length
`code has two sets of codes. One is a variable wordlength code
`set with the code length ranging from 1
`to 12 bits. The unit
`length code is assigned to insignificant pels. This variable
`length code set
`is similar to a Huffman code and provides in-
`formation amounts almost equal to entropy values. The other
`is a fixedrlength code set with 6 bit length, which is used in
`conjunction with the variable one in order to avoid the con-
`tinuation of long codes. ’l'ransition between the two code sets
`is determined by comparing the prediction error amplitude
`with a certain threshold level. The transition information is
`not needed at
`the receiver: The use of these. two code sets is
`particularly effective for encoding pictures with violent mo~
`tions or detailed patterns because long codes are otherwise
`generated in these pictures.
`
`PMCAPL02442664
`
`Butter Memory occupancy (ENG)
`O 102030405060708090l00l99
`Ht—t—iv—ifi—t—r—Hq—t
`
`«———p —m
`175 [‘—n Y‘modc
`S
`H 00
`i
`o—. 0;
`0—. (:2N 03.— Dd
`‘—‘ 05
`H06
`..
`._.l_. 07
`
`I
`a i
`.
`1
`
`¢——u Ni.
`<*NLE'
`
`NL[
`
`N 3
`Fig. 6. Mode control diagram.
`
`Furthermore, continuation of the insignificantly changed
`picture element code is deleted by block coding. Picture ele—
`ments are divided into unit blocks. If all the samples in the
`unit block are changed insignificantly, the unit block is de-
`leted. The positions of the deleted unit blocks are represented
`by block address codes. The use of block address codes is quite
`useful for efficiently encoding still parts of pictures.
`_ Thus, the variable wordlength coder is capable of reducing
`themamount of encoded data with high efficiency, for all
`still. moderate, and active pictures.
`
`D, Coding Parameters Control
`
`Since the rate of encoded information varies with picture
`content, it
`is necessary to smooth out the irregular data gen-
`eration'by using the buffer memory. The capacity ofthe.buf-
`fer memory is determined from the propagation delay time
`tolerated by communication links. The buffer capacity used is
`about 1 Mbit.
`‘
`‘
`
`The control of the coding parameter combination among
`quantizers. nonlinear
`functions. and T/S-mode is made ac-
`cording to buffer memory occupancy (BMO) values. The con-
`trol diagram is shown in Fig. Elite BMO value is expressed in
`terms of precentage occupancy. The combination of NL] and
`Q0, for instance, is used for the BMO value ranging from 0 to
`15 percent. This combination provides the best picture quality,
`equivalent
`to the 8 bit PCM accuracy. As BMO increases, the
`coarser quantizers are used. That is, the significance determina-
`tion level varies depending upon BMO values. Transition be-
`tween T—mode and S-mode operations occurs at the BMO value
`of about 55 percent. When S-mode operation occurs for large
`BMO values, only a slight degradation may be perceived in the
`horizontal resolution, although it
`is rare. The combinations of
`the coding parameters are determined, after examination for
`varieties of pictures, so that the relation between picture qual-
`ity and the rate of encoded information can be made optimum.
`
`III. NETECZQH CODING PERFORMANCE FOR
`BROADCAST TV PROGRAMS
`A. Statistical Measurements
`
`Coding performance is often represented by the resulting
`SNR. However. in irrteri‘ramc coding. SNR varies with picture,
`contents to be“ encoded because quantizing characteristics are
`changed according to the rate of generated information. As
`there is no qualified ohicctivc cvalualioi‘r method for motion
`
`PMC3683171
`
`PMC Exhibit 2035
`Apple v. PMC
`IPR2016-00755
`Page 4
`
`
`
`
`
`ProbabilityorinformationRateThe!Exes-ct:xoMal:
`
`v [- Long—firm
`[I
`Avemql 'PUlo)
`I
`.E.
`l
`'0.|——-—-——ri'
`0
`to
`
`Long-Jim Average
`
`"PLAYING cancers”
`
`Il
`
`
`
`information Raf.
`
`lXal
`
`Fig. 7.
`
`Statistical information rates (Q1, NLI). Peru) is a long—time
`average probability density function.
`
`PMCAPL02442665
`
`' video at present, coding performance is evaluated from source
`signal statistics about‘information rate “and coding parameter
`statistics.
`‘
`in order to estimate information content, the coder param-
`eters are fixed so that a certain picture quality is provided.
`then the probability distribution of the rate of generated infor-
`mation is measured for actual television signals for many hours.
