356
`
`PROCEEDINGS or THE I]-ZEE. VOL. 55. NO. 3, MARCH. 1967
`
`Results of a Prototype Television
`Bandwidth Compression Scheme
`
`A. H. ROBINSON AND COLIN CHERRY
`
`III. DI-.‘l'AI]. DEl'EC‘l'ION
`
`A band-limited television signal may be completely de-
`scribed by samples taken at the Nyquist rate. The “detail
`detector“ is required to locate successive Nyquist samples
`of equal amplitude (i.e., which fall on a common “run”) in
`order that run-length coding may be implemented. Infer-
`ential methods are exploited, aiming to discriminate against
`noise present in the signal. In this case the “runs” constitute
`successive samples of equal amplitude (within prescribed
`fidelity constraints), although "runs" of other signal fea-
`tures are feasible as possible alternatives.
`The detail detector used in the work described here is an
`
`improved version of the one built by Kubba [2] and used
`by Cherry et al. [1]. The improved performance of the new
`detail detector is described by Vieri [3], who reports a reduc-
`tion of approximately 20 percent in the number of samples
`needed, compared with the earlier version, with very little
`change in quality for typical half-tone pictures. He also
`notes that the improved version gives a run-length distribu-
`tion which is more nearly exponential; such a result con-
`forms with an exponential shape for the autocorrelation
`function as measured by Kretzmer
`[4]. An alternative
`analysis of the detail detector operation, proposed by Pear-
`son [5], regards the problem as a significance test on a null
`
`foodhnnk link
`
`
`
`
`GU33 NB
`(nuns! urn)
`
`mu-tum
`nsnxcron
`
`I-um-an
`axurra
`
`
`
`nrvrnn
`an
`
`uumn nmmu.
`ccnsrm
`
`no an
`acorn
`
`uvtr- r
`
`uncr -
`
`-
`
`xusue noon:
`
`I
`
`an-91 ima-
`nine).
`
`Fig. l. Transmitter encoder.
`
`ORACLE 1006
`
`Page 1 ofll
`
`
`
`Abstract-‘Ike transmitter/receiver system for bandwidth or data-rate
`eonpressiononeIevisiondeIIk.descnbadherem.saproco:ypemode1or
`theexperlmentalsysteInofCherryetnl. [l].ThesysteInissIIitn.b|el'orbot1I
`black-and-white or half-tone pictures, in realistic noise conditions. Thesystem
`panneteu-sunyheodjnstedsothatnnoptimunru-lusgdieneodingnlay
`befoudztlnegreatndnntagesofrui-luigthqnnnfizlngnndlowfl-Csflfichfly
`withregudtopncflcnlhstn-wnanafion,!eadlngtodieueofbufl'erstores
`ofnodstenpnclty.0nepu1icnlnrehapfor-nofreeeivet-opentesonn
`quantized-varhblevelocity principle and, being much more single and
`dneapfluntheun|smitter,isstIitah|eforI:eina'tnndonsreqI!rhgnnny
`receivers.
`
`I. Imaonucriou
`
`HIS PAPER EXTENDS the experimental and theo-
`
`Tretiml work of the open-loop compression system
`
`proposed and partly described in an earlier paper
`(Cherry et al.
`[1 ]), leading to the description of a flexible
`prototype television bandwidth compression system for in-
`dustrial and commercial purposes, or for long-distance
`links in broadcast systems. of full broadcast quality, for
`“on-line” trials. The system is basically that described by
`Cherry et al. [1 ], to which extensive reference will be made,
`though developments have led to differences of detail, and
`include an increase in flexibility by provision for variation
`of certain system parameters.
`
`11. Drsciupriow or SYSTEM
`
`The block diagram shown in Fig. 1 contains the essential
`details of the prototype transmitter encoder. The separate
`blocks of Fig.
`l have been outlined in [l], but a brief de-
`scription will be given here for completeness.
`The open-loop system combines attempts to exploit cer-
`tain statistical and subjective redundancies in the television
`signal, with the aim of reducing the average rate of binary
`digits required to specify the picture information with given
`fidelity. Attempts are made to extract the essential signal
`data in a noise-combating fashion for subsequent statistical
`run coding. The resultant nonuniform data rate is then
`equalized by feeding into a buffer store (called an “elastic
`encoder”) for uniform transmission through a suitable
`channel. Two sets of data must be transmitted, the bright-
`ness data and the run-length data.
