GB2347039A

GB2347039A - Estimating encoding parameters

Info

Publication number: GB2347039A
Application number: GB9903801A
Authority: GB
Inventors: Alois Martin Bock; Charles Baudouin
Original assignee: NDS Ltd; Tandberg Television AS
Current assignee: Synamedia Ltd; Ericsson Television AS
Priority date: 1999-02-19
Filing date: 1999-02-19
Publication date: 2000-08-23
Also published as: GB9903801D0

Abstract

A method/apparatus for controlling the encoding process of an encoder comprises decoding the encoded signal and estimating from the decoded signal, using a statistical process, the parameters used to encoded it. These parameters may relate to the type of coding(Intra or inter), the quantisation levels used, motion vectors and mode decisions and the estimation is done. The estimation circuit accepts an input signal 18 to a synchronisation detector 21 and a horizontal pixel operator 22. The absolute magnitude of the output of the operator 22 is determined by module 23. An accumulator 24 accumulates the absolute magnitude of the operator 22 over a picture period such as a picture frame or picture field with the end of such a period being detected by the synchronisation detector 21. The output of the accumulator is provided as an input to one or more modules 25, each of which comprise a number of stores 26(1) to 26(n). The output of each store is input to a decision logic module 27 which provides a signal related to the picture coding type parameter of the input signal 18.

Description

METHOD AND APPARATUS FOR ENCODING DATA This invention relates to encoding of data, and is particularly suited to successive encoding and decoding of data.

Successive encoding and decoding of data is common within data compression systems. One example of such systems is the compression of digital television data. Broadcasters may wish to perform additional processing of programme material or transmit data at a different bitrate to that previously used to compress the data. In order to be able to meet such requirements the programme material is often reencoded. Re-encoding of data often results in a degradation of the integrity of subsequently encoded data. The degree of degradation increases with successive (or cascade) encoding and decoding.

With the growth of digital television networks, cascade encoding and decoding over several generations becomes inevitable. Recently, a number of proposals have been published to improve the cascade performance of video and audio encoding systems: [1]'Real time transcoding of MPEG-2 video streams', P. N. Tudor, O. H.

Wemer, IBC'97.

[2]'Flexible switching and editing of MPEG-2 video bit streams', P. J.

Brightwell, S. J. Dancer, M. J. Knee, IBC'97.

[3]'Proposed SMPTE Standard for Television-MPEG-2 Re-Coding Data Set', SMPTE draft proposal PT20.022-1842B, 20.10.1998.

[4]'Proposed SMPTE Standard for Television-Compressed Stream Format of the MPEG-2 Re-Coding Data Set, SMPTE Draft Proposal PT20.02/021-1843B, 20.10.1998.

The common feature of these proposas is the requirement to convey previous encoding parameters to down-stream encoders through a side channel which is carried in addition to the video or audio signal. In particular, the use of the two least significant chrominance bits has been proposed as a means of carrying re-coding information (4). While such systems work well when the integrity of the side channel can be guaranteed, in practice this is not always the case and is sometimes difficult or impossible to achieve.

A second disadvantage of the prior art is the data capacity required for the side channel. In fact, for the full re-coding data set of MPEG-2 video [3] a side-channel capacity of approximately 20 MBit/s is required.

There is a need to be able to provide an improved method and apparatus for cascade encoding/decoding.

According to the first aspect of the invention there is provided a method of controlling the encoding process of an encoder comprising : estimating an encoding parameter from a signal that has been encoded and subsequently decoded; and encoding the signal in the encoder in accordance with the encoding parameter.

According to a second aspect of the invention there is provided apparatus for controlling the encoding process of an encoder comprising: means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded; and means for encoding the signal in the encoder in accordance with the encoding parameter.

The invention will now be described by way of example only, and with reference to the following figures, in which: Figure 1 is a block diagram of a cascade encoding/decoding system for successive encoding and decoding of broadcast television data; Figure 2 is a block diagram of a first embodiment of an estimator in accordance with this invention; Figure 3 is a block diagram of an encoder comprising a second embodiment of an estimator in accordance with this invention; Figure 4 is a histogram of energy levels of a television video signal having a GOP structure of 9; Figure 5 is a graph of averaged energy correlation for varying levels of quantisation as measured by probability density function (pdf), blockiness and a combination of pdf and blockiness ; Figure 6 is a graph showing the standard deviation of pdf measured averaged energy correlation ; Figure 7 is a graph showing the standard deviation of blockiness measured averaged energy correlation ; Figure 8 is a graph showing the standard deviation of a combination of pdf and blockiness measured averaged energy correlation; and, Figure 9 is a block diagram of an estimator for estimating the frame start point comprised within an audio television signal.