`From the probability distribution thus obtained, the proba-
`bility of providing the picture quality at a given transmission
`bit rate is calculated. In the measurements, information rates
`per frame were measured at the output of the variable word-
`length coder (point A shown in Fig. l) with coding parameters
`fixed; i.e., no feedback control. One quantizing characteristic,
`Q1, which provides an SNR of 50 dB unweighted slightly
`better than broadcast picture quality for long-haul transmis- ,
`sion, is used constantly in conjunction with NLl under T—mode
`operation. Measurements were made for about 36 h applying
`off-the-air TV signals from four different broadcasting sta-
`tions, received by a TV receiver. The SNR of the received TV
`signals was about 40 dB, and random noise was perceptible.
`The cumulative probabilities of information rates obtained
`for three typical programs and the long-time average distribu-
`tion are shown in Fig. 7. Here, the information per frame is
`multiplied by 30 (frames/s), and the horizontal axis is expres-
`sed in Mbits/s. The three fine solid lines are examples showing
`that generated information rates are cdnsiderablydiffcrent be—
`cause cf different image activities.
`“Art Lecture" is an example of inactive pictures, in which
`most of the time is occupied by still or almost still pictures.
`This cumulative prbbahility distribution shows that 99 percent
`of the program material can be transmitted at 20 Mbits/s by
`using Ql (NLl) because the probability accumulated from 0
`to 20 Mbits/s is 99 percent. The average information rzrtein
`this program is as low as 8.2 Mbits/s. In “Playing Children,“
`
`KIEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-7.9, NO. 12. DECEMBER 1981
`
`many children are always running around, and cameras are fol-
`lowing them or frequently switched. Becausa of the motion of
`the children and of the camera, the cumulative probability at
`20 Mbits/s decreases to 73'percent. However, the cumulative
`probability at 30 Mbits/s is 99 percent. “Super Bowl ’79” was
`the most active program encountered in the measurement.
`This program is a football game, which was broadcast in Japan
`in January 1979 after reception via satellite relay and was re-
`corded’in a U-matic VTR from the off-the-air signal. For this
`program, the cumulative probability at 20 Mbits/s is 35 per~
`cent; however, it is still 96 percent at 30 Mbits/s. The average
`information rate is as high as 23.5 Mbits/s, being 15.3 Mbits/s
`higher than that of “Art Lecture.“ The above data indicate
`the range of broadcast TV signal source variations.
`The long-time average probability density function P(Xo)
`over 36 h is shown with a broken linerin Fig. 7. The mean ‘
`value of this average distribution is no more than [5.3 Mbits/s.
`The probability of the rates greater than the average value
`rapidly decreases. and a small peak is seen at around 30—40
`Mbits/s. This small peak is due to scene changes. It is interesting
`that the shape of this distribution is quite similar to that ob-
`tained by Seyler [19] ._ where the number of significant frame
`difference pets was used as ameasure of information rate.
`The bold solid line, representing the cumulative probability
`of the average distribution, shows that 93 percent of the pro-
`gram materials can be encoded at 20 Mbits/s and 99 percent
`of them can be encoded at 30 Mbits/s when a quantizing char-
`acteristic Ql (NLI) is used. With a quantizer coarser than Q],
`the rate of generated“ information becomes smaller. Therefore,
`when the adaptive control is operated, the residual probability
`of 7 percent at 20 Mbits/s and 1 percent at 30 Mbits/s will be
`transmitted by using coarser quantizers providing slightly
`lower SNR,
`"
`""“
`
`H. Information Generation Control
`lit the encoder under normal operation, quantizing charac-
`teristics are adaptively changed in order to control the rate of
`the information generation. As “Super Bowl ’79” is an ex-
`treme case and gives a very large information rate, it is used as
`an example to show the performance under feedback control.
`The probablity distributions of information rates with three
`different quantizers, Q0 (NLl), Ql (NLl), and Q2 (NLZ), are
`compared in Fig. 8. The average values with the three quan-
`tizers are 28.6, 23.5, and 16.7‘Mbits/s, respectively. Then the
`difference between ’Q0 and Q1 is, about 5 Mbits/s and that
`between Q1 and Q2 is about 7 Mbits/s. These results show the
`effectiveness of changing quantizing characteristics to control p.
`information generation. It can be calculated from Fig. 8 that
`when the adaptive coder parameter control
`is made at the
`transmission bit rate of 30 Mbits/s, 82 percent of the program
`is transmitted with Q0, 14 percent with Q1, and 3 percent‘
`with QZ, respectively. The residual percent is transmitted by
`using coarser quantizing characteristics
`than these three.
`About half of the residual
`1 percent is considered to be due
`to scene changes.
