WO2003083859A2 - Watermark time scale searching - Google Patents

Abstract

Method and apparatus are described for compensating for a linear time scale change in a received signal, so as to correctly rescale the frame sequence of the received signal. Firstly, an initial estimate of the sequence of symbols is extracted from the received signal. Successive estimates of correctly time scaled sequences of the symbols are then generated by interpolating the values of the initial estimates.

Description

Watermark time scale searching

The present invention relates to apparatus and methods for decoding information that has been embedded in information signals, such as audio, video or data signals.

Watermarking of information signals is a technique for the transmission of additional data along with the information signal. For instance, watermarking techniques can be used to embed copyright and copy control information into audio signals.

The main requirement of a watermarking scheme is that it is not observable (i.e. in the case of an audio signal, it is inaudible) whilst being robust to attacks to remove the watermark from the signal (e.g. removing the watermark will damage the signal). It will be appreciated that the robustness of a watermark will normally be a trade off against the quality ofthe signal in which the watermark is embedded. For instance, if a watermark is strongly embedded into an audio signal (and is thus difficult to remove) then it is likely that the quality ofthe audio signal will be reduced. In digital devices, it is typically assumed that there exists up to a 1% drift in sampling (clock) frequency. During transmission ofthe signal through an analog channel, this drift is normally manifested as a stretch or shrink in the time domain signal (i.e. a linear time scale change). A watermark embedded in the time domain (e.g. in an audio signal) will be affected by this time stretch or shrink as well, which can make watermark detection very difficult or even impossible. Thus, in the implementation of a robust watermarking scheme, it is extremely important to find solutions to such time scale modifications.

In known time domain watermarking schemes, any linear time scale change within the signal is resolved by repeatedly running the watermark detection (including repeating the extraction ofthe watermark from the host signal) for the different possible time scales, until all the possible time scales are exhausted, or detection is achieved. Performing such searches over the possible time scaling ranges requires a large computational overhead, and is thus costly in terms of both hardware and computational time. Consequently, real time implementation of a watermark detector utilizing such a time scale search technique is not feasible. In watermarking schemes implemented within the frequency domains, it is common to perform the scale search by modifying the frequency domain coefficients. For instance, this can be achieved by carefully shrinking or stretching the frequency domain samples. In principle, such a frequency domain solution could be directly applied to time domain watermark signals. However, since the watermarks are directly embedded in the time domain samples, the time scale search needs to be performed in the time domain as well. Normally, there are only a few thousand frequency domain samples, whilst the time domain signals contain samples in the order of millions. Consequently, such an application of the frequency domain solution to time domain signals is computationally too expensive.

It is an object ofthe present invention to provide a watermark decoding scheme for time domain watermarked signals that utilizes a time scale search that substantially addresses at least one ofthe problems ofthe prior art.

In a first aspect, the present invention provides a method of compensating for a linear time scale change in a received signal, the signal being modified by a sequence of symbols in the time domain, the method comprising the steps of: (a) extracting an initial estimate ofthe sequence of symbols from said received signal;(b) forming an estimate of a correctly time scaled sequence ofthe symbols by interpolating the values of said initial estimate. Preferably, step (b) is repeated so as to provide a range of estimates corresponding to different time scalings.

Preferably, said interpolation is at least one of zeroth order interpolation, linear interpolation, quadratic interpolation and cubic interpolation.

Preferably, the method further comprises the step of processing each estimate as though it were the correctly time scaled sequence ofthe symbols, so as to determine which estimate is the best estimate.

Preferably, the method further comprises the steps of correlating each of said estimates with a reference corresponding to said sequence of symbols; and taking the estimate with the maximum correlation peak as the best estimate. Preferably, said initial estimate ofthe sequence of symbols is stored in a buffer. Preferably, said buffer is of total length M, the total

bf siskin ..searghjs/

M conducted is N_η = — (77_max - 77_min ) where η_mi_n, η_mzx correspond respectively to the mimmum

and maximum likely time scale modifications ofthe signal.

Preferably, said initial estimates ofthe sequence of symbols comprises a sequence of N_b estimates for each symbol, each ofthe N_b estimates corresponding to a different time offset of a symbol.

Preferably, the scale search in the next detection window is adapted based on the information acquired during the current detection window.

Preferably, the scale space is searched using an optimal searching algorithm. Preferably, the searching algorithm is the grid refinement algorithm.

In another aspect, the present invention provides a computer program arranged to perform the method as described above.

In further aspects, the present invention provides a record carrier comprising the computer program, and a method of making available for downloading the computer program.

In another aspect, the present invention provides an apparatus arranged to compensate for a linear time scale change in a received signal, the signal being modified by a sequence of symbols in the time domain, the apparatus comprising: an extractor arranged to extract an initial estimate ofthe sequence of symbols from said received signal; and an interpolator arranged to form an estimate of a correctly time scaled sequence ofthe symbols by interpolating the values of said initial estimate.

Preferably, the apparatus further comprises a buffer arranged to store one or more of said estimates.

In another aspect, the present invention provides a decoder comprising the apparatus as described above.