`
`Manuscript received December 1. 1966; revised January 27. 1961'.
`This work was supported by the ‘National Research and Development
`Corporation and the Science Research Council.
`The authors are with the Department of Electrical Engineering. Im-
`perial College. University of London, London, England.
`
`

`
`ROBINSON AND CHERRY: TV BANDWIDTH COMPRESSION SCHEME
`
`357
`
`hypothesis that two samples belong to the same population.
`The resultant strategy is the same as proposed by Kubba
`[2] except that the thresholds can only be optimized accord-
`ing to subjective tests perfonned on the resultant pictures.‘
`
`This run-length restrictor is further provided with a ca-
`pability of modified restriction when commanded by the
`elastic encoder during “underload" and “overload" con-
`ditions; see Fig. l.
`
`IV. RUN—LENGTH Rssnucrton
`
`V. THE ELASTIC Encoosa
`
`The detail detector output corresponds to a train of
`amplitude-modulated samples obtained by removing those
`Nyquist samples which do not begin new runs. These runs
`may be of any length (put each an integral number of
`Nyquist intervals), and are suitable for Shannon-Fano cod-
`ing, though the number of run lengths possible in practice
`is large, thereby creating formidable instrumentation prob-
`lems. Measurements of the run-length probability distribu-
`tions (Vieri
`[3] and others) indicate an exponential dis-
`tribution with negative exponent, so that an upper limit on
`the maximum pennissible run length would be desirable
`and reasonably efficient. An even more economical method
`of instrumenting run coding is to restrict the permissible
`run lengths to a small subset of the original distribution, all
`other run lengths being broken up into suitable combina-
`tions of these standard runs by insertion of additional
`samples. The process described by Cherry et al. [1] is cailed
`“run-length restriction." It is envisaged that the optimum
`restricted run-length probability distribution will be almost
`flat (in which case variable-length coding will be unneces-
`sary), but this can only be experimentally determined.
`The “run-length restrictor" consists of a shift register of
`eleven stages and associated set-reset logic, and is capable
`of restricting an input train of runs, labelled by a Nyquist
`train of l‘s (start of runs) and 0's (remaining members of
`runs), to a train of standard runs by the insertion of extra
`l‘s into the run-length sequence as it passes down the shift
`register. The standard runs, labelled 1. p, qr, and r. where
`l spsqsr. are permitted to assume a range of values so
`that the experimental detennination of an optimum set is
`possible. Measurements and calculations, Vieri [3]. suggest
`that the following combinations of runs will
`include an
`optimum set.
`
`The run-length restrictor output now corresponds to
`amplitude-modulated samples obtained by removing those
`Nyquist samples that do not begin any of the now standard
`runs. This irregular sample train is next converted into a
`regular one, having a reduced Nyquist rate, using a shift
`register technique (Fig. 2). This is done by stacking the
`samples in a storage device, as they occur, and emptying it
`at a uniform rate by the regular removal of samples at the
`head of the stack. The situation is one frequently en-
`countered in queueing theory and it will be discussed in
`more detail in the Appendix.
`A practical storage device. called an “elastic encoder."
`for equalizing the sample spacing has been described by
`Cherry et al. [1]. If the amplitude-modulated samples are
`first converted into binary PCM form, then the elastic en-
`coder can be realized by a bank of shift registers operating
`in parallel (see Fig. 2). Each amplitude-modulated sample
`is converted into parallel binary PCM form and the digits
`inserted into the “stacking stage" of the “brightness shift
`registers“ (see Fig. 2) so that one shift register is required for
`each digit of the PCM signal. Simultaneously, the corre-
`sponding standard run length is coded into a binary PCM
`signal and inserted in the same “stacking stage“ of the
`“position registers." Each set of corresponding storage
`cells in the parallel arrangement of shift registers constitutes
`a stacking stage. In our present case, using 4 standard run
`lengths, 2 “position registers" are required.