`At the moment of scene changes. frame—twframc correla—
`tion will be lost in general and. therefore. the intral‘ranic pre—
`diction will he used in the adaptive prodit‘livn encoder. The in-
`
`PMC3683172
`
`PMC Exhibit 2035
`Apple v. PMC
`IPR2016-00755
`Page 5
`
`
`
`KOGA 22‘ (IL: PERFORMANCE ANALYSIS OF IN TERFRAMEFNCODER
`
`NETEC-22H ENCODER
`
`r")/
`
`
`
`
`
`3
`
`6
`
`
`
`
`
`ProbabilityofInlormntionRotamatExceedsx0Mb/s
`
`.
`
`SUPERBOWL ”
`
`(I: Ouonhzer
`
`lo
`
`20
`Iniarmatlon Slate
`
`30
`(x0)
`
`45
`4c
`(Mb/s)
`
`Fig. 8. Generated information rates for “Super Bowl ’79" by three
`different quantizers.
`
`formation rate in the case of the intraframe prediction ranges
`mostly from 3 to 4 bits/pel when Q1 is applied. This means
`that, at 30 Mbit/s transmission, the excessive information of
`1—2 bits/pel can be suppressed by using quantizers Q2—Q4.
`Then, the scene change influence on picture quaiity is an SNR
`decrease to 46—40 dB unweighted. in this case, the EMU value
`is raised to about 50 percent (see the control diagram in Fig.
`6). After scene changes, the influence will disappear within a
`few frames.
`q<4
`
`PMCAPL02442666
`
`IV. ADAPTIVE BIT SHARING MULTIPLEXER (ABS-MUX)
`A. ABS-MUXPrinciple
`It has been shown that the long-time average of encoded in-
`formation is about 15 Mbi'ts/s. The probability of information
`rate greater than the average rapidly decreases as the rate iti-
`creases. This
`indicates that busy pictures with large-scale
`rapid motion are not frequently encountered, and that they
`rarely appear at the same time among programs from different
`stations.
`Therefore, more effective channelutilization can be achieved
`by multiplexing data from multiple encoders based upon an
`idea similar to TASI *[20]. ABS-MUX multiplexes plural chan-
`- nels by adaptively allocating a bit rate to each channel depend-
`ing upon individual information rates, while the total bit rate
`is kept constant [16].
`The block diagram of the three-channel ABS-MUX system
`is shown in Fig. 9. Receiving the EMU (buffer memory oc-
`cupancy) value from each encoder, ABS~MUX decides the bit
`rate to be assigned to each encoder with the total bit rate kept
`constant at 60 Mbits/s. The average hit rate per channel
`is
`20 Mbits/s in the three-channel ABS-MUX and 30 Mbits/s in
`the tw0~channcl ABS-MUX. The hit rate assignment decision
`is made every 153.6115, one ABS frame time. The ABS frame is
`as”
`'
`
`NETEC-EZH DEOODER '
`DAT-Al
`CLOCKI
`Data 2 '
`CLOCK 2
`DATA 3
`CLOCK3
`
`-
`.
`
`-
`
`I
`
`~
`
`fig. 9;, Three-channel
`
`transmission system configuration with ABS~
`NIUX.
`
`Photograph of a transmission system with ABS-MUX. (From
`Fig. 10.
`left to right. three NETEC-ZZH encoders, ABS—MUX/DEMUX, four-
`phase PSK modern, and two NETEC-ZZH decoders.)
`
`composed of 18 unit frames, with each unit frame consisting
`of 256 X 2 bits. The bit rate assignment is made by changing
`the number of unit frames assigned to each channel in one
`AB$ frame. Since one ABS frame has I8 unit frames, the as-
`signed bit rates are combinations among 16.7, 20.0, 233,3nd
`26.7 Mbits/s. For instance, 16.7 Mbits/s corresponds to five
`unit frames in one ABS frame.
`
`B. ABS-MUX Operation
`
`A photograph of the ABS-MUX system is shown in Fig. 10. ~
`The photograph shows, from left to right, three NETEC—22H
`encoders, the ABS-MUX, a four-phase PSK modern) and two
`NETEC-ZZII decoders.
`‘
`'
`.
`.
`Two system operations are considered here: two-channet
`and three~channe1 ABS-MUX systems.
`'
`H
`ABS-MUX operation measurements are made for these two
`system configurations by using the hardware system. Diffeient ,
`broadcast TV programs are supplied to the encoders, and the
`bit rate assignment is made by ABS-MUX. ABS-MUX opera
`tion was measured by monitoring the BMO values of the en-
`coders simultaneously for 5 .5 h.
`.