For a better understanding ofthe invention, and to show how embodiments of the same may be carried into effect, reference will now be made, by way of example, to the accompanying diagrammatic drawings in which:

Figure 1 is a diagram illustrating a watermark embedding apparatus; Figure 2 shows a signal portion extraction filter H;

Figures 3 a and 3b show respectively the typical amplitude and phase responses as a function of frequency ofthe filter Η shown in Fig. 2; Figure 4 shows the payload embedding and watermark conditioning stage of the apparatus shown in Fig. 1 ;

Figure 5 is a diagram illustrating the details ofthe watermark conditioning apparatus H_e of Fig. 4, including charts ofthe associated signals at each stage; Figure 6a and 6b show two preferred alternative window shaping functions s(n) in the form of respectively a raised cosine function and a bi-phase function;

Figures 7a and 7b show respectively the frequency spectra for a watermark sequence conditioned with a raised cosine and a bi-phase shaping window function;

Figure 8 is a diagram illustrating a watermark detector in accordance with an embodiment of the present invention;

Figure 9 diagrammatically shows the whitening filter H_w of Fig. 8, for use in conjunction with a raised cosine shaping window function;

Figure 10 diagrammatically shows the whitening filter H_wof Fig. 8, for use in conjunction with a bi-phase window shaping function; Figure 11 shows details ofthe watermark symbol extraction and buffering processes in accordance with an embodiment ofthe present invention;

Figure 12 illustrates a sequence in which estimates of watermark symbols are collected from four buffers when there is no time scale modification;

Figures 13a and 13b illustrate the different sequences, according to an embodiment ofthe present invention, in which estimates of watermark symbols can be collected from four buffers when there is respectively a time stretch and a time shrink time scale modification;

Figure 14 shows an example of an efficient scale search technique based on the concept of grid refinement; and Figure 15 shows a typical shape ofthe correlation function output from the correlator ofthe watermark detector shown in Fig. 8.

Fig. 1 shows a block diagram ofthe apparatus required to perform the digital signal processing for embedding a multi-bit payload watermark w into a host signal x. A host signal x is provided at an input 12 ofthe apparatus. The host signal x is passed in the direction of output 14 via the adder 22. However, a replica ofthe host signal x (input 8) is split off in the direction ofthe multiplier 18, for carrying the watermark information. The watermark signal w_c is obtained from the payload embedder and watermark conditioning apparatus 6, and derived from a reference finite length random sequence w_s input to the payload embedder and watermark conditioning apparatus. The multiplier 18 is utilized to calculate the product ofthe watermark signal w_c and the replica audio signal x. The resulting product, W_cX is then passed via a gain controller 24 to the adder 22. The gain controller 24 is used to amplify or attenuate the signal by a gain factor α.

The gain factor α controls the trade off between the audibility and the robustness ofthe watermark. It may be a constant, or variable in at least one of time, frequency and space. The apparatus in Fig. 1 shows that, when α is variable, it can be automatically adapted via a signal analyzing unit 26 based upon the properties ofthe host signal x. Preferably, the gain α is automatically adapted, so as to minimize the impact on the signal quality, according to a properly chosen perceptibility cost- function, such as a psycho- acoustic model ofthe human auditory system (HAS) in case of an audio signal. Such a model is, for instance, described in the paper by E.Zwicker, "Audio Engineering and Psychoacoustics: Matching signals to the final receiver, the Human Auditory System", Journal ofthe Audio Engineering Society, Vol. 39, pp. Vol.115-126, March 1991.

In the following, an audio watermark is utilized, by way of example only, to describe this embodiment ofthe present invention.

The resulting watermark audio signal y is then obtained at the output 14 ofthe embedding apparatus 10 by adding an appropriately scaled version of the product of w_c and x to the host signal: y[n] = x[n] + aw_c[n]x[n] . (1)

Preferably, the watermark w_c is chosen such that when multiplied with x, it predominantly modifies the short time envelope of x. Fig. 2 shows one preferred embodiment in which the input 8 to the multiplier

18 in Fig. 1 is obtained by filtering a replica ofthe host signal x using a filter H in the filtering unit 15. If the filter output is denoted by x_D, then according to this preferred embodiment, the watermark signal is generated by adding the product of xι and the watermark w_c to the host signal x: y[n] = x + cxw_c[n]x_b[n] . (2) Let x_b be defined such that x_b = x - x_b , and y_b be defined such that y = y_b + x_b , then the envelope modulated portion y_b ofthe watermarked signal y is given as y_b[n] = (\ + w_c[n])x_b[n] (3)

Preferably, as shown in Fig. 3, the filter H is a linear phase band pass filter characterized by its lower cut-off frequency.//, and upper cut-off frequency fy. As can be seen in Fig. 3b, the filter H has a linear phase response with respect to frequency/ within the pass- band (BW). Thus, when H is a band pass filter, x_b and x_b are the in-band and out-of-band components ofthe host signal respectively. For optimum performance, it is preferable that the signals x_b and x_b are in phase. This is achieved by appropriately compensating for the phase distortion produced by filter H. In the case of a linear phase filter, the distortion is a simple time delay.

In Fig. 4, the details ofthe payload embedder and watermark conditioning unit 6 is shown. In this unit, the initial reference random sequence w_s is converted into a multi-bit watermark signal w_c. Firstly a finite length, preferably zero mean and uniformly distributed random sequence w_s, from now on also referred to as the watermark seed signal, is generated using a random number generator with an initial seed S. As will be appreciated later, it is preferable that this initial seed S is known to both the embedder and the detector, such that a copy ofthe watermark signal can be generated at the detector for comparison purposes. This results in the sequence of length L_w w_s[k] e [-1,1], for k=0,l,2 L_w-1 (4)

It should be noted that in some applications, the seed can be transmitted to the detector via an alternate channel or can be derived from the received signal using some predetermined protocol. Then the sequence w_s is circularly shifted by the amounts dj and d using the circularly shifting unit 30 to obtain the random sequences W_JJ and w_d2 respectively. It will be appreciated that these two sequences (waj and vttø) are effectively a first sequence and a second sequence, with the second sequence being circularly shifted with respect to the first. Each sequence w*, i = 1 ,2, is subsequently multiplied with a respective sign bit r_h in the multiplying unit 40, where η = +1 or -1. The respective values of ri and r₂ remain constant, and only change when the payload ofthe watermark is changed. Each sequence is then converted into a periodic, slowly varying narrow-band signal w_t of length L_WT_S by the watermark conditioning circuit 20 shown in Fig. 4. Finally, the slowly varying narrow-band signals wj and w₂ are added with a relative delay T_r (where T_r<T_s) to give the multi-bit payload watermark signal w_c. This is achieved by first delaying the signal w₂ by the amount T_r using delaying unit 45 and subsequently by adding it to wj with the adding unit 50.

