PREDICTIVEDIFFERENTIAL PULSE-CODEMODULATIONAPPARATUS
"Digital Transmission of Information"
The present invention relates to the digital transmission and reception of information. It is convenient to describe the present invention in relation to its use for television signals but the invention is not limited to such a use as the techniques and apparatus to be described are of general application and could be used for audio as well as video signals.
Digital apparatus has already made an impact on broadcast systems, but usually as specific parts of what is still basically an analogue network. However, there are signs that a digital network will eventually emerge.
When transmitted in its simplest form, one form of serial digital television signals require a high bit rate, of the order of 106Mbit/s for an 8 bit pulse code modulation (PCM) system, with sampling locked to e.g. 3 times the PAL sub-carrier frequency. With the addition of error protection incorporated in a ninth bit, this would be suitable for use on a
120Mbit/s digital transmission system. However, a new digital hierarchy for European use has been proposed. One of the bit rates in this hierarchy is 140Mbit/s and it seems reasonable that two television signals should be transmitted in this channel in order to make
full use of the bandwidth. Each television signal would need to be coded at a bit rate somewhat less than 70Mbit/s in order to do this. Further reduction to approkimately 60Mbit/s would allow possible transmission on the proposed European satellite. These bit rates can be cahieved by using differential pulse code modulation (DPCM) in plate of PCM.
When a television signal is sampled at a frequency greater than the colour sub-carrier frequency, e.g. at three times the sub-carrier frequency, the magnitude of the difference from one sample to the next is generally small compared to the magnitude of the original samples; this results from the correlation which exists between adjacent parts of a television picture. This is particularly true for monochrome or low saturation colour signals. This fact can be used to predict the required sample from the immediately preceding sample. However, for colour signals of high saturation where there is a large amplitude of sub-carrier, the sample one cycle of sub-carrier earlier, in this case the third previous sample, will be closer in magnitude, as the sub-carrier would then be sampled at the same relative phase angle. By predicting the expected level of the sample and subtracting the expected, or predicted level from the actual level of the sample, a differential signal will be produced which will be normally much, smaller than the absolute values of the samples. By utilising this fact a reduction in the overall amount of information to be transmitted can be achieved. A further reduction can be made by improving the accuracy of the predictor, using not just one sample, but at least two previous samples which may be on the same line to give a measure of both luminance and chrominance components. The present invention
provides digital transmision apparatus comprising; means for receiving a digital signal comprising a series of data samples; means for predicting data samples from said received digital signal; means for deriving a differential signal indicative of the difference between each predicted data sample and the received data saumple corresponding thereto;
characterised in that said predicting means is arranged to temporarily store an indication of a plurality of preceding received data samples and to predict each data sample from the stored indication of preceding data samples; and in that quantizing means is provided for quantizing each said differential signal into a quantized signal having a lesser number of bits than said differential signal, said quantized signal being arranged to be transmitted as an indication of the received data sample corresponding thereto.
In the embodiments to be described, a prediction is based on a plurality of previous samples which, in the case of a television transmission system,may be in the same line. A predictor working on this basis is hereinafter called a complex predictor. The number of previous samples used depends on the sampling frequency and it is considered that 6 is the practical upper limit. The embodiments to be described use 5 previous samples. As an alternative, or in addition to using at least two previous samples from the same lines for prediction, additional samples from the previous two lines and/or from the previous two fields may be used. As mentioned above, the difference signals,
in this case in the form of digital words, are generally small in magnitude i.e. the occurrence of large magnitudes is relatively infrequent. Fortunately the eye is less sensitive to errors adjacent to large transitions so the embodiments of the present invention transmit large differences with less accuracy than small differences, thus reducing the bit rate; this may be achieved by using a non-linear quantizing law when producing the differential signals for transmission. Obviously, this introduces errors in the output signal to be transmitted so these must be allowed for, otherwise they will be propagated for the remainder of the line of picture information transmitted until the system is reset. Identical predictors are incorporate at both the transmitter and received in the embodimen and the operation of the DPCM system is such as to allow an inherent correction of the longer term effects of the non-linear quantizing law to be transmitted as part of the DPCM signal.
In order that the present invention may be more readily understood, embodiments thereof will now be described, by way of example, with reference to the accompanying drawings in which:- Figure 1 shows a block diagram of a transmitter and receiver;
Figures 2, 3 and 4 shows diagrams for assisting understanding of the operation of a part of the apparatus shown in Figure 1; Figure 5 shows one arrangement for use as the transmitter in Figure 1;
Figure 6 shows a diagram for assisting under standing of the operation of a part of the apparatus in Figure 5; and Figure 7 shows another arrangement for use as the transmitter in Figure 1.
