WO1988006390A1

WO1988006390A1 - Coder

Info

Publication number: WO1988006390A1
Application number: PCT/GB1988/000087
Authority: WO
Inventors: Graham R. Seabrook; David I. Crawford
Original assignee: British Telecommunications Public Limited Company
Priority date: 1987-02-13
Filing date: 1988-02-11
Publication date: 1988-08-25
Also published as: GB8703354D0

Abstract

A coder comprising a differential pulse code modulation loop (1) including a quantizer (4), followed by a variable length coder VLC (5) and a buffer (8). The VLC (5) assigns words from a codebook to the different output levels from the quantizer (4), and the buffer (8) smooths the bit rate of the signals received from the VLC prior to transmission over a fixed rate channel (9). When the buffer approaches the overflow condition it signals the quantizer (4) to cause a reduction in the number of quantization thresholds. This causes only a subset of the codebook words to be assigned, and a reduction in the bit rate if a codebook is used which is sub-optimized in normal operation but gives a better overflow performance.

Description

CODER

Television signals contain considerable spatial and temporal redundancy due to the high correlation between adjacent pels - a fact that is exploited in predictive coding schemes such as Differential Pulse Code Modulation (DPCM). However, the output signal levels produced by DPCM have a highly peaked probability distribution which indicates high statistical redundancy.

The most common technique employed for the removal of this statistical redundancy is Variable-Length Coding (VLC) often using codes derived by Huffman's algorithm [Huffman D., "A Method for the Construction of Minimum Redundancy Codes," Proc. IRE, pp 1098-1101, September 1952]. VLC attempts to allocate shorter length codewords to high probability levels and longer length codewords to low probability levels: the VL Codebook is thus dependent on the statistics of the input signal. It can be shown [Jayant N.S. and Noll P., "Digital Coding of Waveforms", Prentice and Hall, 1984] that the minimum entropy achievable with a VLC/uniform quantizer combination is lower than that of a VLC/non-unifor quantizer combination. However, this is the case only if. the VL Codebook is perfectly matched to the statistics of the input signal - clearly a difficult constraint to fulfill if the signal is non-stationary.

A perfectly matched non-uniform quantizer aims to equalise the probability of all levels. In such a case it is obvious that, since all levels become equiprobable, the number of bits required by the respective VL Codewords (given by n_k = -log₂P_k, where n_k is the theoretical minimum number of bits required and P_k is the probability of the kth level) will be equal: application of VLC thus achieves zero reduction in bit-rate. However, the ideal situation is in practice never reached and in a non-ideal case a combination of VLC/non-uniform quantization yields an average bit-rate very close to the minimum as given by:

P_kn_k bits/sample

where N is the number of quantization levels.

Given that the use of non-uniform quantizers is justifiable in attempting to achieve the required channel rate it has been shown [Schafer R., "DPCM Coding of the chrominance Signals for the Transmission of Colour T.V. Signals at 34 Mbit/s", Signal Processing 6, pp 187-199, 1984] that the minimum number of levels required to produce subjectively acceptable coded pictures are 13,9, and 5 for the luminance (Y) and chrominance (U and V) Bands respectively. Since these numbers are not powers of 2 it is necessary to employ VLC to attain the minimum achievable average bit-rates.

Any VLC system must employ a buffer to smooth the variable data rate for transmission over a fixed rate channel. However any buffer will, almost inevitably, eventually exceed its limits; so some form of buffer control is imperative. The traditional way of preventing buffer overflow is to employ feedback from the buffer to control selection of a more, or less, coarse quantization characteristic: the system is thus adaptive. The main penalties incurred by an adaptive system are: (a) increased complexity, and hence cost,

(b) less tolerance to transmission errors; since overhead information regarding quantizer selection must be received uncorrupted.