Fig. 5 shows the watermark conditioning apparatus 20 used in the payload embedder and watermark conditioning apparatus 6 in more detail. The watermark seed signal w_s is input to the conditioning apparatus 20.

For convenience, the modification of only one ofthe sequences w^,- is shown in Fig. 5, but it will be appreciated that each ofthe sequences is modified in a similar manner, with the results being added to obtain the watermark signal w_c.

As shown in Fig. 5, each watermark signal sequence W_di[k], i=l,2 is applied to the input of a sample repeater 180. Chart 181 illustrates one ofthe sequences w^ as a sequence of values of random numbers between +1 and -l_s with the sequence being of length L_w. The sample repeater repeats each value within the watermark seed signal sequence T_s times, so as to generate a rectangular pulse train signal. T_s is referred to as the watermark symbol period and represents the span ofthe watermark symbol in the audio signal. Chart 183 shows the results ofthe signal illustrated in chart 181 once it has passed through the sample repeater 180. A window shaping function s[n], such as a raised cosine window, is then applied to convert the rectangular pulse functions derived from w^ and vitø into slowly varying watermark sequence functions wjfnj and w₂[n] respectively.

Chart 184 shows a typical raised cosine window shaping function, which is also of span T_s. The generated watermark sequences wj[n] and w₂[n] are then added up with a relative delay T_r (where T_r<T_s) to give the multi-bit payload watermark signal w_c[n] i.e.,

w_c[n] = w,[n] + w₂[n -T_r] (5)

The value of T_r is chosen such that the zero crossings of w_t match the maximum amphtude points of w and vice-versa. Thus, for a raised cosine window shaping function T_r=TJ2, and for a bi-phase window shaping function T_r=T 4. For other window shaping functions, other values of T_r are possible. As will be appreciated by the below description, during detection the correlation of w_c[n] will generate two correlation peaks that are separated by pL ' (as can be seen in Fig. 15). pL ' is an estimate ofthe circular shift pL between W_di and wj₂, which is part ofthe payload, and is defined as

In addition to pL, extra information can be encoded by changing the relative signs ofthe embedded watermarks.

In the detector, this is seen as a relative sign r_sign between the correlation peaks. It may be defined as:

r,_lg, - ²- ^{+ z + 3} _S {0.1,2,3} (7)

where

are respectively estimates ofthe sign bits ri (input 80) and r₂ (input 90) of Fig. 4, and cLi and cL₂ are the values ofthe correlation peak corresponding to wj) and W_d2 respectively. The overall watermark payload pL_w, for an error- free detection, is then given as a combination of r_Sjg_n and pL:

pL_w = (r_sisn,pL) . (8)

The maximum information (I_m<Y), in number of bits, that can be carried by a watermark sequence of length L_w is thus given by:

In such a scheme, the payload is immune to relative offset between the embedder and the detector, and also to possible time scale modifications.

The window shaping function has been identified as one ofthe main parameters that controls the robustness and audibility behavior ofthe present watermarking scheme. As illustrated in Figs. 6a and b, two examples of possible window shaping functions are herein described - a raised cosine function and a bi-phase function. It is preferable to use a bi-phase window function instead of a raised cosine window function, so as to obtain a quasi DC-free watermark signal. This is illustrated in Figs. 7a and 7b, showing the frequency spectra corresponding to a watermark sequence (in this case a sequence of afk] = {1,1,-1,1,-1,-1,}) conditioned with respectively a raised cosine and a bi-phase window shaping function. As can be seen, the frequency spectrum for the raised cosine conditioned watermark sequence has a maximum at frequency/ = 0, whilst the frequency spectrum for the bi-phase shaped watermark sequence has a minimum at/= 0 i.e. it has very little DC component.

Useful information is only contained in the non-DC component ofthe watermark. Consequently, for the same added watermark energy, a watermark conditioned with the bi-phase window will carry more useful information than one conditioned by the raised cosine window. As a result, the bi-phase window offers superior audibility performance for the same robustness or, conversely, it allows a better robustness for the same audibility quality. Such a bi-phase function could be utilized as a window shaping function for other watermarking schemes. In other words, a bi-phase function could be applied to reduce the DC component of signals (such as a watermark) that are to be incorporated into another signal.

Fig. 8 shows a block diagram of a watermark detector (200, 300, 400). The detector consists of three major stages: (a) the watermark symbol extraction stage (200), (b) the buffering and interpolation stage (300), and (c) the correlation and decision stage (400).

In the symbol extraction stage (200), the received watermarked signal y '[n] is processed to generate multiple (N_b) estimates ofthe watermarked sequence. These estimates ofthe watermark sequence are required to resolve time offset that may exist between the embedder and the detector, so that the watermark detector can synchronize to the watermark sequence inserted in the host signal.

In the buffering and interpolation stage (300), these estimates are demultiplexed into N_b separate buffers, and an interpolation is applied to each buffer to resolve time scale modifications that may have occurred, e.g. a drift in sampling (clock) frequency may have resulted in a stretch or shrink in the time domain signal (i.e. the watermark may have been stretched or shrunk).