Referring to the drawings, and particularly to Figure 1, DPCM apparatus is shown to comprise an analogue to digital converter 10 which samples the input video signal at a rate greater than the video sub-carrier frequency (fsc). In this case it is preferred to use a sampling frequency of three times the video sub-carrier frequency. The video sub-carrier frequency is typically 4.43361875 MHZ. The output from the converter 10 is a number of samples, each in the form of a digital word, of, typically, 8 bits which is fed as an input signal S to a quantizing and predicting circuit, indicated generally by a reference numeral 11. The output of the circuit
11 is a differential signal indicative of the difference between the sample word and its predicted value. In the circuit 11, a predicted value P of each sample, derived from a predictor 14, is subtracted from the input signal S to produceanother 8 bit word. If the prediction system used is efficient then most of the time these differences will be small. However, it is necessary to allow for occasional large differences which can occur. In order to minimise the bit rate required, a non-linear quantizer
12 is provided. The output of the quantizer 12 is directly proportional to the input for small differences but gets progressively more non-linear for larger differences so that it introduces a quantizing error E. This will be explained in more detail later. The resultant differential signal S-P+E is coded in a coder 13 and then transmitted.
At a receiver 20, the differential signal is decoded using a decoder 21 and the predicted value of the sample derived from a predictor 22 is added. The resulting digital signal is passed through a digital-to-analogue converter 24 and a low pass filter
25 to provide the video signal, which ignoring transmission and quantizing errors, is the signal S. The predictor 22 is identical to the predictor 14 used in the transmitter. To ensure that the receiver does not produce a D.C. offset compared with the transmitter, it is often necessary to reset the predictors 14 and 22 periodically, and it. is convenient to do this each line during the line blanking interval. If the gain of the predicted path is less then unity resetting may be unnecessary.
The quantizing and predicting circuit 11 will now be described in more detail. Dealing firstly with the quantizing part of the circuit, basically the difference signal (S-P) is quantized into a number of levels and preferably the quantization is performed in a non-linear manner depending on the magnitude of the difference signal. In this case 32 levels have been used and each level can be represented by a 5 hit code word. This 5-bit code word is then converted to another 8-bit word, but this will not necessarily be identical to the input 8-bit word due to a number of adjacent input 8-bit words being assigned the same 5-bit code word; hence an error E may be introduced by the quantizing circuit. This other 8-bit word will hereinafter be termed a quantized 8-bit word and is used as the input signal to the predictor l4. In order to simplify the quantizer and reduce the number of errors, it is proposed to utilize modulo 256 arithmetic with all negative numbers having 256 added to them, in which case all will become positive, any carries generated being ignored. This allows for transitions up to ±255
without increasing the non-linearity of the system. Providing the difference between the input 8-bit signal S and the predicted value P is within ± 255, the output of the quantizer 12 will be correct. Furthermore, the input signal will be sampled in such a manner that the bottom few levels and top few levels are not assigned to an input signal so that gross quantizing errors will be avoided. Figures 2, 3 and 4 are useful in understanding the operation of the quantizer 12 with Fig. 2 showing a non-linear quantizer law for prediction on the basis of a plurality of previous samples and using modulo 256 arithmetic and 32 levels, Figure 3 showing a signal sampled at a number of points, and Figure 4 being a table exemplifying the operation of the quantizer. The figures in brackets shown in the column "S-P" of Figure 4 relate to the actual values S-P before conversion into modulo 256 form.
It has been found that the quantizer 12 can take the form of two programmable read only memories, one for putting the 8-bit input signals into 32 levels, this being called a compressor, and one for converting the 5-bit, 32 level signals into quantized 8-bit signals, this being called an expander. In order to reduce the bandwidth required for transmitting the DPCM signals, it is proposed to transmit the 5-bit, 32 level signals and provide each of the receivers with an expander in addition to a predictor. Such a system will now be described in relation to Figure 5 which also illustrates one form of the predictor 14.
Referring now to Figure 5, it will be seen that the input 8-bit word Si has subtracted from it in the adder 51 its predicted value Po, and the resulting 8-bit differential signal is fed to a compressor 52 which assigns the 8-bit differential signal to one
of 32 levels and produces a 5-bit code word. This code word is fed both to a transmission system where it is encoded and transmitted, and to an expander 53 which. converts the 5-bit code word to a quantized 8-bit word. To the quantized 8-bit word is added the predicted value of the input 8-bit word and this is fed to an output as a video output signal So which can be used for monitoring purposes. lt can therefore be seen that the compressor 52 and expander 53 possess mutually reciprocal non-linear characteristics, the "expanded" signal at the output of the expander having substantially the form of the differential signal input to the compressor. The video output signal or summed signal So is also fed to a predictor circuit 55where it is used together with the summed signals corresponding to other preceding samples to provide a predicted signal for the next sample. In this embodiment, the preceding samples are derived from the same line. The predictor 55 utilizes 5 previous samples and each of these samples is multiplied by a co-efficient K according to a prediction law so as to produce at the output the desired predicted value being aweighted mean of the previous samples. In order to step the 8-bit words through the predictor 55, a number of latches 56 are provided which operate at the sampling frequency.