`An extra "control register" is necessary to operate the
`associated logic which moves the stacking stage up and
`down the registers in step with the input and output se-
`quences. The input and output sequences must be arranged
`so that input and output pulses never occur at the same
`instant. and this can only happen when the Nyquist rate is a
`
`Run
`
`N-Q?-
`
`Length (Nyquist units)
`
`to
`
`bike)
`
`#31
`
`Until
`
`O\ON
`
`~.I~J
`
`storage all
`
`m_‘“"“
`Inputs
`
`HEIEIEIGEEDDI-JEIUIIEIII
`
`Rflflhu.
`legusu-I
`
`We have considered it necessary to consider a wide range
`of standard run lengths, even though the overall compres-
`sion for any particular set can be predetermined by numer-
`ical calculations on the run-length probability distribution,
`because the subjective effects of restriction needed assess-
`ment.
`
`‘ An alternative theoretical approach to detail detection has also been
`reported by Sekey [6].
`
`*°'"*°°
`
`Tllfllte
`
`Oahu
`
`I
`
`I---'
`
`I
`raaaaramamanunsi
`
`.
`T2323
`
`r~'
`—t:':':'
`It-coking INC! '3' :
`
`reduce
`
`Iolieten
`
`}
`
`cut:-at
`:
`EIE.IE1ElE!ElElElEIlIlllll } M-*°*
`Fig. 2. Elastic encoder.
`
`Page 2 ofll
`
`

`
`358
`
`PROCEEDINGS OF THE IEEE
`
`MARCH
`
`multiple of the output reduced Nyquist rate. (This is the
`sampling ratio It to 1, n an integer.)
`The position “X” of the stacking stage is monitored at
`both ends of the control register so that “underload“ and
`“overload” conditions can be detected and used to modify
`the operation of the run—length restrictor in an optimum
`fashion. The underload and overload facilities should not be
`confused with "elastic encoder feedback," described in
`Section VI, which considers the required elastic encoder
`capacity.
`
`Vl. CAPACITY REQUIREMENTS
`
`The equilibrium behavior of a finite-size elastic encoder
`with a stationary random input is considered in the Ap-
`pendix, where it is suggested that a run—coded television
`signal can be usefully considered as a random sequence of
`samples having "short-tenn" stationarity (“short-term”
`meaning ofthe order of one line length). It is suggested that
`the mean rate of this random sequence exhibits “long-
`term” variations only.
`The “long~tenn“ variations of mean sampling rate can
`be minimized in the practical situation since the detail de-
`tector constitutes the sample source. It merely requires
`that the detail detector thresholds be adjusted by feedback
`from the elastic encoder.
`
`It should be noted that a change in the detail detector
`thresholds, corresponding to a change in the detection
`fidelity constraints, results in a change in the proportion
`of Nyquist samples selected. but the resulting run—length
`distribution remains essentially exponential in “shape.”
`The detailed behavior of the elastic encoder, and in par-
`ticular the required storage capacity, depends upon the
`actual form of feedback, the allowable proportion of time
`in the “overload" condition, i.e., with the elastic encoder
`completely full, and also upon the scanning standards em-
`ployed.
`The requirement of constant line-scan period is equivalent
`to restricting the elastic encoder to be in the same state at the
`end of each line, i.e., the same number of stages are filled
`at the end of each line. The basic efi'ect of this form of feed-
`
`back is to coarsely quantize the scanned lines containing
`excessive detail and to finely quantize lines containing little
`detail, and is thus broadly equivalent on a macroscopic
`scale to the coarse quantization of high-frequency com-
`ponents of the signal described by E. R. Kretzmer [10].
`A less drastic form of feedback is possible if a variable
`line-scan period is permitted, as is the case in the results
`reported here. When the number of samples in a line is in
`excess of the average number per line, then, in one line
`period, the elastic encoder will fill a net amount equal to this
`excess. Conversely, when the number of samples in a line is
`less than the average, then, in one line period. the elastic
`encoder will empty a net amount equal to this deficiency.
`The probability of a large net increase or decrease in the
`number of filled stages in one line period can be controlled
`by the amount of feedback employed. In this case, the
`“slow" feedback should aim at keeping the trafiic intensity
`(see Appendix) just less then unity. The actual movement
`
`of the stacking stage during one line-scan period can use-
`fully be used to serve as the "slow" feedback signal by tak-
`ing advantage of line-to-line correlation.
