KR20050023454A - Watermark detection - Google Patents

Abstract

A watermark detection method is disclosed that is based on the calculation of cross-correlation between a guess signal and a watermark. The sequence 61 of signal samples correlated with the watermark is sub-sequences to be stronger for prolonged dominant signal components that adversely affect the correlation. Divided by (A (k)). The sub-sequences individually represent initial signal changes, but collectively (62) a weighting function to obtain modified sub-sequences B (k) which represent a flatter distribution of sample values. Is handled by As such, the fundamental peaks in the signal are substantially reduced.

Description

Watermark Detection {WATERMARK DETECTION}

The present invention relates to a method and apparatus for detecting a watermark in a signal, the method comprising calculating a correlation between a sequence of signal samples and a predetermined watermark and detecting whether the correlation exceeds a given threshold Steps.

Watermarks are undetectable messages embedded in the content of information signals such as audio or video. Watermarks support various applications such as monitoring and copy control. In general, a watermark is embedded in the signal by modifying the signal samples according to each sample of the watermark. The term "samples" refers to signal values in the region where the watermark is embedded.

Conventional watermark embedding and detection systems for audio are described by Jaap Haitsma, Michiel van der Veen, Ton Kalker and Fons Bruekers, October 30, 2002-November 4, 2002, pages 199-122 of the ACM Multimedia Conference. Title "Audio Watermarking for Monitoring and Copy Protection". The audio signal is divided into frames and converted into a frequency domain. The watermark sequence is embedded with the magnitudes of the Fourier coefficients of each frame. The detector receives a time domain version of the watermarked audio signal. The received signal is divided into frames and converted into a frequency domain. The magnitudes of the Fourier coefficients are cross-correlated with the watermark sequence. If the correlation exceeds a given threshold, then there is a watermark. The expression "sequence of signal samples" defined at the outset represents the magnitudes of the Fourier coefficients of the audio frame in this case.

A conventional watermark embedding and detection system for video is entitled, “A Video watermarking System” by Ton Kalker, Geert Depovere, Jaap Haitsma and Maurice Maes, January 1999, Vol.3657, Proceedings of SPIE, pages 103-112. for Broadcast Monitoring ". In the system, the watermark is embedded in the pixel area. The watermark sequence is a 128x128 watermark pattern tiled through the image. The watermark detector correlates 128x128 image blocks with the watermark pattern. If the correlation is large enough, there is a watermark. The expression "sequence of signal samples" defined at the outset represents the image blocks of 128x128 pixels in this case.

Watermark detection algorithms can be based on a strong single tone that is present in or appended to an audio signal, or a strong logo present that is on a fixed position every frame of video, or the base of every frame. It is sensitive to certain signal conditions or attacks, such as the white captioned characters in.

1 is a schematic illustration of a conventional apparatus for embedding a watermark to provide background information for watermark embedding processing;

2 schematically shows a preferred embodiment of an apparatus for detecting a watermark according to the present invention;

3 is a graph showing correlation peak values of an audio signal illustrating the performance of the method according to the invention.

4 to 6 are views showing the operation of the watermark detection apparatus shown in FIG.

Fig. 7 is another graph showing correlation peak values indicating performance of the watermark detection method according to the present invention.

An object of the present invention is to improve the performance of the conventional watermark detection method.

Thus, the method according to the invention comprises the step of preprocessing said sequence of signal samples, said preprocessing:

Dividing the sequence of signal samples into sub-sequences;

Weighting all signal samples of the sub-sequences with equal weights and changing the weights between the sub-sequences to obtain a substantially uniform distribution of signal samples over the sequence; And

Concatenating the weighted sub-sequences to obtain a preprocessed sequence of signal samples.

The method according to the invention effectively suppresses large signal peaks while maintaining small signal changes indicative of a watermark. This is accomplished without knowing or detecting the location of the disturbing component in the signal.

The present invention is particularly efficient when the watermark detection method involves the accumulation of multiple signal sequences. Typically such accumulation improves detection reliability (watermark sequences are reasonable but the signal is averaged), but not where the signal contains the same disturbing component in all the accumulated sequences. In a preferred embodiment of the method according to the invention, the pretreatment step is applied to the accumulated sequences. Thus, it has been achieved that the disturbing component can be effectively removed from the accumulated sequences.

