GB2238444A

GB2238444A - Digital video coder quantization control

Info

Publication number: GB2238444A
Application number: GB9019903A
Authority: GB
Inventors: Andrew David Raine Cotton
Original assignee: British Broadcasting Corp
Current assignee: British Broadcasting Corp
Priority date: 1989-09-21
Filing date: 1990-09-12
Publication date: 1991-05-29
Anticipated expiration: 2010-09-12
Also published as: GB2238444B; GB8921319D0; GB9019903D0

Abstract

In a digital coder of the type in which a signal is formed and subject to a transform, for example, a discrete cosine transform 12, the signal is applied to a quantizer 14 with variable coarseness of quantization. The quantization coarseness is controlled by control circuit 32 in response to the occupancy of a buffer 26 following the quantizer so as to prevent buffer overflow or underflow. The control circuit 32 is responsive to the buffer occupancy to generate a quantizing step size control signal which varies relatively slowly with buffer occupancy over a central range of buffer occupancies and varies relatively rapidly with buffer occupancy in top and bottom ranges respectively above and below the central range. <IMAGE>

Description

Digital Order The present invention relates to a digital coders, for example, coders of the type wherein an error signal is formed from the difference between an input signal and a predicted signal and is, possibly after application of a transform, applied to a quantizer with variable coarseness of quantisation which is controlled by the occupancy of a buffer following the quantiser and optional variable length coder so as to prevent buffer overflow or underflow. The invention may be applied, for example, to a DPCM (differential pulse code modulation) coder or to a video coder employing the DCT (discrete cosine transform).

The invention is concerned in particular with improving the way in which the quantiser coarseness is adapted to the buffer occupancy. The proposed scheme makes good use of the buffer size and can offer an improvement in the overall signal-to-noise ratio of up to 1.5 dB when compared with codec simulations employing more primitive schemes.

A currently proposed DCT video codec incorporates a variable length coder (VLC), so each block of pixels is coded in an irregular number of bits. As the codec output is at a fixed bit-rate a buffer, in the form of a FIFO, is required to smooth out these variations. To prevent the buffer from over- or underflowing, the average buffer input bit rate must match the fixed buffer output bit rate. To ensure that this is the case the coarseness of coefficient quantisation must be adapted in accordance with the buffer occupancy.

The quantising scheme simulated at the BBC is similar to that outlined in the CMTT/2 DCT codec specification (Digital transmission of component-coded television signals at 30-40 Mbit/s and 45 Mbit/s using Discrete Cosine Transforms. CMTT/2 DCT Group chairman report, document CMTT/2-66, July 1988). Here the coarseness of the quantiser is determined in dependence on a quantity known as the transmission factor, TXF. This is defined to lie in the range -24 to +39. A value of -24 represents a very fine quantiser, whilst a value of +39 represents a coarse quantiser. In fact, we have found that +39 is not sufficiently coarse to code some sequences, so our simulations allow it to reach a maxImum value of +63.The quantiser step size may vary linearly in accordance with TXF for example, or in accordance with 2(+A) /B, where A may vary according to coefficient position and B may be, for example, 8.

The simulest law relating buffer occupancy to transmission factor is a linear law where the transmission factor is set to -24 when the buffer is empty, an occupancy of 0.0, and set to +63 when the buffer is full, an occupancy of 1.0. All intermediate levels are defined by a straight line joining these two points. A similar linear law is shown in Figure 1, and may be used for reference control purposes. In actual fact the end points do not represent occupancies of 0.0 and 1.0, but 0.02 and 0.9. This leaves a safety margin to prevent buffer overflow and underflow. The upper safety margin needs to be much larger than the lower safety margin as the buffer is capable of being filled at a far greater rate than it can empty.

In an ideal codec the quantiser coarseness, and hence transmission factor, would remain constant over as long a period as possible, the buffer smoothing out local variations in bit rate.

This could be obtained by using a very large buffer store. A large, localised, increase in the output bit-rate would have little effect on the buffer occupancy and hence the transmission factor. In practice, because of factors of codec cost, equipment size and restricting the signal delay through the codec, the upper limit on the buffer size has been set at one frame at the coded bit-rate.

With a simple occupancy/quantiser law, such as the one described, this can lead to quite dramatic changes in the quantiser coarseness across a frame. This is demonstrated by a specially constructed test sequence, in which the top fifty and the bottom fifty lines of the sequence have been replaced by Gaussian noise at 0 dB relative to peak video, an addition often used as a video effect. This sequence represents one containing both difficult and easy areas to code within a single frame. Figure 2 shows how the buffer occupancy and transmission factor vary whilst coding this sequence with the simple law of Figure 1. The peak-to-peak change in transmission factor within a frame is about 15, representing a change in quantising step of about 4:1. When viewing the coded sequence this is quite visible.The top portion of the picture insert, directly after the noise, is markedly noisier than the lower portion.

The object of the present invention is to overcome these problems. While the invention is defined with particularity in the appended claims, the basic concept lies in replacing the linear law of Fig. 1 with a law composed of a shallow central portion, such that quantising step size changes only slowly with buffer occupancy, and two or more steeper end portions where the quantising step size changes significantly more abruptly. In an important development of the invention the central portion is shifted up and down adaptively in dependence upon an integral function of buffer occupancy.