`The limiting case of the "slow“ feedback is that in which
`the feedback link is completely removed. The elastic encoder
`will then fill and empty according to the picture content,
`the sampling ratio :2, and the setting of the detail detector
`thresholds. The maximum variation in line duration in this
`
`case is proportional to the elastic encoder capacity. The
`“long-term" variations in mean sampling rate in this
`limiting case determine the amount of “underload“ and
`“overload" in any particular picture; an increase in the
`detail detector thresholds generally causes an increase in
`the amount of “underload" and a decrease in the amount of
`
`“overload." Generally "overload” reduces picture quality,
`while “underload“ inserts redundant samples with little
`benefit; i.e., the latter process is wasteful.
`The experimental results in Section IX, using an elastic
`encoder of 30-stage capacity and with the feedback link
`disconnected, together with further results (Robinson {8])
`in which an infinite capacity store is simulated, again with no
`feedback, suggest that storage approximately equal to the
`average number of runs per line together with a small
`amount of “slow" feedback would be adequate. An indica-
`tion of the variability of the line statistics (i.e., the probabil-
`ity distribution of the number of run-coded samples per
`television line) has been given by Vieri [3], who found that
`“only 5 percent of the measured lines used a number of
`runs more than 50 percent greater than the mean.” This
`tends to support the estimate of the required storage capac-
`ity, given above.
`Further benefit might accrue if a form of "fast" feedback
`is also used, in which the queue length exerts some con-
`tinuous detail detector control.
`
`It should be noted that if the run-coded sample train can
`be considered to have independent successive run lengths
`and a probability distribution other than exponential, then
`the variance of the number of samples arriving over a long
`period of time is smaller than that for the completely ran-
`dom situation. Note that run-length restriction changes the
`probability distribution; the limiting case is of course the
`distribution in which all run lengths are equal. It can be
`concluded that the assumption of a completely random
`distribution errs on the conservative side.
`
`VII. VARIABLE Vsnocrrr Rracarvrzn
`
`The most obvious form of decoder consists essentially of
`an elastic decoder of equal capacity to the elastic encoder
`at the transmitter, in which case conventional receivers
`would be used. However, more economical and simple
`receivers are especially attractive wherever broadcast
`“many-receiver" situations exist. Accordingly, an experi-
`mental study using a quantized-variable-velocity principle
`is described (Robinson [8]) in which decoding and display
`are perfonned on a single CRT. A 536 Telrtronix oscillo-
`scope was used for convenience because it had a wide band-
`width deflection system, so that the displayed raster was
`limited to 3-by-4 inches in size.
`
`Page 3 ofll
`
`

`
`I967
`
`ROBINSON AND CHERRY: TV BANDWIDTH COMPRESSION SCHEME
`
`359
`
`The “quantized-variable-velocity receiver” is outlined in
`Fig. 3. The run-coded samples will be correctly positioned
`on the final display. provided the horizontal time base is
`controlled so that its velocity is switched between the three
`or four quantized velocities proportional to the restricted
`run lengths. The time-base is conventional, with switched
`velocity control obtained by feeding into the miller time
`base grid as time-switched current having the three or four
`values proportional to the quantized run lengths. The con-
`trol circuit current sources are accurate to approximately
`1 percent. Brightness compensation is considered in the
`next section.
`
`It should perhaps be emphasized here that continuous
`beam velocity control is not required; only a number of
`distinct velocities are needed, corresponding to the three or
`four quantized (restricted) run lengths. Thus, the corre-
`sponding transmitted “position signal," having only one of
`three or four values, is highly resistant to channel noise.
`
`VIII. Bruor-rrmns COMPENSATION
`
`Brightness compensation (Robinson [8]), required to
`correct the variations in display brightness due to changes in
`horizontal scan velocity, is considered in Figs. 4 and 5.
`The necessary dynamic range of a variable-velocity dis-
`play tube is larger than that for a conventional receiver, so
`that nonlinear tube characteristics assume greater sig-
`nificance. Various practical approximations of typical CRT
`
`unrplnudo expel
`
`petition Iifill
`
`characteristics are considered and in each case the necessary
`compensation is described.