In an advantageous embodiment according to the invention, the sequence of signal samples is divided into overlapping, preferably windowed, sub-sequences. Suitable windows are the well-known Hanning window or the square root of the Hanning window. 50% overlap was found to produce good results. The concatenation sequence to be correlated with the watermark is obtained by adding weighted sub-sequences.

Preferably, the weighting step comprises Fourier transforming a sub-sequence of signal samples, normalizing the Fourier coefficients, and inversely transforming the normalized coefficients. Alternatively, the weighting step includes dividing all signal samples of the sub-sequence into the maximum signal samples of the sub-sequence. The second option, ie scaling, has a lower computational complexity than the first option, which is weighted by normalizing the magnitudes in the frequency domain. In embodiments, the sequence is adaptively weighted based on the signal characteristics.

These and other aspects of the invention will be apparent from and elucidated with reference to the accompanying drawings.

The present invention now describes the detection of a watermark embedded in an audio signal. First, the embedding apparatus will be described to provide background information. 1 schematically shows the device. The apparatus includes an adder 101 for receiving an audio signal in the form of audio samples x (n) and adding a watermark w (n) to the signal. The dominant part of the watermark w (n) is derived from the Fourier domain. The apparatus includes a splitting unit 102 that splits an audio signal into sequences or frames of 2048 samples. The sequences are transformed using Fourier transform 103. The random watermark W (k) in the frequency domain is obtained separately from the normal distribution with mean and standard deviations 0 and 1. The watermark W (k) is periodically shifted in the shifting circuit 104 by an amount representing the 10-bit payload d. The magnitudes of the Fourier coefficients are modified by a multiplier 105 as follows.

W _i (k) = W _s (k) X _i (k)

Where i represents a frame or sequence number, X _i (k) is a spectral representation of the frame (x _i (n)), W _s (k) is a periodically shifted version of W (k), and Wi ( k) shows the result of the frequency domain watermark. Inverse Fourier transform 106 is used to obtain a time domain watermark representation w (n).

2 schematically shows a preferred embodiment of an apparatus for detecting a watermark according to the present invention. As shown in the figure, the apparatus comprises three main steps: accumulation (1), pre-processing (2), and correlation (3).

In the dividing unit 11 of the accumulating step, the apparatus divides the speculative audio signal y (n) into frames or sequences y _i (n) of 2048 audio samples. Each sequence is Fourier transformed 12 and the magnitudes of the Fourier coefficients Y _i (k) are calculated 13. The magnitudes of the Fourier coefficients of frame i constitute a sequence of 1024 real numbers (| Y | _i (k)) in which watermark information is embedded. In a preferred embodiment of the apparatus, a plurality of said sequences | Y | _i (k) are accumulated by accumulator 14 to obtain an accumulated sequence Y (k). The number of accumulating sequences is chosen to represent a period, that is, an audio signal period of two seconds.

The correlation step 3 is now briefly initiated. For a detailed description of watermark detection using correlation, see international patent application WO 99/45707. The correlation step calculates a correlation C between the accumulated sequence of signal samples and all possible shifted versions of the watermark sequence W (k) ("signal samples" in this example represent magnitudes of Fourier coefficients). Note). The correlation step receives the sequence Z (k). The correlation step is initially assumed to directly receive the accumulated sequence from the accumulation step 1 (ie, Z (k) = Y (k)).

The cross-correlation for all possible shifted versions of W (k) is most efficiently calculated using the Fourier transform. The conventional cross-correlation can be recorded as follows.

C = F ^-1 (F (Z (k)) × F ^* (W (k)))

Where F (.) Denotes a Fourier transform, F ^* (.) Denotes a Fourier transform that includes the use of complex Fourier coefficients, and F- ¹ (.) Denotes an inverse Fourier transform. Respective transformations are performed by the Fourier transform circuits 31, 32, and 33 of FIG. 2. Multiplication is performed by multiplier 34.

Detection performance is enhanced by symmetrical phase only filtering (SPOMF). In the cross-correlation procedure, only phase information of signals F (Z (k)) and F ^* (W (k)) is used. Phase-only operation is defined as follows and performed by the respective phase extraction circuits 35 and 36 in FIG.