Figure 1 is a block diagram of a cascade encoding/decoding system for a television signal. The system comprises two encoder/decoder pairs 1 and 2. Pair 1 comprises an up-stream encoder 3 and a decoder 4, and similarly encoder/decoder pair 2 comprises down-stream encoder and decoder 5 and 6 respectively. Pairs 1 and 2 may form part of a larger cascade encoding/decoding system.

Encoder 3 receives an input television signal 7. The signal is encoded to produce a bit-rate reduced data stream 8 representative of the input television signal 7. The data stream 8 usually conforms to a standardised encoding format known as MPEG (Moving Pictures Expert Group). Decoder 4 decodes the data stream output of encoder 3 to provide an output decoded signal 9. Encoding often introduces degradation into the resultant signal. It is preferable that the subsequent encoding by encoder 5 is carried out in substantially the same manner as that of encoder 3. This reduces the amount of degradation introduced during cascade encoding/decoding.

There are various encoding parameters associated with standardised formats of data compression. Examples of video encoding parameters are the picture coding type [ie, whether a picture is intra-coded (I), predictively encoded from previous frames (P) or bi-directionally predicted from past and future frames (B)], quantisation levels, motion vectors, mode decisions etc,. The importance of these parameters in terms of their effect on the picture quality of the decoded picture in cascade encoding/decoding stages varies greatly. For example, the picture coding type of a previously encoded video signal is significantly more important than all other video compression parameters put together. This information can be carried in less than 50 bit/s Figure 2 is a block diagram of an estimator for estimating picture coding type parameter (s). The estimator has an input video signal 18 such as that provided as an output from decoder 4 of Figure 1. The input signal is fed to a synchronisation detector 21 and a horizontal pixel operator 22. The absolute magnitude of the output of the operator 22 [defined in terms of, for example, energy] is determined by a module 23. An accumulator 24 accumulates the absolute magnitude of the operator 22 over a picture period such as a picture frame or picture field. The end of each picture period is detected by the synchronisation detector 21.

The output of accumulator 24 is provided as an input to one or more modules 25.

The, or each, module 25 comprises a number of stores 26 (1) to 26 (n). The output of each store of the, or each, module 25 is provided as an input to a decision logic module 27. The module 27 provides an output signal 19 that is related to the picture coding type parameter of the input signal 18. This output is then available to enable a down-stream encoder to subsequently encode a decoded signal in substantially the same manner as that used by the up-stream encoder.

The estimator of Figure 2 uses measurements of certain statistical properties of the input signal 18 which are affected by up-stream encoding in a known manner. Video compression algorithms use temporal prediction, DCT transformation followed by quantisation and run-length coding as the main means of reducing the data rate of the input datastream. In order to achieve the best picture quality for a given video sequence, the quantisation which is applied to any particular picture depends on its coding type. Typically, bi-directionally predicted pictures (B) are more heavily quantised than intra-coded (I) or forward-predicted (P) ones. Furthermore, the quantisation matrix used in (I) pictures is normally different to that in (P) and (B) pictures. The differences in the coding of the three picture types lead to measurable statistical differences in the decoded video signal. These statistical differences can be detected in the down-stream signal.

The pattern of (I), (P) and (B) pictures is detected by the estimator of Figure 2 in the following way. The horizontal pixel operator 22 is a one-dimensional mathematical function applied in the horizontal direction of a video signal. This function may be one of the following : high-pass filter, e. g. as implemented by Finite Impulse Response Filter (FIR) =(0.5,-0.5) band-pass filter, e. g. FIR= (-0.25,0,0.5,0,-0.25) median filter one-dimensional DCT coefficient, i. e. ci = 0.125 ; 47 f (x) cos [(2x+1) in/16].

The number of modules 25, and also the stores contained therein, comprised with an estimator are dependent upon predetermined assumption (s) conceming the periodicity of the picture coding type sequence. Where a group of pictures (GOP) structure of 12 pictures is assumed, the output of accumulator 24 is provided to a module 25 having 12 stores. The output of accumulator 24 may be provided to a number of modules 25, each comprising a different number of stores 26. By running several modules in parallel different GOP structures may be detected. Figure 2 shows a module 25 having an assumed GOP structure of 3.

Where, for example, a GOP structure of 12 pictures is assumed the output of the accumulator is averaged in 12 different stores, with each successive picture contributing to one of the 12 stores in tum. Each time a new value is added to one of the stores, i. e. at the end of each picture, a sliding average is calculated over a large number of pictures. Each sliding average is provided as an input to the logic module 27. Where the input signal to the estimator had been previously encoded using the hypothesised GOP structure of 12, then a pattern will emerge across the 12 stores. The logic module 27 detects the presence or absence of such a pattern within the, or each, module 25. The output of module 27 may be indicative of the GOP structure detected, or altematively may be limited to designation of the presence or absence of a pattern relating to a particular GOP structure.