`The brightness signal 12 is added to a pedestal co, the sum
`amplified in a linear amplifier, and connected to the grid of a
`display tube.
`Thus, V= Ke= K(v+ vo).
`Assuming that screen brightness is proportional to beam
`current I, and that I, and 1, correspond to black level and
`white level, respectively, at any one velocity, we require 1,
`and 11 to change in direct proportion to any change in veloc—
`ity. This change is obtained by appropriate changes in the
`gain K of the linear amplifier. Consider the following
`CRT characteristics:
`
`a) linear CRT characteristic—the whole system is linear
`and K is directly proportional to velocity
`b) constant gamma CRT characteristic—if the time base
`velocity increases by a factor k, then correct compensa-
`tion requires an increase in amplifier gain of k‘”
`c) piecewise linear CRT characteristic—correct com-
`pensation requires changes in amplifier gain matched
`to the tube characteristic, and also requires a different
`pedestal v0 for each velocity
`d) a piecewise constant gamma CRT characteristic can
`be postulated in which the compensation method is as
`in c).
`
` video innit V
`
`V - [g - [(vw°]
`
`13.6. pedestal
`
`Fig. 4. Brightness compensation.
`
`hauler-lotto
`
` llpltflor
`
`I
`
`Maul: ¢.I.'l
`
`_._nu-no‘‘ _I_c;h_1ie_ _
`1 - cl
`
`I
`
`Gusltnt 01: 0.1.1.
`manner-hH.u
`[IV
`I
`
`,-
`/
`
`"536'
`0812114409007!
`
`
`
`\§ gm
`Ive and nut! to c.|.!.
`
`Iva
`
`input up to e.|.7.
`
`v
`
`Fig. 3. Prototype variable-velocity receiver.
`
`Fig. 5. Amplifier and CRT characteristics.
`
`Page 4 of 11
`
`

`
`PROCEEDINGS OF THE IEEE
`
`MARCH
`
`IX. EXPERIMENTAL RESULTS
`
`The transmitter performance can be assessed without
`building an elastic decoder system, by reconstructing the
`pictures from the samples at the run-length restrictor out-
`put. though actual pictures taken from an experimental
`quantized~variable-velocity receiver are also illustrated.
`The reconstruction uses a "box-car generator” as a sim-
`ple fonn of interpolation (Fig. 6). Although the samples
`were taken from the run-length restrictor output, they were
`also fed into the elastic encoder so that the “underload" and
`“overload” effects were observed.
`The feedback link was not connected in these experi-
`ments. although of course the “underload" and “overload“
`facilities were maintained; the detail detector thresholds
`were set manually to optimize the results by reducing the
`average sampling rate until overload effects largely dis-
`appeared.
`The system signal-to-noise ratio was approximately 40
`dB, and the elastic encoder had a 30-stage capacity. i.e.. it
`had storage for 30 run-coded samples.
`Typical quantized-variable-velocity pictures are shown
`in Fig. 7; run lengths are labelled ((1, p, q, r) and relative
`amplifier gains are labelled g(K,, K1, K3, K4).
`ma-Iunh z-uuiaud.
`nflsloe
`
`Vidlo
`
`
`Initially the CRT was assumed to be linear: however,
`better results were obtained with a gamma of 1.3 (using a
`gamma corrector at the transmitter), and the brightness
`compensations were adjusted to be more in accord with
`the values indicated in b). The actual values selected were
`obtained experimentally, since the possible values had been
`preset according to a).
`The bandwidth of the horizontal defleetion circuits was
`10 MHZ.
`
`X. DISCUSSION or RESULTS
`
`The simulation results in Fig. 8. in which the sampling
`ratio n= 3, and the detail detector was set to minimize the
`effects of “overload." are all single-frame pictures; con-
`sequently. a certain amount of noise in these photographs is
`in excess of that obtained by direct viewing on a television
`monitor. The experimental elastic encoder had been made
`of only a 30-sample capacity: the expected overload distor-
`tion could then be studied subjectively. Full use of the detail
`detector capability therefore was not possible because the
`traffic intensity (see Appendix) was set at an artificially low
`level in order to minimize “overload.” producing an excess
`of samples inserted during “underload." This efiect can
`clearly be seen in Fig. 8(c). The effect of restriction can also
`be seen in Fig. 8(c) as spurious contours in the “detail func-
`tion."