The peak detector 4 has a peak value where the cross-correlation function C is greater than a given detection threshold. ), E.g., 5 is the standard deviation of the correlation function. In this case, the watermark W (k) exists. The peak detector also retrieves the position of the peak value corresponding to the shift amount applied to the watermark W (k), thus representing a 10 bit payload d. However, this aspect is not relevant to the present invention.

3 shows correlation peak values (measured at one second intervals of an audio signal, Shows a graph. Dark line 31 represents the result for a regular piece of music. As you can easily see, each peak value is a threshold (5 E.g., the signal has an embedded watermark. The dashed lines 32 represent peak values for the same pieces of music and are distributed with a strong 15 kHz sine wave. None of the peaks have a threshold (5 ) Is not exceeded. The detector now makes a false decision that the signal does not have an embedded watermark. This problem is described with reference to FIGS. 4 and 5. In Fig. 4, the number 41 represents the general accumulation sequence Y (k) derived from the normal piece of music. In Fig. 5, the number 51 represents the corresponding sequence Y (k) derived from the same but disturbing piece of music. As the 15 kHz tone dominates the signal, the magnitude changes of the Fourier components in the sequence 51 carrying the watermark information are insignificantly reduced compared to the changes in the sequence 41.

A possible solution to overcome this problem is to ignore the signal parts of the video frame parts or the audio spectral parts, for example where there are disturbing components. For example, the position of the logo in the video signal is known in advance so that the corresponding pixels can be ignored. Or, if the audio watermark detector observes an FM radio station, frequencies close to the carrier can be ignored. Ignoring parts of the signal may appear to supply more or less weight functions suddenly. In general, however, the location of disturbing components is not known. Some kind of mechanism wants to adapt the weighting function to the signal.

Thus, the apparatus for detecting a watermark according to the present invention includes a preprocessing step 2 between the accumulation step 1 and the correlation step 3 (see Fig. 2). The pretreatment step comprises a sub-division unit 21, a weighting circuit 22, and a chain circuit 23.

The sub-division unit 21 divides the accumulated sequence Y (k) into as many overlapping and windowed sub-sequences A (k) as possible. For audio signals where the sequence Y (k) contains 1024 signal samples, the sub-sequence length of 16 samples has been found to be a good choice.

The weighting circuit 22 weights each individual sub-sequence as a weighting function. The original changes of the signal samples within each sub-sequence are maintained, while the weighting function is chosen such that the distribution of the signal samples is substantially uniform throughout the entire sequence. The expression “substantially uniform” can mean, for example, that the mean value of the signal samples of the sub-sequences is the same for the entire sub-sequences.

In one embodiment, this is obtained by normalizing the magnitudes of each sub-sequence of the frequency domain. Thus, the weighting circuit performs the following operations.

B (k) = F ^-1 (P (F (A (k))) (1)

Where F (.) Represents the Fourier transform, P (.) Represents the phase-only operation as described above, and F- ¹ (.) Represents the Inverse Fourier transform.

In another embodiment, the weights are performed by the following scaling operation.

Where A _k and B _k represent the samples of the individually weighted sub-sequence B (k) and the original sub-sequence A (k), where A _k | is the sub-sequence A (k Is the maximum absolute value of the signal samples.

Substantially, the weighted sub-sequences B (k) are concatenated by chain circuit 23 to obtain a preprocessed sequence Z (k). If the sub-sequences overlap each other, appropriate windows (eg Hanning windows) are preferably applied to B (k). It is a preprocessed sequence Z (k) that is input to the chaining step 2.

6 shows diagrams schematically illustrating a preprocessing operation. Reference numeral 61 denotes an accumulated sequence Y (k) which is divided into sub-sequences A (k). Reference numeral 62 indicates that the sequence Z (k) is obtained by concatenating the weighted sub-sequences B (k). As shown, each sub-sequence A (k) has been weighted. The same weight has been applied to all signal samples of the sub-sequence, but different weights are applied to the other sub-sequences. While the change in the signal samples remains local, a more uniform distribution of the signal samples is the result.