According to the invention there is provided a coder comprising a differential coding loop including a quantizers variable length coder;a buffer for adapting the bit rate of the variable length coder output to the characteristics of a transmission channel; control means responsive to the condition of the buffer to control the switching of the quantizer between a first mode and a second mode, the quantizer and the variable length coder being; a) in the first mode responsive to the quantizer input signal lying within respective ranges defined by a first set of threshold values to generate respective codewords of a set of codewords; b) in the second mode responsive to the quantizer input signal lying within respective ranges defined by a second, smaller set of threshold values to generate respective codewords from a subset of the set of codewords; wherein; c) the subset is such that for each codeword of the subset, the value ranges giving rise to generation of that codeword in the two modes at least partially overlap; and d) the set and subset of codewords are so selected that the variable length coder output exhibits a lower average redundancy in the second mode than in the first, thereby effecting a bit rate reduction in the second mode.

Preferably, the thresholds employed by the quantizer in the second mode are a subset of those employed in the first mode and the quantizer outputs are correspondingly a subset of those used in the first mode. This advantageously simplifies the construction and operation of the quantizer.

Preferably, the codebook is structured so that one or more relatively shorter codewords are assigned to quantizer output levels which, in the first mode, have a relatively lower probability of recurrence than a level or levels to which relatively longer codewords are assigned; the said shorter codeword(s) being members of the subset.

Preferably, the second mode subset of codewords is an optimal subset for the second mode quantizer output levels, with an added bit to allow expansion of the set, and the remainder of the codebook is an optimal set for the first mode, non-second mode, quantizer output levels, each codeword of the remainder having a binary prefix to distinguish it from the subset.

In another apect, the invention comprises a transmission system employing such a coder and a decoder with a single, fixed codebook.

The invention will now be described by way of example with reference to the following drawings in which:

Figure 1 shows a coder according to an embodiment of the invention;

Figure 2 shows an example of a quantizer characteristic that may be used in an embodiment of the invention.

Referring now to figure 1 a coder is shown comprising a Differential Pulse Code Modulation (DPCM) loop 1 having an input terminal 2 connected to a quantizer 4 via a subtractor 3; the output of the quantizer 4 is connected both to a variable length coder 5, and to a feed-back predictor 6 via a second adder 7. The output of predictor 6 is connected to the inverting input of first adder 3 so as to complete a negative feed-back loop, and the output of the predictor 6 is also connected to a second input of the second adder 7. The output of variable length coder 5 is connected via data buffer 8 to a fixed rate transmission channel 9. The buffer 8 also has a feed-back connection 10 to the quantizer 4. In operation the coder processes the luminance (Y) component s(t) of a colour television signal in the DPCM loop 1, to produce a quantized innovation signal E_g(t). The quantizer 4 has a non-uniform quantization characteristic for example as indicated by a dotted line ll in figure 2. Columns 1 and 2 of Table 1, which follows below, indicate the quantized outputs obtained from inputs to a quantizer having this characteristic.

The quantized output levels (Table 1 column 2) are then fed to the variable length coder 5 where high probability levels are coded with shorter words and low probability levels are coded with longer words in order to produce an output data stream, of variable bit rate, having a low average bit rate. The buffer 8 then smooths the bit rate prior to transmission. when the buffer 8 approaches an overflow condition a signal is sent via a feed-back line 10 to the quantizer 4 to alter the quantizer characteristic 11. Strictly speaking the quantizer is therefore adaptive, however, the characteristic is not in fact changed to a different characteristic, but a simplified sub-set, (the dash line 12 in figure 2) of the existing characteristic 11 is brought into play. The simplified characteristic 12 uses only five of the original thirteen output levels, as indicated in Table 1 column 3, and makes use of (some of) the same quantization thresholds that the original characteristic used, namely those between input levels -14 and -13, -3 and -2, 2 and 3, and 13 and 14 (although it could use thresholds which differed slightly from those of the original without seriously impairing the operation of the invention). The variable length coder 5 sees the quantized output signals, in the "overflow" mode, no differently from the signals received during normal operation and uses the same codewords to code the same levels as previously. This means that the same codebook is used for the same quantizer output levels in the overflow mode as in normal operation although, in practice, only a sub-set of the codewords present in the codebook are actually utilised in the second mode.