In the correlation and decision stage (400), the content of each buffer is correlated with the reference watermark and the maximum correlation peaks are compared against a threshold to determine the likelihood of whether the watermark is indeed embedded within the received signal y'[n].

In order to maximize the accuracy ofthe watermark detection, the watermark detection process is typically carried out over a length of received signal y'fnj that is 3 to 4 times that ofthe watermark sequence length. Thus each watermark symbol to be detected can be constructed by taking the average of several estimates of said symbol. This averaging process is referred to as smoothing, and the number of times the averaging is done is referred to as the smoothing factor Sf. Let L_D be the detection window length, defined as the length of the audio segment (in number of samples) over which a watermark detection truth-value is reported. Then, L_D=SJL_WT_S, where T_s is the symbol period and L_w the number of symbols within the watermark sequence. During symbol extraction, a factor T_s decimation takes place in the energy computation stage. Thus, the length (L_b) of each buffer 320 within the buffering and interpolation stage is L =SfL_w.

In the watermark symbol extraction stage 200 shown in Fig. 8, the incoming watermark signal y '[n] is input to the optional signal conditioning filter H_b(210). This filter 210 is typically a band pass filter and has the same behavior as the corresponding filter (H, 15) shown in Fig. 2. The output ofthe filter Hi, is y'bfnj and, assuming linearity within the transmission medium, it follows from equations (1) and (3):

Y_b ["] « _t[n] = (1 + a )x_b[ (10)

Note that in the above expression, the possible time offset between the embedder and the detector is implicitly ignored. For ease of explanation ofthe general watermarking scheme principles, from now on, it is assumed that there is perfect synchronism between the embedder and the detector (i.e. no offset). Explanation is given however below in reference to Fig. 11 of how to compensate for time offset in accordance with the present invention.

Note that when no filter is used in the embedder (i.e., when Η=l) then H_b in the detector can also be omitted, or it can still be included to improve the detection performance. If Hb is omitted, then y_b in equation (10) is replaced with y. The rest ofthe processing is the same.

We assume that the audio signal is divided into frames of length T_s, and that y'b,m[n] is the n-th sample ofthe m-th filtered frame signal. The energy Efm] corresponding to the m-th frame is thus:

Combining this with equation 10, it follows that:

+ ocw_e[m])x_b nf (12)

where w_e[m] is the m-th extracted watermark symbol and contains N_b time-multiplexed estimates ofthe embedded watermark sequences. Solving for w_e[m] in equation 12 and ignoring higher order terms of α, gives the following approximation:

In the watermark extraction stage 200 shown in Fig. 8, the output y ' [nJ ofthe filter H_b is provided as an input to a frame divider 220, which divides the audio signal into frames of length T_s i.e. into y '_b.mfnj, with the energy calculating unit 230 then being used to calculate the energy corresponding to each ofthe framed signals as per equation (12). The output of this energy calculation unit 230 is then provided as an input to the whitening stage H_w (240) which performs the function shown in equation 13 so as to provide an output w_e[m] . Alternative implementations (240A, 240B) of this whitening stage are illustrated in Figs. 9 and 10.

It will be realized that the denominator of equation 13 contains a term that requires knowledge ofthe host (original) signal x. As the signal x is not available to the detector, it means that in order to calculate w_e[m] then the denominator of equation 13 must be estimated. Below is described how such an estimation can be achieved for the two described window shaping functions (the raised cosine window shaping function and the biphase window shaping function), but it will equally be appreciated that the teaching could be extended to other window shaping functions.

In relation to the raised cosine window shaping function shown in Fig. 6(a), it has been realized that the audio envelope induced by the watermark contributes only to the noisy part ofthe energy function EfmJ. The slowly varying part (i.e. the low frequency component) is predominately due to the contribution ofthe envelope ofthe original audio signal x. Thus, equation 13 may be approximated by:

where "lowpassQ" is a low pass filter function. Thus, it will be appreciated that the whitening filter H_w for the raised cosine window shape in the function can be realized as shown in Fig. 9.

As can be seen, such a whitening filter H_w (240A) comprises an input 242A for receiving the signal EfmJ. A portion of this signal is then passed through the low pass filter 247 A to produce a low pass filtered energy signal EipfmJ, which in turn is provided as an input to the calculation stage 248 A along with the function EfmJ. The calculation stage 248 A then divides EfmJ by EipfmJ to calculate the extracted watermark symbol w_efmj.

When a bi-phase window function is employed in the watermark conditioning stage ofthe embedder, a different approach should be utilized to estimate the envelope ofthe original audio, and hence to calculate w_e[mj. It will be seen by examination ofthe bi-phase window function shown in Fig.

6b, that when the audio envelope is modulated with such a window function, the first and the second halves ofthe frame are scaled in opposite directions. In the detector, this property is utilized to estimate the envelope energy of the host signal x.

Consequently, within the detector, each audio frame is first sub-divided into two halves. The energy functions corresponding to the first and second half-frames are hence given by

and

T.-i

£ ]= Σ f (16) n=T 2

respectively. As the envelope ofthe original audio is modulated in opposite directions within the two sub-frames, the original audio envelope can be approximated as the mean of EjfmJ and E₂ m/. Further, the instantaneous modulation value can be taken as the difference between these two functions. Thus, for the bi-phase window function, the watermark w_efmj can be approximated by:

Consequently, the whitening filter H_w (240B) in Fig. 8 for a bi-phase window shaping function can be realized as shown in Fig. 10. Inputs 242B and 243B respectively receive the energy functions ofthe first and second half frames EifmJ and E₂ . Each energy function is then split up into two, and provided to adders 245B and 246B which respectively calculate EifmJ - E₂[m], and EifmJ + E_∑fmJ. Both of these calculated functions are then passed to the calculating unit 248B which divides the value from adder 245B by the value from 246B so as to calculate w_e[rø/, containing N_b time-multiplexed estimates ofthe embedded watermark sequences, in accordance with equation 17.