4 and 5 actually illustrate the effect of the preprocessing step 2 on a particular piece of music. As mentioned above, the number 41 in Fig. 4 represents the accumulated sequence Y (k) derived from the normal musical piece. Number 51 in Fig. 5 represents the accumulated sequence Y (k) derived from the same piece of music distributed in a strong 15 kHz tone. The sequences include 1024 accumulated signal samples. Reference numerals 42 and 52 denote corresponding weighting sequences Z (k) obtained by normalizing the magnitudes of each sub-sequence in the frequency domain as defined by equation (1). Reference numerals 43 and 53 denote corresponding weighting sequences Z (k) obtained by scaling as defined by equation (2). For both music pieces, especially except for pieces of distributed music, the figures show that a significantly larger correlation peak can be expected to be detected in the correlation step.

The improvement obtained with the watermark detection method according to the invention is shown in FIG. 3. In this figure, the dark lines represent normal pieces of music, and the dashed lines represent distributed pieces of music. The dark line 31 and the broken line 32 have already been described above. Dark lines 33 and 35 illustrate the performance of the weighting operation according to equation (1). The dashed lines 34 and 36 illustrate the performance of the weighting operation according to equation (2). As can be readily seen, all peak correlation values are above the threshold 5σ used by the peak detector 4. For the sake of completeness, FIG. 7 shows the same graph with the same legends and reference numbers for the same musical pieces that are mp3 that are now encoded and then decoded.

In the above embodiments, the watermark is represented by the micro modifications of the magnitudes of the Fourier coefficients, ie in the frequency domain. However, it is understood that the invention is equally applicable to the detection of watermarks embedded in temporal or spatial (video) regions.

The watermark detection method is initiated based on the calculation of the cross-correlation between the speculative signal and the watermark. To be stronger for the extended basic signal components against the correlation, the sequence 61 of signal samples correlated with the watermark is divided into sub-sequences A (k). The sub-sequences are processed by a weighting function to obtain modified sub-sequences B (k), individually representing the original signal changes, but collectively representing a more uniform distribution of sample values (62) ). As such, the fundamental peaks of the signal are substantially reduced.

Claims (9)

A method of detecting a watermark in a signal comprising computing a correlation between a sequence of signal samples and a predetermined watermark, and detecting whether the correlation exceeds a given threshold; The method includes pre-processing the sequence of signal samples, The pretreatment step is: Dividing the sequence of signal samples into sub-sequences; Weighting all signal samples of the sub-sequences with equal weights and varying the weights by sub-sequences to obtain a substantially uniform distribution of signal samples over the sequence; And Concatenating the weighted sub-sequences to obtain a preprocessed sequence of signal samples. 2. The method of claim 1, further comprising accumulating a plurality of sequences of signal samples prior to correlation, wherein said preprocessing is applied to said cumulative sequences. 2. The method of claim 1, wherein dividing the sequence of signal samples into sub-sequences comprises dividing into overlapping sub-sequences. 4. The method of claim 3, wherein the overlap is 50%. 4. The method of claim 3, wherein dividing into overlapping sub-sequences applies a window function to the overlapping sub-sequences. 2. The method of claim 1, wherein the weighting step comprises Fourier transforming a sub-sequence of the signal samples, normalizing the magnitudes of Fourier coefficients, and inversely transforming the normalized coefficients. 2. The method of claim 1, wherein the weighting step comprises dividing all signal samples of a sub-sequence into a maximum signal sample of the sub-sequence. 11. An apparatus for detecting a watermark in a signal comprising computing means for calculating a correlation between a sequence of signal samples and a predetermined watermark, and threshold means for detecting whether the correlation exceeds a given threshold; The apparatus comprises preprocessing means for preprocessing the sequence of signal samples, The pretreatment means is: Means for dividing the sequence of signal samples into sub-sequences; Weighting means for weighting all signal samples of a sub-sequence with equal weights and changing said weights between sub-sequences to obtain a substantially uniform distribution of signal samples over said sequence; And -Concatenating means for concatenating said weighted sub-sequences to obtain a preprocessed sequence of said signal samples. A computer program product arranged to cause a computer executing a computer program to perform the method claimed in any one of claims 1 to 7.