In the receiver, codewords of both normal and overflow transmissions are decoded by the same codebook, and since it is not even necessary for the receiver to know which mode of operation the transmitter is in, no overhead information need be sent. The only effect on the decoded signal, when the coder is in overflow mode, is reduced resolution.

Although it is advantageous to employ in the second mode a set of quantizer output levels which is a subset of the set used in the first mode, and to arrange the variable length coder to respond in a fixed manner to these output levels, it would of course be possible to envisage other arrangements of the variable length coder and the quantizer in which in the second mode, codewords are assigned to quantizer output levels which correspond to input signals to which those codewords would have been assigned in the first mode; the essential thing is that the decoder at the receiver produces an output in response to a codeword which bears some resemblance to the input to the quantizer at the coder which produced that codeword. The average bit rate in overflow mode should be below that of normal operation, or else the buffer will not empty. Using a codebook produced by Huffman's algorithm that is "optimum" in the sense of minimum redundancy could lead to an increase in average bit rate when operating in the overflow mode. This is because, to a first approximation, in the overflow mode the system operates by adding the probabilities of occurrence of the missing levels to the closest of the remaining levels. Thus if the statistics of the coded picture are not altered then a reduction in average bit rate must ensue. However in practice restriction of the quantizer to a lower number of levels results in poorer predictions and thus increased errors: hence the probability of the remaining levels increases beyond the simple addition of the former probabilities. 'Thus the increase in average bit rate resulting from the increased probability of the outer levels may be come greater than the reduction in average bit rate resulting from the use of the overflow quantizer. It has been shown empirically that to avoid this condition when employing an optimum Huffman code one must employ only the inner three levels, which does not provide adequate picture quality. Therefore it is preferable to use a "sub-optimum" (non-minimum redundancy) codebook which, although resulting in a slight increase in bit rate with thirteen level quantization, provides a significant reduction with five levels, the picture quality of which is acceptable. such a "sub-optimum" codebook should preferably provide: an average bit rate for the majority of typical source material that is less than the effective video transmission bit rate (in bits per pel); as large a reduction in average bit rate as possible when operating in overflow mode; and synchronous code. For example if the available channel rate is 29 Mbit/s then the effective video transmission rate is given by:

transmission rate (Mbits/s) x line period fu sec) number of pels per line ie 29 x 64

(10.125 E6 X 52 E-6) + (3.375 E6 X 52 E-6)

= 2.644 bits per pel

when the number of pels per line is calculated assuming a 3:1:0 sampling structure, ie sampling rates of 10.125 MHz for the luminance and 3.375 MHz for the chrominance with alternate line transmissions of the U&V signals. The code is therefore required to have an average bit rate of less than 2.644 bits per pel during normal operation. The optimum Huffman codebook gives typical average bit rates of 2.5 bits per pel for luminance and 1.4 and 1.2 bits per pel for U&v respectively, giving an overall bit rate of 2.2 bits per pel, ie:

[3 X 2.5 + 1 X (1.4 + 1.2) / 2]/4 when 3 out of every 4 pels in the line are luminance pels and the last in every 4 is a chrominance pel being used in alternate lines for U or V values.

Using the optimum Huffman code, a typical bit rate reduction of less than 0.03 bits per pel is achieved in overflow mode (five levels). Clearly since the effective transmission bit rate is 2.644 bits per pel the code can be sub-optimised considerably in normal operation, so as to provide a large reduction in the average bit rate when in overflow mode (and hence a more rapid return to the normal buffer condition). This sub-optimisati:n is performed by allocating a relatively short codewcrd for a low probability level; simulations reveal that use of the +/- 28 levels produce least subjective degradation, and a codebook has been developed as shown in T= le 2 column 2. The codebook is preferably generated ir the following manner. First, the quantized levels produced in the overflow mode (Overflow 1 mode) are taken and ranted by magnitude (0, +/- 5, +/- 28). Then short codewords are assigned in order to these levels to form an 'optimal' Huffman codebook for Overflow 1 mode. In order to allow this codebook to be expanded, an extra bit is added t: the longest two codes (the +/- 28 codes), rendering the c debook slightly sub-optimal. Now the remaining normal mofe levels, not used in the Overflow 1 mode (+/- 10, +/- 17, +/- 40, +/- 58), are ranked by magnitude and codewords are assigned to them to form an optimal (Huffman) codebcok. These codes are then prefixed by a fixed prefix corresponding to the initial part of the expanded codewords from the Overflow 1 mode (the +/- 28 codes), with the added rit inverted to distinguish them from the overflow mode codes, and the two codebooks are combined, as shown in T=ile 3, to form the subopti al codebook.