`
`The results shown in Fig. 8 are those one would expect
`from an ideal receiver. provided there are no errors in the
`transmission channel. except that here the “box-car“ form
`of interpolation is employed, which can only produce a
`poorer picture than a full decoder; it cannot give a better
`one.
`
`Further results (Robinson [8]) indicate that with a store
`of larger capacity (simulated by disconnecting the “under-
`load" and “overload” links to the run-length restrictor).
`sampling ratios of about six can be achieved with results
`similar to those illustrated in this paper.
`
`
`
`(a) Variablewelocity reconstruction. RI. 2. 4. I0), g(l. 2. 4, S). The trans-
`mitted signal had an artificially low signal-to-noise ratio of 30 dB:
`n=4.
`
`(D)
`(b) Variable-velocity reconstruction. [(1. 2, 4, l0), g(l, 2. 4. 9). n =3.
`
`Nate : Quantized-variable-velocity reconstructions. 32-level amplitude quantization only (see text). and 30—stage elastic encoder used at tra ns-
`rnitter (photographic exposure I £5 second). In ch ofthese photographs. the received signal, including synchronization. was of5-bit quality only.
`Fig. 7.
`
`Page 5 ofll
`
`

`
`I967
`
`ROBINSON AND CHERRY: TV BANDWIDTH COMPRESSION SCHEME
`
`36|
`
`The most surprising result of this investigation has been
`the effect of run-length restriction on the quality of the final
`picture. The calculations on the number of extra samples
`inserted by this process (Vieri [3]) were confirmed, but their
`subjective effect has been one of improved quality and not
`the reverse as might have been anticipated. Indeed,
`the
`contrary is the case; the extra samples appear to give even
`better results than would a similar number inserted by firs:
`removing the run-length restrictor and then increasing the
`detail detector sampling rate to compensate. The slight con-
`
`touring effect due to run-length restriction is less noticeable
`than the streaky nature of the pictures without restriction.
`The slight flicker effect, due to the difference between
`successive fields, also disappears. These results suggest that
`the detail detector has greatest difficulty in handling long
`runs (confinned by a band~splitting system suggested by
`Vieri [3]).
`Typical variable-velocity pictures are shown in Fig. 7. The
`noisy structure in these photographs is due to two limita-
`tions. both unconnected with the variable-velocity principle.
`
`(3) Original picture.
`
`
`
`
`
`(c) Detail function. I(l, 2, 4. I0).
`
`(3) Original picture.
`Note.‘ All these photographs are single-frame (U25 second); n=3 in all recon-
`stmctions. In all cases, the encoder store was restricted to a capacity of 30 samples.
`Fig. 3.
`
`Page 6 ofll
`
`

`
`362
`
`PROCEEDINGS OF THE IEEE
`
`MARCH
`
`The analog-to-digital converter at the transmitter was only
`capable of handling five bits, so that 32 level signals were
`used in this prototype system (including synchronization);
`however, a seven-bit analog-to-digital converter is a modest
`requirement by present-day standards. Also, the phosphor
`structure of the display CRT was the limiting resolution
`factor in the small variable-velocity display.
`XI. CONCLUSIONS
`
`The results generally confirm the proposals for a basic
`bandwidth compression system made by Cherry et al. [1].
`The analysis in Section VI together with the experimental
`results using our very limited elastic encoder store of 30
`samples, suggests that an elastic encoder with l00—l 50-stage
`capacity would be adequate for most purposes. It is essential
`to minimize "overload," even at the expense of an average
`deterioration in picture quality over large regions, if neces-
`sary. It is suggested that the best use is made of the elastic
`encoder if a variable line-scan period is allowed. If the
`elastic encoder is made unnecessarily large, then the cor-
`responding possible line-period variations will require the
`use of a ratchet time base.