This output w_efmj is then passed to the buffering and interpolation stage 300 (Fig. 8), where the signal is de-multiplexed by a de-multiplexer 310, buffered in buffers 320 of length L_b, so as to resolve a lack of synchronism between the embedder and the detector, and interpolated within the interpolation unit 330 so as to compensate for a time scale modification between the embedder and the detector.

In order to maximize the possible robustness of a watermark, it is important to make sure that the watermarking system is immune to both time offsets and drifts in sampling frequency between the embedder and the detector. In other words, the watermark detector must be able to synchronize to the watermark sequence inserted in the host signal. Fig. 11 illustrates the process carried out by the buffering and interpolation stage 300 to resolve the offset issue. The example described illustrates the process for resolving offset when a raised cosine window shaping function has been employed in the watermark embedding process. However, in principle the same technique is applicable when the bi-phase window shaping function has been used.

Referring to Fig. 11, after filtering by the filter H_b 210, the incoming audio signal stream y ' fnJ is separated into preferably overlapping frames 302 of effective length T_s by the frame divider 220. Preferably, to resolve possible offset between the embedder and the detector, each frame is divided into N_b sub-frames (304a, 304b,...,304x), and the above computations (equations (12) to (17)) are applied on a sub-frame basis.

Preferably, each sub-frame overlaps with an adjacent sub-frame. In the example shown, it can be seen that there is a 50% overlap (T_s/N_b) of each sub-frame (304a, 304b, ..., 304x), with each ofthe sub-frames being of length 2T_S/N_b. When overlapping sub- frames are considered, the main frames are preferably longer than the symbol period T_s so as to allow inter-frame overlap as shown in Fig. 11.

The energy ofthe audio is then computed for each sub-frame by the whitening stage 240, and the resulting values are de-multiplexed into the N_b buffers 320 by the demultiplexer 310. Each one(i?Λ B2, .... Bm) ofthe buffers 320 will thus contain a sequence of values, with the first buffer Bj containing a sequence of values corresponding to the first sub- frame within each frame, the second buffer B2 containing a sequence of values corresponding to the second sub-frame within each frame etc. If w_Di is the content of the i-th buffer, then it can be shown that: ^w _D,[k] = w_e[k - N_b

e {0,...,L_b - 1} (18)

where L_b is the buffer length.

For a raised cosine window shaping function, the energy ofthe embedded watermark is concentrated near the center ofthe frame, such that the sub-frame best aligned with the center ofthe frame will result in a distinctly better estimate ofthe embedded watermark symbol than all the other sub-frames. Effectively, each buffer thus contains an estimate ofthe symbol sequence, the estimates corresponding to the sequences having different time offsets. The sub-frame best aligned with the center ofthe frame (i.e. the best estimate ofthe correctly aligned frame) is determined by correlating the contents of each buffer with the reference watermark sequence. The sequence with the maximum correlation peak value is chosen as the best estimate ofthe correctly aligned frame. The corresponding confidence level, as described below, is used to determine the truth-value ofthe detection. Preferably, the correlation process is halted once an estimated watermark sequence with a correlation peak above the defined threshold has been found. Typically, the length of each buffer is between 3 to 4 times the watermark sequence length L_w, and is thus typically of length between 2048 and 8192 symbols, and Na is typically within the range of 2 to 8.

The buffer is normally 3 to 4 times that ofthe watermark sequence so that each watermark symbol can be constructed by taking the averages of several estimates of said symbol. This averaging process is referred to as smoothing, and the number of times the averaging is done is referred to as the smoothing factor Sf. Thus, given the buffer length L_b and the watermark sequence length L_w, the smoothing factor Sf is such that:

In another preferred embodiment, the detector ref nes^■ the parameters used in the offset search based upon the results of a previous search step. For instance, if a first series of estimates shows that the results stored in buffer B₃ provide the best estimate ofthe information signal, then the next offset search (either on the same received signal, or on the signal received during the next detection window) is refined by shifting the position ofthe sub-frames towards the position of the best estimate sub-frame. The estimates of the sequence having zero offset can thus be iteratively improved.

As previously mentioned, there can exist a drift in sampling (clock) frequency in digital devices, which results in a stretch or shrink in the time domain signal.

For instance, consider an audio segment s of length L that is time scaled such that it's new length becomes L, = L(l+ 17) where η is the time scaling factor, with η being a constant such that 1+τj >0; for a time stretch η>0, and for a time shrink η<0.

When the signal is not time scale modified (η =0), Ν_b estimates ofthe watermark sequence are constructed by collecting the symbols stored in the Ν_b buffers separately. Fig. 12 illustrates four buffers (Bl, B2, B3, B4), each buffer shown as a row of boxes, with each box within a row indicating a separate location within the respective buffer. The sequences w_π, W12, wι₃, w₁ are respective estimates ofthe watermark sequence. In the example shown in Fig. 12, it is assumed that the signal is not time scale modified, and hence each estimate (wπ, w^, w_I3, w_] ) represents an estimate ofthe watermark sequence with different time offset.