The corresponding typical average luminance bit rates using this codebook are: normal mode 2.61 bits per pel; overflow mode (Table 2 column 3) 2.20 bits per pel. The typical average chrominance bit rates are: U-band 1.4 bits per pel and V-band 1.2 rits per pel thus the typical average overall bit rates ie for all three Y, U and V bands are: normal mode 2.286 bits per pel; overflow mode 1.975 bits per pel, a reduction in bit rate of 0.311 bits per pel. Advantageously a second, and even a third, overflow mode can be provided. The second overflow mode will come into play when the buffer approaches the overflow condition in the first overflow mode. In the second overflow mode the quantizer outputs only three levels as shown in Table 1 column 4 (shown as Overflow 2 mode'); the variable length coder in this case will only output three different codes as shown in Table 2 column 4. The third overflow mode ( 'Fallbac ' mode) which is brought into play when the buffer approaches overflow in second overflow mode is a fall-back position in which only one quantizer output level is used, as shown in Table 1 column 5, for all input signals; and in this case variable length coder outputs only one codeword one bit long (Table 2 column 5).

The corresponding typical luminance average bit rates in overflow mode 2 will provide 1.93 bits per pel; with the same typical average chrominance bit rates as before this will provide an average overall bit rate for all three bands of 1.776 bits per pel, a reduction in bit rate of 0.51 bits per pel.

The third preferred requirement of the sub-optimum codebook, that the output code is synchronous, requires that the total probability of obtaining a synchronising codeword must be as large as possible; there must therefore be as many synchronous codewords as possible each with as high probability of occurrence as possible. The codebook as shown in Table 2 is very synchronous - any codeword ending in the sequence 100 is a synchronising codeword: the total probability of obtaining synchronization (when coding typical source material) is thus 0.26, ie approximately every fourth codeword. Up to now the invention has been described with reference to luminance signals (Y), but clearly such a codebook could be used for the chrominance (U and V) bands as well, although use of Huffman codebooks for these bands provides such a high probability of obtaining zero error, especially when enhanced by a non-uniform quantization that the application of a codebook having overflow modes in the chrominance bands is of negligable benefit in terms of bit rate reduction. The increased hardware complexity/bit rate reduction trade-off falls in favour of a non-variable codebook.

In practice the buffer may operate line by line, switching the quantization characteristic at the end of any particular line in which overflow occurs. When coding typical source material a buffer of 100 K bits will usually be sufficient. However, if a particular line causes a buffer to overflow (iέ during an "busy" area) then since the number of subsequent frames are likely to have a very similar picture content (in particular the areas of busyness are likely to be spatially coincident) there is a high probability of overflow during contiguous fields. The area of low resolution thus become sub ectively apparent as a rolling bar.of low resolution lines. It may therefore be preferable to initiate switching on a field by a field basis - clearly a much larger buffer would then be required, in the order of 3 Mbits. However in hardware this can easily be implemented at minimal cost with DRAMs. In terms of subjective quality the low resolution fields are hardly discernible. Table 1

Luminance (Y)

Input Normal Overflow 1 Overflow 2 Fallback Output

1 2 3 4 5

51-255 58 28 28 0

35- 50 40 28 28 0

23- 34 28 28 28 0

14- 22 17 28 28 0

8- 13 10 5 28 0

3- 7 5 5 0 0

-2- 2 0 0 0 0

-7-{- 3) -5 -5 0 0

-13-(- 8) -10 -5 -28 0

-22-(-14) -17 -28 -28 0

-34-(-23) -28 -28 -28 0

-50-(-35) -40 -28 -28 0

-255-(-55) -58 -28 -28 0

Table 2

Quantization Codewords Overflow 1 Overflow 2 Fallback Level Normal

Operation

4 5

0 0 0 0 0 -5 11 11 +5 100 100 -10 101100 +10 101110 -17 101111 +17 1011011 -28 10100 10100 10100 +28 10101 10101 10100 -40 10110100 +40 101101011 -58 1011010100 +58 1011010101

Table 3.