`
`It can be seen, by comparison with Cherry et al. [1], that
`run-length restriction has a beneficial subjective effect when
`used in conjunction with the present detail detector. The
`variable-velocity receiver results suggest that a larger dis-
`play and a seven-bit quality amplitude signal could give
`broadcast quality pictures. In particular, the positional
`accuracy of the display should be quite acceptable.
`The final data reduction ratio is obtained by multiplying
`the sampling ratio by 7/9 in the case of seven-bit brightness
`and two-bit run-length information. The compression ratio
`is the same as the data ratio in the case of binary PCM sys-
`tems, but otherwise depends upon the proposed fonn of
`modulation.
`
`The major problems associated with a complete system
`have been examined in this paper, except for the effect of
`channel noise on the run-coded signal. The latter has been
`investigated in relation to this system by Pearson [9], using a.
`quantitative assessment of the subjective efiects of position
`errors due to noisy channels in run-coded television signals.
`Further,
`the restriction of run lengths to three or four
`quantized values renders the position signal very noise-
`resistant.
`It appears that the development of variable-
`velocity receivers will possibly result in cheaper systems
`than could be obtained with elastic decoders. The variable-
`velocity time base, control, and brightness compensation
`circuits are relatively cheap and simple. although wide band-
`width defiection circuits, or some alternative, must be pro-
`vided for large display systems. Improved variable—velocity
`receivers should result from the use of more accurate bright-
`ness compensation as outlined in Section VIII, together
`with better forms of interpolation. (Note that “box-car"
`interpolation is used in this system, whereas linear inter-
`polation obtained by joining up the tops of the received
`amplitude-modulated samples with straight lines is particu-
`larly simple at the receiver, because here the samples arrive
`at a regular rate.)
`
`Future work on detail detection is indicated by the effect
`of run-length restriction. In particular, a simple improve-
`ment might be achieved if the average picture brightness of
`each run is used as the run-coded sample, instead of the
`first sample as in this system. Further work on systems em-
`ploying detail detection in the vertical as well as horizontal
`direction is proceeding, and should yield further compres-
`sions. A combination of the noise-combating procedure
`used here, together with the tapered quantization suggested
`by Graham [ll], would appear to form a useful extension
`of detail detection.
`
`The choice of the specific values of restricted run lengths
`is not critical and similar quality and compression values
`are obtained for a range of run lengths. However,
`it is
`obvious that once a particular run-length set has been
`chosen, the maximum possible compression is immediately
`bounded to a value somewhat below the maximum per-
`mitted run length, regardless of the setting of the detail
`detector thresholds. Thus, for the pen-nitted set used here,
`reasonable quality pictures are possible but compression
`ratios above five or six are unrealistic.
`Further work could include an extension of the restricted
`
`run-length values, especially for use with black-and-white
`picture material where larger compressions are practical.
`The authors regard the implementation and perhaps ex-
`tension of the feedback proposals made here to be essential
`for efficient usage of elastic encoders.
`APPENDIX
`
`QUEUE STORAGE (ROBINSON [8])
`
`Budrikis [7] and others have shown that run—coded sig-
`nals have exponential probability distributions and that
`successive run lengths along a television line are inde-
`pendent. In other words.
`the run-length samples occur
`randomly in time. It is useful to consider a run—coded televi-
`sion signal as a random sequence exhibiting “short-term"
`stationarity, i.e., as a random sequence in which the mean
`rate undergoes “long-term" variations. Here, the equilib-
`rium behavior of a finite capacity storage device which
`takes account of necessary “underload" and "overloa “
`facilities is considered.
`
`Consider a queue of maximum size M, its input a train of
`independent pulses occurring at the Nyquist rate 1/ T, each
`with a probability of occurrence p, 0 <p<l (i.e., a random
`sequence). Let the queue empty by one pulse at every nth
`Nyquist interval, giving a uniform output rate of l,!nT.
`(Note that n, the sampling ratio, is an integer.) In practical
`stores it is necessary to stagger the possible input and output
`instants.
`
`The queue can best be analyzed by considering the in-
`stants of time just prior to each output pulse.
`An “underload" situation occurs whenever the queue is
`empty just prior to an output pulse. If no action is taken.
`then the scheme of run-length coding will break down. This
`is prevented by inserting an extra pulse into the queue
`whenever “underload" occurs.