The derivation of the co-efficients K1 to K5 will now be explained in relation to Figure 6. The proportion of each of the 5 previous samples which are required to obtain the optimum prediction depends on the characteristics of each of the samples.
Consider the five samples shown in Figure 6. These are to be used to predict the next sample, the
predicted value being called 'P' (Note: The sampling frequency is exactly 3 times the sub-carrier frequency therefore the phase of the chrominance component of sample 'A' is the same as that of sample 'D'). Now let Ac , Bc, Cc, Dc and Ec be the chrominance components of samples A,B,C,D,E, respectively and let AL,BL,CL,DL, and EL be the luminance components of samples A.B.C.D.E. respectively. The predicted sample P is most closely related to the luminance component of sample A and the chrominance component of sample C.
This is the basic equation for the complex predictor and gives the coefficient values for K1 to
K5 of:
n practice, the values of are difficult to
implement in a digital system so they will be modifie slightly; but it can be shown that the adjustment of the ⅓ value will not significantly reduce the overall performance of the predictor. It is proposed to use either ¼ or ½ in place of the ⅓ values as these are suitable for a practical design since co-efficients of ¼ and ½ can be readily handled in a digital system. It has been found that ½ is the better replacement. An alternative approach is to say that, as a first step, we will use the third previous sample i.e. C as apredictor for the next sample.
This is a good predictor for chrominance but not very good for luminance (A would be much better).
This may be improved by putting in a correction for the anticipated difference between
P and C. The information for next sample has not arrived at the receiver so this cannot be used. However, the difference between P and C (i.e. P - C) represents a gradient. If this gradient is constant then A - D will be the same as P - C .
A better predictor is therefore C + (A - D) since this would work perfectly for a uniform gradient.
However gradients are not necessarily uniform and a more accurate forecast for (X - C) could be based on the assumption that the gradient
P - C will differ from the gradient A - D by the same amount as gradient A - D differs from B - E
(P - C) - (A - D) = (A - D) - (B - E)
(P - C) =2(A - D) - (B - E)
This perfectly allows for not only a uniform gradient (slope) but also for a uniform rate of chang of slope.
lt is therefore an alternative basis for a predictor:-
P = C + 2 (A - D) - (B - E) Although this would be an admirable predictor for noise free signals providing the slope and rate of change of slope were constant, it would be upset by random noise due to the coefficient of 2 for A and D. Additionally, actual pictures have less correlation and this decreases the optimum value of the coefficients.
We use P = C + (A - D) - ½ (B - E) It is considered that the operation of the circuit shown in Figure 5 will be clear without a detailed description. Suffice to say that the immediately preceding sample is multiplied by co-efficient K1 and added to the second preceding sample which is multiplied by co-efficient K2 etc. A problem with this embodiment is that the predicted value (Po) must be calculated in less than 75n Secs, so that the next difference signal (Si-Po) can be generated. A saving in time can be achieved by combining the compressor and the expander in a single programmable read only memory (P.R.O.M.) which thus has eight address inputs and eight outputs. Even using a single P.R.O.M. the operations which must be completed in this time are three additions and the non-linear quantisation. The three additions comprise that performed in the adder 51, the addition of the "expanded" signal from the expander 53 with the predicted signal Po, and the final addition in the predictor 55 which adds the immediately preceding sample weighted by the coefficient K1 to the other weighted preceding samples. The fastest convenient PROM available at the time of construction has an address to output time delay of
typically 25n Secs. In practice, 30n Secs has to be allowed. This only leaves 45n Secs for all three addition operations, as well as any latching that is required. With emitter coupled logic (E.C.L.) this might be possible but the P.R.O.M. is T.T.L. so there would be extra delay in conversion between E.C.L. and T.T.L, and vice versa. Further it is more expensive to build in E.C.L. and the power dissipation is higher. When one uses a combined compressor and expander it is necessary to separately produce the 5-bit code word for transmission. This is best achieved by connecting a further compressor, with identical characteristics to that of the P.R.O.M., so as to receive the difference signals from the adder. Alternatively, the further compressor may be connected to receive the output from the P.R.O.M. but this is usually less desirable. In the first case, the further compressor, when considered in conjunction with the expander at the receiver, must have the same law as the compressor and expander in the P.R.O.M.
If therefore it is desired to build a DPCM system utilizing a complex prodictor and using Schottky T.T.L. care must be taken to allow time to complete the necessary addition and latching operations in 45n Secs.