The advantage of the invention is demonstrated in that, for the same test sequence referred to above, we are able to reduce the change in transmission factor for any one frame from 15 to 3. This reduces the higher quantising errors and allows the buffer to absorb the local variations in bit rate more effectively.

The invention will be described in more detail, by way of example, with reference to the accompanying drawings, in which: Fig. 1 shows a known law relating TXF to buffer occupancy, described above, Fig. 2 shows the fluctuations in buffer occupancy BO and transmission factor TXF over the test sequence of twenty frames, also described above, Fig. 3 shows one example of an improved law relating TXF to BO, Fig. 4 shows an adaptive version of the improved law, Fig. 5 is a similar drawing to Fig. 2 but showing the improvement obtained using the law of Fig. 4 Fig. 6 shows a simplified version of the adaptive law of Fig.

4, and Fig. 7 shows a DCT coder embodying invention.

The basic shape of the improved quantiser/buffer law is shown in Figure 3. This particular law is split into three zones. The gradient of the middle zone is relatively shallow, allowing large changes in buffer occupancy with little change in transmission factor and hence quantiser coarseness. The lower and upper zones are fall-back zones where, should the buffer over- or underfill, the transmission factor changes rapidly to correct the situation. The gradients of the law where the different zones meet are matched to ensure a smooth transition, i. e. there are no first order discontinuities.

It will be seen that although TXF varies from -24 to 63 it only varies by 5 between BO = 0.1 and BO = 0.7. Thus appreciably less than 10% of the variation of TXF takes place over the linear central range.

A law of this nature would not be universally sufficient on its own because of the large variation in picture characteristics. The mean transmission factor needed to code a sequence at 17 Mbit/s can vary by as much as 30 or 40 from one sequence to another. This is evident even from a limited range of test material. The mean transmission factor should lie on the shallow part of the law: with a single law for all sequences this is clearly not possible. For many sequences the coder would be constantly operating in one of the fall-back zones, yielding very large variations in transmission factor for only small changes in buffer occupancy. To overcome this problem the control law can be made adaptive.

The buffer occupancy is constantly monitored and its difference from a reference occupancy, the occupancy error, is integrated. This integral term is used to shift the transmission factor law up and down to suit the current picture characteristics.

This ensures that the law is operating in its central shallow zone as often as possible. This is demonstrated in Figure 4.

Because the buffer can always fill faster than it can empty we have chosen the ideal buffer occupancy, and hence the reference occupancy from which the occupancy error is measured, as 0.3, rather less than a half. The integral term is updated once every few line periods or at other convenient intervals when the transmission factor and buffer/quantiser law are recalculated.

An example of the computer code to calculate the new law is quite simple and is given in the Appendix. The gradient of the middle region and the integral scaling factor or time constant have been carefully chosen to maintain stability consistent with a fast response time. As with any computer control system, it is important that the integral term is limited, or desaturated, to prevent it reaching very high values. This might happen when there is a sudden change in picture content from an easy scene to a difficult scene to code, causing a prolonged buffer occupancy error. If the integral were allowed to reach very high values it would lead to unacceptably long settling times and excessive occupancy error overshoot. In our calculation we limit the integral term to +30.0 allowing for a mean transmission factor in the range of +17 to +47.

Figure 5 shows how the test sequence referred to above codes with the new transmission factor law. The quantiser coarseness remains almost constant throughout the frame. The improvements are clearly visible on the coded sequence. The coding impairments in the sequence with the new law are far less severe and spread at a lower level across the frame. The new law offers an improvement in measured signal-to-noise ratio of 1.25 dB compared with the original simple law. With another sequence, an improvement of 1.5 dB was recorded.

Using such a characteristic it should be possible to reduce the buffer size to one television field at the coded bit rate and still offer an improvement in performance over a simpler buffer strategy using a one frame buffer store.

Figure 6 shows a simplified law in which three straight-line segments are used with the break points therebetween shifted up and down in dependence upon the integral term. Such a law can be derived very simply and may contain more than three segments. This description concludes with an Appendex expressing derivation thereof in mathematical terms while a second Appendix gives a Fortran listing implementing the more complex law of Fig. 4.

Figure 7 shows, by way of example, a DCT predictive coder embodying the invention. The coder is based upon a known predictive coder loop comprising a subtractor 10 which subtracts a prediction signal P from an input video signal V to form a difference signal D.

This is transformed to (V-P)* by a DCT transform unit 12 of known form operating on a block or strip of 8 x 1 pixels (from one video line) or a block of 8 x 8 pixels in eight adjacent lines, to give two of many possible examples. (V-P)* is a set of coefficients in the spatial frequency domain.

(V-P)* is fed to a quantiser 14 whose output is (V-P+E)* where E represents the quantizing error. This output is the information which is transmitted to the decoder, via a buffer 26.