`
`An “overload" situation arises whenever the queue is of
`length M. in which case any input pulses which occur at this
`
`Page 7 ofll
`
`

`
`1967
`
`ROBINSON AND CHERRY: TV BANDWIDTH COMPRESSION SCHEME
`
`363
`
`101
`.
`
`.I
`
`.
`.
`
`,
`
`L
`
`10°
`.
`'
`I
`
`-110
`.
`.
`
`:
`
`‘
`10.:
`.
`
`I
`
`to’)
`
`lo”.
`
`tr’-
`
`1a'5
`
`n.95
`
`—- (ma) msrvrcnrrusm (-:1.-ax-J
`1.00
`1.n2
`141;
`1.06
`1.03
`
`0.56
`
`1.1a
`
`"
`
`—
`_
`..
`
`..
`39,39
`r 09°‘
`
`C
`"
`
`C
`_
`
`_
`
`i"
`
`no.3!)
`3.30
`
`n
`
`3.5::
`
`man)
`
`anmmmn ltmo I. I
`
`(n,.)
`
`cuxnrAa.u¢-rats
`sun-Lnvo mm In
`mum
`turn it
`
`Fig. 9.
`
`)0
`
`as
`
`26
`
`23;
`
`22
`
`-0: sun 01 cm: x
`13
`16
`1!;
`12
`so
`6
`
`20
`
`6
`
`t.
`
`2
`
`time are not accepted. but are lost to the system.
`We define:
`
`= p;'T.
`queue input data rate
`= 1;'nT
`queue output data rate
`queue data rejection rate (due to overload) = L,!nT
`queue data insertion rate (due to underload) = IIn T.
`
`Also we define (by normalizing with respect to the output
`data rate):
`
`p/T—l/nT
`= (np — 1}
`proportional overdrive rate = U" T
`traflic intensity [from telephone theory)=np
`
`.
`.
`.
`proportional rejection rate
`rt" nal insertion rate
`propo 10
`
`UHT
`= U"T = L
`— “HT -1
`— “HT — .
`
`Let P, be the probability that the queue is of length 3:,
`called the queue state. and consider
`rt!
`x!(n — x)!px-(1— pr”.
`
`B(x)EB(x:p,n)=
`
`Then the equilibrium conditions are obtained as solutions
`to the following difference equations :
`
`3(0) -P2 =[1~{B(o)+ B(1)}]P1
`B(o>P,+. = [1 — Bu)]P,, — B(2)P,-. -
`— BiX)Pt
`
`— B(x — 1)P2
`
`forx=2,3."',n.
`
`B(0)P..+t = [1 - B(1)]Px - B(2)Px-1 -
`forx=(n +1),'--,(M — l).
`
`- 3(H}P..—..+1
`
`P, = 1.
`
`M Z
`
`x=l
`
`and
`
`Also
`
`n-I
`
`L= E a{P,.,B(a +1)+ P,.,_,B(a + 2) +
`u=l
`
`+ PM‘+a+1-n'B{n}}
`
`and
`
`I = P,B(0).
`
`The solutions to these equations are plotted for typical
`parameters of interest in Figs. 9 and 10.
`The overload performance of the queue in equilibrium
`suggests that the most significant parameter is the traflic
`intensity up. When the traffic intensity is greater than unity
`the overload is largely independent of queue capacity M
`and of the sampling ratio n. When the traffic intensity is less
`than unity, the overload is generally less than 1 percent (at
`least for M > 30). The queue stare, i.e., the P, distribution,
`is plotted in Fig. 10 for a capacity of M = 30.
`
`Page 8 ofll
`
`

`
`364
`
`PROCEEDINGS or ‘rue IEEE. VOL. 55. NO. 3. MARCH. i967
`
`ACKNOWLEDGMENT
`
`The authors wish to thank W. A. Gurnhill for his patient
`and skillful construction of much of the equipment.
`
`REFERENCE
`[l] C. Cherry. M. H. Kubba, D. E. Pearson. and M. P. Barton. “An ex.-
`perimental study of the possible bandwidth compression of visual
`image signals,“ Proc. IEEE. vol. 51, pp. l507—l5l7. Novemb