Level Code Length

0 0 1 )

-5 11 Added 2 )

+5 100 /Bit 3 )) Overflow

+28 10100 5 ) Mode 1 ]

-28 1010 1 5 ) Codebook ]

+10 ion;oo 6 )) Normal

-10 1011J 10 6 ι Codebook

-17 ion; 11 6

-40 101l| 01011 9

+58 lOllJ 010100 10

-58 1011, 010101 10

/ \

Prefix Optimal Codebook

Claims

1. A coder comprising: a differential coding loop including a quantizer; a variable length coder; a buffer for adapting the bit rate of the variable length coder output to the characteristics of a transmission channel; control means responsive to the condition of the buffer to control the switching of the quantizer between a first mode and a second mode, the quantizer and the variable length coder being; a) in the first mode responsive to the quantizer input signal lying within respective ranges defined by a first set of threshold values to generate respective codewords of a set of codewords; b) in the second mode responsive to the quantizer input signal lying within respective ranges defined by a second, smaller set of threshold values to generate respective codewords from a subset of the set of codewords; wherein; c) the subset is such that, for each codeword of the subset, the value ranges giving rise to generation of that codeword in the two modes at least partially overlap; and d) the set and subset of codewords are so selected that the variable length coder output exhibits a lower average redundancy in the second mode than in the first, thereby effecting a bit rate reduction in the second mode.

2. A coder according to claim 1 in which the thresholds employed by the quantizer in the second mode are, or do not differ substantially from, a subset of those employed in the first mode and the output levels produced by the quantizer in the second mode are a subset of those produced in the first mode.

3. A coder according to claim 1 or claim 2, in which the set of codewords comprises; a) the said, first, subset of codewords generated in both quantizer modes comprising a sequence of short codes assigned in inverse order of length to the quantizer input signal ranges ranked by probability of recurrence in the second mode, calculated to form a near minimum redundancy codebook, and b) a second subset of codewords generated only in the quantizer first mode comprising a sequence of short suffix codes assigned in inverse order of length to those input signal ranges, other than those giving rise to generation of a codeword of the said first subset, ranked by probability of recurrence, so that the suffix codes are calculated to form a minimum redundancy codebook for such ranges, each prefixed by a prefix code so that codewords of the said second subset have a longer average length than those of said first subset.

4. A coder according to claim 3 in which each codeword of the said second subset is longer than the codewords of the said first subset.

5. A coder according to any one of claims 2 to 4 in which the quantizer has a third mode employing a set of quantization thresholds which is a subset of said subset of said set; whereby the codewords assigned by the coder when the quantizer is in the third mode are a subset of those which can be assigned when the quantizer is in the second mode, and the set of codewords is so selected that the variable length coder output exhibits a lower average redundancy in the third mode than in the second mode.

6. A coder as claimed in claims 1 to 5 in which the control means is responsive to the level of fullness of the buffer to control the switching of the quantizer from the first mode to the second mode when the buffer approaches overflow in the first mode.

7. A luminance band coder according to any preceding claim for coding the luminance band signals of a television signal.

8. A coder substantially as hereinbefore described with reference to the accompanying tables.

9. A coder substantially as hereinbefore described with reference to the accompanying figures.

10. A transmission system comprising a coder according to any preceding claim, and a decoder adapted to decode values representing the codewords assigned by the variable length coder, wherein the decoder employs only a single set of codewords so that a change in the mode of the quantizer at the coder need not be signalled to the decoder.