Consequently, each estimate (that is passed to the correlator 410) is formed by sequentially collecting the entries from each buffer. For example, the first value in sequence π (wπ [1]) is collected from the first location of Bl, the second (w_π [2]) from the second location of Bl etc, with the final value (w_π [Lb]) being collected from the final location ofthe buffer. It will be appreciated that the arrows, which connect each box in a row to the neighboring box, show the direction in which values ofthe sequence estimates are collected from the buffer locations. It will also be appreciated that, whilst only eleven buffer locations are shown for each buffer, the size ofthe buffers in practice is likely to be significantly larger than this. For example, in the preferred embodiment, the length of each buffer is typically between 2048 and 8192 locations, with the number of buffers typically being between 2 and 8. However, in order to prevent overflow of buffers during time scale search, the actual buffer lengths are set to (l+|r)max|) times the typical lengths specified above, where η_max is the expected maximum scaling factor.

When the received signal y ' [n] has been time scale modified, it is necessary to perform a time scale search in order to correctly estimate the watermark sequence. In the present invention, such a search is performed by systematically combining the extracted watermark sequence estimates (w_e[m]), preferably by systematically combining

(interpolating) the different estimates ofthe watermark sequences stored in the buffers.

Such time scale searches can be performed by utilizing any order of interpolation. In the following two preferred embodiments, two orders of interpolation will be described - the first order (linear) interpolation and the zero order interpolation. However, it will be appreciated that this technique can be extended to higher orders of interpolation e.g. quadratic and cubic interpolation.

In the first embodiment, estimates ofthe time scaled watermark sequence are provided by applying linear interpolation to the previously extracted estimates ofthe watermark sequence. To this end, it can be assumed that the intermediate values w_e[k] generated by the symbol extraction step shown in Fig. 8 are sequentially stored in a single buffer of length M in place ofthe N_b buffers. In other words, that the N_b buffers are multiplexed into a single buffer of length

where L_w and sr are as defined earlier. Let the so stretched sequence be represented by w_/ It can now be assumed that w_D represents discrete samples of an otherwise continuous function. During time scale modification, these discrete points are either pushed towards each other or stretched out. This in turn is translated to re-sampling of the watermark function. In this embodiment, re-sampling is realized via a linear interpolation technique. That is, given the watermark sequence

...,M, an interpolated watermark sequence wjfmj is generated as

Wj[m] = μw_D( l(l + η)mj ) + (l -μ)w_D( [(l +

) (20)

Where μ = f(l+ η)m 1- (1 + η)m, and [^• 7 and bJaxe the floor and the ceiling operators, respectively. After the interpolation, the watermark sequences are folded back into the N_b buffers in a similar way to that shown in Fig. 11. Let the interpolated watermark sequence folded into the buffer b e {0,...,Nb-l} be denoted by w/_bfkj, then it can be shown that ,M = _r™_D( l(N_bk + b)(\ + η)i )+ (l -μ)w_D{ [(N_bk + b)(l +

). (21)

Let for b=l, .... Nb, wo.bfkj be the pre-interpolation sequence stored in the ό-th buffer, and q_pk e {], ...sjL_w} and r_pt e{l, ...N_bJ be defined as

and

Then, it can be shown that w_D ( (N_bk + b)(\ + η)j ) = w_Dιl}Λ [g_bk ] . Putting this into equation (21), it follows that w_Itb[k] = μw_DΛk [q_bk[k]] + (1 - μ)w_D,(r_hk+i)[g_bk + 1] (22)

Thus, the interpolated buffer entries can be calculated directly from the N_& sequences wo_b, b=l,..,N_b (as shown in Fig. 8, being passed to the correlator 410), by solving equation (22). A further embodiment ofthe present invention will now be described, in which estimates ofthe time scaled watermark sequence are provided by applying zero order interpolation to the previously extracted estimates ofthe watermark sequence. This approach can be represented with equation (22) with μ = 1. In this case, the interpolation function can be written as

^W/.*[*] = w. _ttø*[*]] . (23)

where q_pk e (J, ...sL_w} and r_pk efl, ...N_b} are as defined above.

A graphical interpretation of equation (23) is shown in Figs. 13a & b. Fig. 13a shows how the different estimates ofthe correct watermark sequences (wπ, w^, W_I3, W₁ ) are extracted from the buffers for a time stretch, whilst Fig. 13b shows similar information for a time shrink. As in Fig. 12, each row of boxes represents a respective buffer, with each box representing a location within each buffer. The arrows indicate the order in which the buffer contents are collected from the estimates ofthe watermark sequences. When the audio signal is time scale modified, the start and the end ofthe framing will gradually drift backward or forward, depending respectively upon whether the signal is time scale stretched or compressed. The watermark symbol combining stage according to this embodiment tracks the size ofthe drift. When the absolute value ofthe cumulative drift exceeds T_s N_b (where N_b is the number of buffers i.e. the number of consecutive symbols that represent a single watermark symbol), then the symbol collection sequence from the buffers is adjusted to provide the next best estimate ofthe symbol from the buffers. In other words, the buffer counters are incremented or decremented (depending on drift direction), and a circular rotation ofthe buffer pointer for each watermark sequence estimation (wn, w_I2, w>_/j, wu) is performed. Let k be the buffer entry counter, where £ is an integer representing each location within each buffer i.e. k=l represents the first location within each buffer, k=2 the second etc. If the estimates ofthe watermark sequence are being taken from the buffers with no time scale modification (as shown in Fig. 12), then it will be appreciated that the values in the first sequence can be represented by wnfkj. However, for time scaled estimates, assuming that an estimate η is being made

ofthe time scale, then when

the counter values and the buffers from which the watermark estimates are taken are changed.