In practice it is preferable to test for the particular GOP structure for which the downstream encoder is configured. If a downstream encoder finds its input signal has been previously compressed with the same GOP structure it is configure for, then it is desirable to time-shift the GOP structure to achieve alignment. If, however, the input signal to the downstream decoder had been previously encoded with a different GOP structure it would not be appropriate to re-configure the downstream encoder. This is because the previous GOP structure might not be appropriate for the down-stream encoder. The downstream encoder therefore encodes the input signal in accordance with the estimated GOP structure encoding parameter by ignoring the estimated GOP structure and encoding according to its own GOP structure configuration.

Figure 3 is a block diagram of an alternative embodiment of the present invention. It shows the estimator of Figure 2 and further comprises a scene cut detector 30. The estimator of Figure 3 is shown as comprised within a downstream encoder 32. The downstream encoder also comprises compression engine 31. The output of the encoder 32 is a synchronised encoded video signal.

As described above, where an assumed GOP structure of a decoded signal matches the actual structure of the estimator input signal, then over a period of time a pattern will emerge. This period is longer for video broadcast material having numerous scene changes. The scene cut detector 30 may be used to reset the sliding averages within the, or each, module 25. This enables faster detection by the logic module 27 of any pattern that may emerge.

Figure 4 is a diagram of the energy levels of 9 frames assuming a GOP structure of 9. It can be seen that the energy levels of (I) and (P) frames are between energy boundaries 41 and 42. For some down-stream encoding purposes it may be possible to ignore the statistical difference between (I) and (P) frames. This can be implemented by designating any store 26 having a sliding average within particular boundaries as being either an (I) or (P) frame. All other stores are then designated as (B) frames.

From Figure 4 it can be determined that the GOP structure is: IBBPBBPBBIBBP However, where the statistical difference between (I) and (P) frames is ignored, such a sequence may be detected as: P B B I B B P B B P B B I Or PBBPBBIBBPBBP Where the difference in energy levels between (I) and (P) frames is within the boundaries 41 and 42, the differences between the above sequences is quite small. This means that there is little difference in alignment and so no time shifting for alignment is required.

The second most important video encoding parameter for downstream encoding after picture coding type is that of quantisation. The most appropriate pixel operators to measure previously applied quantisation parameters are one-dimensional horizontal DCT. Unlike picture coding type, quantisation is rarely applied uniformly over an entire picture. Reducing the measurement areas to slice or macroblock sizes, however, leads to statistically unreliable results. Nevertheless, an encoder can get statistical information about previously applied quantisation levels from a picture-based analysis.

A direct measurement of quantisation can be carried out as follows : 1. Calculate the one-dimensional horizontal DCT transform; 2. Derive histograms of all DCT coefficients, one for each coefficient, over the entire measurement area; 3. If the signal (within the measurement area) had been previous compressed with a particular quantisation value, the histograms will show distinct local maxima; 4. From the pattern of local maxima across all histograms the quantisation value can be derived.

Other methods can also be used to estimate quantisation levels. For some applications it may be useful to determine upper and lower bounds of quantisation parameters.

Examination of the noise floor probability density function (pdf), such as is described in W098/10595 reveals that MPEG quantisation results in a gradual distortion of the originally smooth pdf with distinctive peaks and troughs occurring at regular intervals. In particular, the step from a noise level of 3 (corresponding to an average absolute difference of 1.0) to a noise level of 4 (average absolute difference = 1.33) becomes bigger as the relative effect of MPEG quantisation increases. The transition between analogue and digital noise occurs at a quantisation level of approximately 5 (new qscale~type) for sequences with low source noise. However, as MPEG quantisation increases still further, the ratio pdf (3)/pdf (4) saturates and is thus less sensitive at quantisation levels of more than 13.

At such levels of quantisation a new method can be used, based on a measurement of blockiness. Blockiness is measured as the ratio of absolute differences between pairs of adjacent inter-block pixels as compared to intra-block pixels. This method is somewhat complementary to the method described above in so far as it mainly measures heavy MPEG distortion, but is less sensitive at lower quantisation levels. The best overall measure of MPEG noise (highest correlation to MPEG quantisation) is, therefore, taken as a weighted sum of the two measured values. Figure 5 shows a comparison of each of the two methods together with the weighted sum of both. A mix of 0.35 pdf + 0.65 blocky gives the highest correlation to quantisation parameter (qu).

Figure 5 gives a comparison of the mean values of the two methods (plus a mixture of the two) as measured over 31 test sequences. However, it does not show the spread of the measured values. The spread of values is important if upper and lower bounds for qp should be derived from the measured quantisation noise.

Figures 6,7 and 8 show the same graphs, this time with standard deviation. It can be seen that the best estimate for qp is obtained by a mixture of the two methods.