The embodiment shown in Figure 7 is suitable for this purpose and it will be seen that this embodiment, though generally similar to that of Fig. 5, utilizes two predictors and a further feedback loop from the output of the quantizer to its input.
The predictor 71 is identical to the predictor 55 and has the same co-efficients. However, the predictor 72 differs from it in detail
and also its co-efficients are different. It has been found that suitable values for the co-efficients KA to KE are related to the values for the co-efficients of the predictor 71 as follows:- KA = K1 + K2 KB = K2 + K3 KC = K3 + K4 KD = K4 + K5 KE = K5
Using the values of co-efficients K1 to K5 given above, the values for KA to KE are ½,½,0,-½,½. The embodiment provides a circuit whose timing is not critical since it has no more than 2 additions and one latch delay in 75n Secs (or 1 addition, the PROM delay and a latch delay in 75n Secs). This timing is well within the capabilities of Schottky type TTL and in some parts ordinary TTL.
An important feature of this circuit is the provision of the two feedback loops around the quantizer. The first loop is a complex loop comprising the two predictors and the second loop comprises only a latch. The second loop involves only the immediately preceding sample and there is only 1 addition operation at the input to the compressor and the PROM delay which can be achieved within the required time of 75n Secs. The first loop involves the second and other preceding samples and has a loop time of 150n Secs.
It can be seen that the latch at the output which retains the 8-bit output signal So, effectively delays this output by the timing of one sample. Since the predictors 71 and 72 receive this delayed output signal, an extra period of time is made
available, but the predictors are working on one sample later than the P.R.O.M. compressor/expander arrangement. The predictors therefore are responsive to the second preceding and earlier samples; the second feedback loop comprising the latch involves the immediately preceding sample and the minimal delay involved ensures the required time of 75n. Secs.
It will be appreciated that for both theembodiments of both Figs. 5 and Fig. 7 a receiver will include an expander and predictor (and additional feedback loop) identical to those shown for the transmitter so that the receiver can reconstuct the signal from the differential signal transmitted in the mam ar shown in Fig. 1.
The above arrangement using a complexpredictor produces a better overall prediction of the required sample than a system using one previous sample be it either the immediately or third previous sample. The complex predictor, has virtually no difference values in excess of 70. whereas the third previous sample predictor is significantly worse. This means that the complex predictor will have less edge business for a given number of bits per sample, In transient and frequency response tests, the complex predictor, with its modified coefficients, performs nearly as well as the previous and third previous sample predictors respectively. When the predictor was used in DPCM system the PROM laws were adjusted to give what appeared to be an optimum system - this gave subjectively higher quality pictures than pictures using an immediately or third previous sample prediction transmitted at 5 bits per sample. The complex predictor could be operated using 4½ bits or 4 bits for the quantizing levels for a slight quality impairment.
The use of DPCM enables the bit rate to be reduced from 106Mbit/s to less than 70Mbit/s. This is achieved using 5 bits per word ar a sampling frequency of three times the PAL sub-carrier frequency; further reduction can be achieved by reducing the number of bits per sample to 4½ and even 4 giving bit rates of 6θMbit/s and 53.3M/bit/s respectively.
A further advantage of these embodiments is that they could be modified to work effectively with an NTSC colour signal.
The embodiments described above make use of a number of immediately preceding samples in order to predict the sample being received. However, by providing suitable gating means to the predictor (s) it could be arranged for a number of preceding samples on the same line to be selected on a basis of e.g. every second or third sample received.
In addition, or as an alternative, samples
2 from preceding lines could be used. A suitable a shirt register containing data samplesfroma number of eg/ arrangement would be to provide/preceding lines.
The samples could then be serially output in a suitable fashion and fed to the predictor(s). If the shift register were to store the summed output signals So as samples, these could be applied directly as inputs to the predictor(s). The register may be arranged to output the sample which is positioned directly above the sample being predicted and/or to either side thereof. This arrangement can be readily implemented under the NTSC system, but precautions would need to be taken with a PAL system due to the phase reversal of the sub-carrier between adjacent lines.
As a further alternative, or in addition to some or all of the above embodiments, a prediction
may be based on preceding fields. Such a prediction should be very accurate due to the slow rate of change, on average, of detail within a television picture. A suitable random access memory (R.A.M.) means would need to be provided to store samples from one or more previous fields. One arrangement is to provide two R.A.M.'s for storing alternative fields; in a 625 line system, one may be arranged to store samples from 312 lines, and the other to store sample from 313 lines. Both of the R.A.M.'s could then add their components to the prediction on the basis of alternate fields. It is belived that the memories should each store an integral number of lines, hence the division into 312 and 313 lines. Each arrangement must allow for any discrepancies between the phase of the subcarrier of the samples chosen if an accurate prediction is to be made.