`'
`,The’ probability distribution of bit rates assigned to one of
`the encoders is shown in Fig. 11. In the three-channel ABS~
`MUX case, the probabilities ofqtransmission at 16.7; 20.0, 23.3,
`and 36.7'Mbits/s were 9.08 percent, 77.71 percent. 12.07 pet- "
`cent, and H4 percent, respectively. The probabilities ofbit;
`rates assigned for
`the three encoders are similar, The average _
`bit
`rates of the three encoders are 30.”. 20.0.1. and 19.81
`MbiIs/s, nearly equal to the fixed bit rate.‘l'hesc indicate that ,
`a”;
`‘
`N
`
`PMC3683173
`
`PMC Exhibit 2035
`Apple v. PMC
`IPR2016-00755
`Page 6
`
`
`
`.
`‘
`u
`IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. COM-29, NO. 12, DECEMBER 1931
`
`IDO
`
`3—chonnel ABS
`
`‘
`
`at
`
`o.
`
`chobilityoiASSIunGflBitRate
`
`
`
`PROBABILITYDENSIYY
`
`we“... r ; -
`0
`5O
`'-
`5M0 Value
`
`
`
`PROBAEiLITVDENSITY
`
`EMU value
`
`(‘4),
`
`Fig. 12.
`
`Performance improvement by ABSvMUX. (a) Three-channel
`ABS-MUX. (b) Twrxchannel ABS-MUX.
`
`IG.7
`
`200
`
`Avenue - 20.15 Mb/s
`'2-07
`|.l4__-4t_~A—_l_.
`235
`26.7
`30.0
`33.3
`36.?
`Ra's
`(Mb/s)
`Assigned Bi!
`(3)
`J
`
`' Zecnonnel ABS
`
`
`
`ProbabilityorAssignedanRate
`
`PMCAPL02442667
`
`D. SA? Representation
`,0 As stated before, evaluation of the interframe coding per-
`formance is a difficult problem, since SNR varies with picture
`contents and SNR is partly improved through the nonlinear
`temporal filtering for still parts of pictures.
`'
`Here, in order to obtain an objective measure, an SNR esti-
`mation is made from the BMO statistics. Fig. 13 shows the
`estimated SNR probability calculated from the BMO statistics
`in Fig. 12(3) and (b), using the relation among BMO values,
`selected quantizers, and corresponding SNR’ values [see Fig.
`6 and (3)] . It is seen by summing probabilities for worse SNR
`that the cumulative pr'élbability of pictures being transmitted
`with SNR values equal to or greater than 50 dB unweighted
`was 99 percent in thes30 Mbit/s transmission case. The SNR
`probability curve estimated for the 20 Mbit/s transmission
`casa is about 6 dB lower than that for the 30‘Mbit/s case This
`is a reasonable value since the bit rate difference is nearly equal
`to l bit/pal, corresponding to 6 dB in SNR.
`As far as the average SNR is concerned, ABS-MUX provides
`only a slight improvement. However, it should be noted again
`that, in '20 Mbit/s coding, the occurrence of such an objection
`ably low SNR as 3340 dB caused by Q4, Q5, etc.. can be
`greatly reduced by using ADS-MLTX.
`Thus. it can be said that .ArllstUX is useful particularly in
`the lower bit rale range. At the higher bit trite, the effect has
`comes less ()lWlOllS. Ilowovvr, a variety of picture contents
`may be involved in broadcast 'l V signals. therefore. the two-
`
`059 004
`333 as7
`(Mb/s)
`
`4
`
`Average - 29.99 Mir/s
`0.55
`I
`DID
`267 300
`6.7
`23.3
`zoo
`Assigned Bit Rate
`(b)
`Probability of bit rates dynamically assivned by ABsiMUX.
`Fig. 11.
`(a) Three-channel ABS-MUX transmission.
`(is) 'l‘wochannel ABS~
`MUX transmission.
`4
`
`the. bit rate sharing operates effectively among the three chan-
`nels, being active about 9 + 12 + l = 22 percent of the time,
`I and that picture quality improvement can be expected for any
`channel over the fixed bit-rate transmission case about 12 +
`»1 = 13 percent of the time. In the two-channel ABS-MUX
`‘ case, however, the bit rate change hardly appears. This is be-
`cause thch are very few pictures whose information amount
`exceeds 30 Mbits/s. It is seen that ABS-MUX operates to alter
`the normal sharing for only 1.41 percent of the time.
`
`C. Coding Performance Improvement by ABS-MUX
`
`Coding performance improvement cangbe achieved by ap-
`plying ABS-MUX. As the BMO values-relate to the bit rate
`assignment and the quantizer selectiongthe perfoimance irri-
`provement can be evaluated from the BMO value statistics.
`Another encoder is op