If η is positive (time stretch), the counter for the first buffer is incremented. The ordering ofthe buffers is also circularly shifted (i.e. the watermark sequence estimate w_u previously being taken from buffer one will now been taken from buffer four, the estimate from buffer two will now be taken from buffer one, the estimate from three will now be taken from buffer two, and the estimate from buffer four will now be taken from buffer three). A similar circular shift is also performed on the buffer counter k. This is shown diagrammatically in Fig. 13 a. If η is negative (time stretch), the counter for the first buffer is incremented, and the ordering ofthe buffers is circularly shifted (i.e. the watermark sequence estimate w_n previously being taken from buffer one will now be taken from buffer two, the estimate from buffer two will now be taken from buffer three, the estimate from three will now be taken from buffer four, and the estimate from buffer four will now be taken from buffer one). A similar circular shift is also performed on the buffer counter k. This is shown diagrammatically in Fig. 13b.

After these circular shifts and adjustment to the buffer counters have been performed the symbol collection to form the different estimates ofthe watermark sequences continues from left to right until

« (n + l)/N_b (i.e. the next interchange position is reached). The process of buffer order interchanging and the sequential symbol collection is then repeated until the end ofthe buffer is reached.

Consequently, it will be appreciated that a zeroth order interpolation ofthe time scaled watermark sequence has been performed. In other words, the time scaled watermark sequence has been estimated by selecting those values from the original, non time scaled watermark sequence estimates that would most closely correspond to the temporal positions ofthe time scaled watermark sequence. By utilizing previously extracted estimates ofthe watermark sequence, such a technique efficiently resolves the problems of estimating correctly time scaled watermarks, with minimal cost in terms of computational overhead. Such estimates ofthe time scaled watermark sequence will then be passed to the correlator (410), so as to determine whether the predicted time shift η accurately represents the time shift ofthe received signal i.e. do the estimates provided to the correlator provide good correlation peaks. If not, then the time scale search will be repeated for a different estimated value i.e. a different value of η. Due to possible time scale modification, the detection truth-value (whether or not the signal includes a watermark) is determined only after the appropriate scale search has been conducted. Let Δη be the scale search step size and let us assume that we want the watermark to survive all the scale modifications in the interval [/_/"min, max]- The total number of visited scales is then given by T ^max ~ ?7min .. .. Aη ⁽²⁴⁾

To rninimize N_η it is preferred to find the maximum value of Δη that can still allow an exhaustive scale search. To this end, experimental results show that the detection performance is not significantly affected if the time scaling does not exceed half of the inverse ofthe buffer length. This means that, for an exhaustive scale search, Δη should be such that 2

Aη ≤

N_bs_fL_w

Putting this into equation (24), it follows that it is preferable to conduct a search over

time scales in order to conduct an exhaustive scale search. Clearly, any scale search can be time consuming. Thus, the complexity issue and cost in computing overhead should be considered when choosing the watermark embedding parameters N_b, S_f and L_w.

In one preferred embodiment the scale search is adapted such that information acquired during detection is utilized to plan an optimum search in the subsequent detection windows. For example, the scale search in the next detection window is started around the current optimum scale.

An alternative embodiment illustrated in Fig. 14 provides a method for efficient walk through the scale space by grid refinement. The most straightforward solution is a linear search from the minimum scale towards the maximum scale by adding up an incremental step. Assuming correlation, and thus confidence level, does not change abruptly from one scale to the next, one can considerably reduce the amount of scales visited during the search by reducing the space granularity. As shown in Fig. 14, the algorithm starts at scale zero and is repeated until a minimum granularity is reached or the watermark is detected (i.e., a local maximum for the confidence level is found) and/or the confidence level exceeds a predetermined threshold. When one has an indication where to start the scale search (e.g. an initial estimation from a previous detection), a random or linear search around this scale may suffice.

As shown in Fig. 8, outputs (WD_I, W_D2, ... w_DNb) from the buffering stage are passed to the interpolation stage and, after interpolation, the outputs (wn, wn, ... w/w,) of this stage, which are needed to resolve a possible time scale modification in the watermarked signal, are passed to the correlation and decision stage. All ofthe estimates (wn, wn,- wmb) ofthe watermark corresponding to the different possible offset values are passed to the correlation and decision stage 400.

The correlator 410 calculates the correlation of each estimate

...,N_b with respect to the reference watermark sequence w_cfkj. Each respective correlation output corresponding to each estimate is then applied to the maximum detection unit 420 which determines which two estimates provided the maximum correlation peak values. These estimates are chosen as the ones that best fit the circularly shifted versions W_di and Wd₂ ofthe reference watermark. The correlation values for these estimated sequences are passed to the threshold detector and payload extractor unit 430. The reference watermark sequence w_s used within the detector corresponds to

(a possibly circularly shifted version of) the original watermark sequence applied to the host signal. For instance, if the watermark signal was calculated using a random number generator with seed S within the embedder, then equally the detector can calculate the same random number sequence using the same random number generation algorithm and the same initial seed S so as to determine the watermark signal. Altematively, the watermark signal originally applied in the embedder and utilized by the detector as a reference could simply be any predetermined sequence.

Fig. 15 shows a typical shape of a correlation function as output from the correlator 410. The horizontal scale shows the correlation delay (in terms ofthe sequence samples). The vertical scale on the left hand side (referred to as the confidence level cL) represents the value ofthe correlation peak normalized with respect to the standard deviation ofthe normally distributed correlation function.

As can be seen, the typical correlation is relatively flat with respect to cL, and centered about cL - 0. However, the function contains two peaks, which are separated by pL (see equation 6) and extend upwards to cL values that are above the detection threshold when a watermark is present. When the correlation peaks are negative, the above statement applies to their absolute values.