The disadvantage of the blockiness measure as defined above is that it is dependent on knowledge of the size and position of coded blocks. For this reason a more general method has been developed which is largely independent of block size and is completely independent of block position. Instead of taking certain positions to be known block boundaries all pixel locations are treated equally. Blockiness is then defined as the ratio of the largest difference over the average difference within a given pixel range. The pixel range is conveniently chosen as a multiple of upsampled block sizes.

Although the above descriptions associated with Figures 2 to 8 are made with reference to encoding parameters for television video data, it will be understood that this invention can be used for estimating encoding parameters for other data such as television audio data.

Fig 9 is a block diagram of an estimator for estimating the frame start point in an audio signal. A memory element 100 stores the audio data associated with one frame, and a second memory element 12 stores another frame of delayed audio data. A cross-correlator 104 receives the output of both frame memories 100 and 102. The cross-correlator 104 cross-correlates the outputs of frame memories 100 and 102 and the result is input to a decision logic module 106 for analysis. The data in the frame memories is refreshed and the cross-correlated results are again input to the decision logic module 106. The decision logic module 106 attempts to estimate the position of a maximum value for each value at the input. This enables the decision logic to estimate the starting point of the frame as it was during the original encoding and decoding process.

Claims

CLAIMS 1 A method of controlling the encoding process of an encoder comprising: estimating an encoding parameter from a signal that has been encoded and subsequently decoded; and, encoding the signal in the encoder in accordance with the encoding parameter.
2 A method according to Claim 1 wherein the step of estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises estimating an encoding parameter of a television video signal.
3 A method according to Claims 1 or 2 wherein the step of estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises measuring a statistical property of the signal.
4 A method according to Claims 2 or 3 wherein the step of estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises estimating a picture coding type parameter.
5 A method according to Claim 4 wherein the step of estimating a picture coding type parameter comprises measuring energy comprised within a picture period.
6 A method according to Claim 5 wherein the step of measuring energy comprised within a picture period comprises detecting presence of a pattern according to a pre-determined GOP structure.
7 A method according to Claims 4 or 5 wherein the step of measuring energy comprised within a picture period comprises detecting whether the measured energy lies within upper and lower bounds.
8 A method according to Claims 2 or 3 wherein the step of estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises estimating a quantisation parameter.
9 A method according to Claim 8 wherein the step of estimating a quantisation parameter comprises measuring the probability density function (pdf) of the signal.
10 A method according to Claim 8 wherein the step of estimating a quantisation parameter comprises measuring the blockiness of the signal.
11 A method according to Claim 8 wherein the step of estimating a quantisation parameter comprises measuring a combination of the pdf and the blockiness of the signal.
12 A method according to Claim 1 wherein the step of estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises estimating an encoding parameter of a television audio signal.
13 A method according to Claim 12 wherein the step of estimating an encoding parameter of a television audio signal comprises estimating a frame start point within the signal.
14 Apparatus for controlling the encoding process of an encoder comprising: means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded; and, means for encoding the signal in the encoder in accordance with the encoding parameter.
15 Apparatus according to Claim 14 wherein the means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises means for estimating an encoding parameter of a television video signal.
16 Apparatus according to Claims 14 or 15 wherein the means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises means for measuring a statistical property of the signal.
17 Apparatus according to Claims 15 or 16 wherein the means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises means for estimating a picture coding type parameter.
18 Apparatus according to Claim 17 wherein the means for estimating a picture coding type parameter comprises means for measuring energy comprised within a picture period.
19 Apparatus according to Claim 18 wherein the means for measuring energy comprised within a picture period comprises means for detecting presence of a pattern according to a pre-determined GOP structure.
20 Apparatus according to Claims 18 or 19 wherein the means for measuring energy comprised within a picture period comprises means for detecting whether the measured energy lies between upper and lower bounds.
21 Apparatus according to Claims 15 or 16 wherein the means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises means for estimating a quantisation parameter.
22 Apparatus according to Claim 21 wherein the means for estimating a quantisation parameter comprises means for measuring a probability density function (pdf) of the signal.
23 Apparatus according to Claim 21 wherein the means for estimating a quantisation parameter comprises means for measuring blockiness of the signal.
24 Apparatus according to Claim 21 wherein the means for estimating a quantisation parameter comprises means for measuring a combination of the pdf and the blockiness of the signal.
25 Apparatus according to Claim 1 wherein the means for estimating an encoding parameter from a signal that has been encoded and subsequently decoded comprises means for estimating an encoding parameter of a television audio signal.
26 Apparatus according to Claim 25 wherein the means for estimating an encoding parameter of a television audio signal comprises means for estimating a frame start point within the signal.
27 An encoder comprising apparatus according to any of Claims 14 to 26.
28 A method as substantially hereinbefore described and with reference to Figures 2to9.
29 Apparatus as hereinbefore described and with reference to Figures 2 to 9.