Block 18 operates inversely on (V-P+E)* to form V-P+E to which P is added in an adder 20 to provide V+E. V+E is stored in a frame store 22 whose output is the basic prediction signal P' which may be used directly as P. The prediction signal may optionally be processed to P by a known motion compensator 24 responsive to a motion vector M which may be provided pixel by pixel or on a pixel block basis.

The quantiser 14 is of known form effecting variable length coding and moreover has a variable quantising step size controlled by the signal TXF, also as is known per se. As well as the quantised coefficients (V-P+E)* the quantiser outputs bit count pulses, one for every bit in the quantised coefficents. These are applied to the UP input of an UP/DoWN counter 28 associated with the buffer 26. The quantised coefficients are clocked out of the buffer 26 at a constant output bit rate, in a well known manner, and clock pulses at this rate are applied to the DCWN input of the counter 28.

The contents of the counter thus represent the buffer occupancy BO.

The buffer occupancy reference value BOR (e.g. 0.3 as suggested above) is subtracted from BO in a subtractor 30 to produce a difference or error value EV. This value EV is employed by a control means 32 which generates TXF therefrom. No further details of the control means 32 are necessary since they merely implement a simple algorithm of which two examples are given in the two appendices.

The subtractor 30 and control means 32 operate once per stripe of pixels transformed by the DCT 12, i.e. once per set of output VLC coefficients generated by the quantiser 14, so as to generate the value of TXF to be used for quantising the next set.

The invention is not restricted to the particular type of coder selected for Fig. 7. For example, transforms other than DCT may be used and the coder need not be of this variety at all but, for example, a DPCN coder operating pixel by pixel rather than a pixel blocks. The coded data is not necessarily video data. The invention is applicable to any kind of coder using a quantiser with variable length coding, a buffer to smooth the output data rate and variable step size control of the quantiser dependent upon the buffer occupancy.

Appendix (1) For BOL < BO <

Where ml, c and k are constants.

For BO < BOL TXF - m0BO + where m0 is given by:

For BO > BOu TXE - m3BO + d where m3 is given by:

and d is given by: d = TXFmax - m3 Appendix (2) The Fortran code listing to calculate the new quantiser/occupancy law subroutine calctxf(occupancy,txf) implicit none integer txf(2) Ithe Y and W transmission factors integer tend Itxf at end of middle range real occupancy Ithe buffer occupancy (0.0 to 1.0) real integral lintegral of occupancy error real a,b,c Iparameters for quadratic real x Ivariable in quadratic for lower range c first find new error integral integral-integral+(occupancy-0.3) c implement integral desaturation integral-min(integral,30.0) integral-max(integral,-30.0) c calculate txf. 3 laws dependant on occupancy if (occupancy.lt.0.1) then Ein lower range tend-nint(30.833+0.5*integral) a=-3645.833-156.25*tend x-0.1-occupancy txf(1)=nint(a*(x**2)-8.333*x+tend) else if (occupancy.gt.0.7) then !in upper range tend-nint(35.833+0.5*integral) a-1533.333-25.*tend b-35.*tend-2138.333 c-745.5-11.25*tend txf(l)-nint(a*(occupancy**2)+b*occupancy+c) else Sin middle txf(l)-nint(8.333*occupancy+30.0+0.5*integral) endif c check bounds txf(l)-max(txf(l) , -24) i tx f(l)-min(txf(l),63) txf(2)-txf(l) !chrominance txf - luminance txf c end of subroutine return end

Claims

1. A digital coder of the type wherein a signal is formed and is, possibly after application of a transform, applied to a quantizer with variable coarseness of quantisation which is controlled by the occupancy of a buffer following the quantiser so as to prevent buffer overflow or underflow, comprising control means responsive to the buffer occupancy to generate a quantising step size control signal which varies relatively slowly with buffer occupancy over a central range of buffer occupancies and varies relatively rapidly with buffer occupancy in top and bottom ranges respectively above and below the said central range.

2. A digital coder according to claim 1 in which the quantiser output is subject to the operation of a variable length coder.

3. A digital coder according to claims 1 and 2, wherein the control signal varies linearly with buffer occupancy in the central range.

4. A digital coder according to claim 3, wherein the control signal also varies linearly with buffer occupancy in the top and bottom ranges but with greater slopes.

5. A digital coder according to claim 3, wherein the control signal varies non-linearly with buffer occupancy in the top and bottom ranges so as to merge smoothly into the variation in the central range.

6. A digital coder according to any of claims 1 to 5, wherein the control means are further responsive to an integral term of buffer occupancy to shift the central range bodily up and down the control signal ordinate so as to tend to maintain operation of the control means in the said control range.

7. A digital coder according to claim 6 in which the top and bottom ranges are recalculated when the central range is shifted so as to avoid discontinuities.

8. A digital coder according to claim 7, wherein the integral term is formed from the difference between buffer occupancy and a reference buffer occpancy.

9. A digital coder according to claim 1, wherein the relationship of control signal to buffer occupancy is substantially in accordance with Fig. 3, Fig. 4 or Fig. 6 and of the accopanying drawings.

10. A digital coder according to claim 1 and substantially as hereinbefore described with reference to and as illustrated in Fig. 7 of the accompanying drawings.