A horizontal line (shown in the Fig. as being set at cL = 8.7) represents the detection threshold. The detection threshold value controls the false alarm rate.

Two kinds of false alarms exist: The false positive rate, defined as the probability of detecting a watermark in non watermarked items, and the false negative rate, which is defined as the probability of not detecting a watermark in watermarked items. Generally, the requirement ofthe false positive alarm is more stringent than that ofthe false negative. The scale on the right hand side of Fig. 11 illustrates the probability of a false positive alarm ?. As can be seen in the example shown, the probability of a false positive p=10^' is equivalent to the threshold cL = 8.7, whilst/? = 10^'83 is equivalent to cL = 20. After each detection interval, the detector determines whether the original watermark is present or whether it is not present, and on this basis outputs a "yes" or a "no" decision. If desired, to improve this decision making process, a number of detection windows may be considered. In such an instance, the false positive probability is a combination ofthe individual probabilities for each detection window considered, dependent upon the desired criteria. For instance, it could be determined that if the correlation function has two peaks above a threshold of cL = 7 on any two out of three detection intervals, then the watermark is deemed to be present. Such detection criteria can be altered depending upon the desired use ofthe watermark signal and to take into account factors such as the original quality ofthe host signal and how badly the signal is likely to be corrupted during normal transmission. The payload extractor unit 430 may subsequently be utilized to extract the payload (e.g. information content) from the detected watermark signal. Once the unit has estimated the two correlation peaks cLi and cL₂ that exceed the detection threshold, an estimate cL' ofthe circular shift cL (defined in equation (6)) is derived as the distance between the peaks . Next, the signs pi and ₂ ofthe correlation peaks are determined, and hence r_Sjg_n calculated from equation (7). The overall watermark payload may then be calculated using equation (8). For instance, it can be seen in Fig. 15 that pL is the relative distance between the two peaks. Both peaks are positive i.e. pj = +1, and ₂ = +1. From equation (7), r_Sj_gn = 3. Consequently, the payload pLw = <3, pL>. It will be appreciated by the skilled person that various implementations not specifically described would be understood as falling within the scope ofthe present invention. For instance, whilst only the functionality ofthe detecting apparatus has been described, it will be appreciated that the apparatus could be realized as a digital circuit, an analog circuit, a computer program, or a combination thereof.

Equally, whilst the above embodiment has been described with reference to an audio signal, it will be appreciated that the present invention can be applied to add information to other types of signal, for instance information or multimedia signals, such as video and data signals. Further, it will be appreciated that the invention can be applied to watermarking schemes containing only one watermarking sequence (i.e. a 1-bit scheme), or to watermarking schemes containing multiple watermarking sequences. Such multiple sequences can be simultaneously or successively embedded within the host signal.

Within the specification it will be appreciated that the word "comprising" does not exclude other elements or steps, that "a" or "and" does not exclude a plurality, and that a single processor or other unit may fulfil the functions of several means recited in the claims.

Claims

CLAIMS:

1. A method of compensating for a linear time scale change in a received signal, the signal being modified by a sequence of symbols in the time domain, the method comprising the steps of:

(a) extracting an initial estimate ofthe sequence of symbols from said received signal;

(b) forming an estimate of a correctly time scaled sequence ofthe symbols by interpolating the values of said initial estimate.

2. A method as claimed in claim 1, wherein step (b) is repeated so as to provide a range of estimates corresponding to different time scalings.

3. A method as claimed in claim 1, wherein said interpolation is at least one of zeroth order interpolation, linear interpolation, quadratic interpolation and cubic interpolation.

4. A method as claimed in claim 1 , the method further comprising the step of processing each estimate as though it were the correctly time scaled sequence ofthe symbols, so as to determine which estimate is the best estimate.

5. A method as claimed in claim 1, the method further comprising the steps of correlating each of said estimates with a reference corresponding to said sequence of symbols; and taking the estimate with the maximum correlation peak as the best estimate.

6. A method as claimed in claim 1, wherein said initial estimate ofthe sequence of symbols is stored in a buffer.

7. A method as claimed in claim 6, wherein said buffer is of total length M, the

M total number of scale searches conducted is N₇ = — (η_max -η_miτ>) where 77_mιn, max correspond respectively to the minimum and maximum likely time scale modifications ofthe signal.

8. A method as claimed in claim 1, wherein said initial estimates ofthe sequence of symbols comprises a sequence of N_b estimates for each symbol, each ofthe N_b estimates corresponding to a different time offset of a symbol.

9. A method as claimed in claim 1, wherein the scale search in the next detection window is adapted based on the information acquired during the current detection window.

10. A method as claimed in claim 1, wherein the scale space is searched using an optimal searching algorithm.

11. A method as claimed in claim 10, wherein the searching algorithm is the grid refinement algorithm.

12. A computer program arranged to perform the method as claimed in claim 1.

13. A record carrier comprising the computer program as claimed in claim 12.

14. A method of making available for downloading a computer program as claimed in claim 12.

15. An apparatus arranged to compensate for a linear time scale change in a received signal, the signal being modified by a sequence of symbols in the time domain, the apparatus comprising: an extractor arranged to extract an initial estimate ofthe sequence of symbols from said received signal; and an interpolator arranged to form an estimate of a correctly time scaled sequence of the symbols by interpolating the values of said initial estimate.

16. An apparatus as claimed in claim 15, the apparatus further comprising a buffer arranged to store one or more of said estimates.

7. A decoder comprising the apparatus as claimed in claim 15.