EP1887565A1 - Method and device for detecting a tattooing identifier in an audio signal - Google Patents

Abstract

The method involves preparing a current block of samples from a digitally tattooed audio signal, and calculating a correlation by a hadamard type or Fourier transform type rapid transformation method, between the samples of the block and a random sequence that is obtained from an initialization key. Correlation values of a specific number of the samples are compared to a predetermined threshold. The preparation, calculation and comparison operations are performed with another block that is shifted by the number of samples. Independent claims are also included for the following: (1) a device for detecting a tattooing identifier in an audio signal (2) a computer program comprising code instructions for implementation of a method for detecting a digital tattooing identifier in an audio signal.

Description

The invention relates to a method and a device for detecting a tattoo or mark identifier in an audiovisual signal and in particular an audio signal.

Tattooing makes it possible to transmit additional information in support data imperceptibly. For audio-visual data this amounts to establishing a concealed data communication channel, in parallel with the conventional information transport infrastructure.

One of the main advantages of watermarking is that the auxiliary data is concealed and attached to the signal and is therefore able to withstand changes in format and sampling frequency giving the tattoo interoperability at the terminals and networks, only the detection-level ends to be compatible with the insertion.

For example, a frame of the communication channel concealed in the carrier signal consists of a synchronization mark which may also be called a tattoo identifier, followed by the message inserted in the signal, the beginning of the message being indicated by the identifier. tattoo.

After transmitting the tattooed signal, when reading the media, the detection first consists in detecting the tattoo identifier that has been inserted at the beginning of the frame. Once this identifier is identified in the audio content, for example, the detection of the bits of the message can begin. The sequences "tattoo identifier + eventual message" are continuously inserted in the media, the identifier starts coinciding with the areas of sufficient energy suitable for insertion. In reception, the detection of each sequence is also done continuously, thus making it possible to resynchronize in the event of a break in the carrier medium (editing of the signal, cuts by a hacker or even frame losses of the transmission).

Tattoo identifiers and messages are formed from random Gaussian sequences from a random number generator or Pseudo-Random sequences (Pseudo-Noise Sequences in English or PN sequences) generated by a shift register whose cells are looped so as to produce an output sequence of maximum length. Any other random sequence resulting from the theory of finite state machines (Gold sequences, Baker codes, Walsh sequences, ...) may be suitable provided that its correlation function is close to a Dirac distribution. In the remainder of the description, we will speak of random sequences as well to designate random sequences as pseudo-random sequences.

A standard procedure for generating a tattooed signal is illustrated in FIG. 1. Thus, the module 11 makes it possible to generate a random sequence of length N from an initialization value K fixed in advance. This gives a random sequence w (n).

In the case where it is desired to transmit an information bit 0 or 1, this sequence is modulated by the information to be transmitted by multiplying it by 1 if the bit to be transmitted is 1 and by -1 if the bit to be transmitted is 0. This generation step is performed in the module 12 and generates a modulated sequence ± w (n).
In case we just want to insert a tattoo identifier, the random sequence is not modulated.

A psychoacoustic model 13 is taken into account by the module 14 to generate the tattoo identifier or the message w _f (n) so as to make it inaudible. A gain adjustment step is performed at 15. This tattoo identifier or message is then added at 16 to the audio signal s (n) to form the tattooed signal s _w (n).

This signal is either stored on a media for later reading by a reader or transmitted over a fixed, mobile or broadcast transmission network.
In reception, the tattoo identifier must be detected to act on the signal received, for example by blocking the reader in the event of detection of pirated content.

A typical tattoo identifier detection procedure is illustrated in FIG.

The tattooed audio signal s _w (n) is first bleached by inverse filtering by the module 21 (corresponding to the inverse of the consideration of the psycho-acoustic model). The operation of detecting the presence of the tattoo identifier and the content of the frame itself is then carried out by calculation of correlation by the module 22 between the filtered signal x _w (n) and the random sequence w (n ) generated by the module 23 in the same way as during the insertion.

This detection determines either a frame mark k _s when the tattoo identifier is detected, or the bits of the inserted information.

Not knowing the start time of the frame in the carrier signal, the correlation calculation is slippery. This is called slippery intercorrelation.

This method consists of performing a correlation calculation between the samples of the tattooed signal and the random sequence for each offset of a sample of the tattooed signal, until the beginning of the tattoo identifier is revealed by the fact that the correlation coefficient Cor (k) exceeds a threshold S. For each sampling instant k, the correlation coefficient is given by: $$\begin{array}{cc}\mathit{Horn}\left(k\right)={\displaystyle \underset{\mathrm{not}=0}{\overset{\mathrm{NOT}-1}{\Sigma }}}{x}_w\left(\mathrm{not}+k\right)w\left(\mathrm{not}\right)& k=0,...,k_s\end{array}$$

sur N_m échantillons du signal porteur pour un bit de message. Le bit détecté sera 0 si la corrélation message Cor_m est négative, il sera de 1 dans le cas contraire.Once the frame start is detected, a single correlation calculation is sufficient to detect the message: $$\begin{array}{c}{\mathit{Horn}}_m={\displaystyle \underset{\mathrm{not}=0}{\overset{{\mathrm{NOT}}_m-1}{\Sigma }}}{x}_w\left(\mathrm{not}\right)w\left(\mathrm{not}\right)\end{array}$$

on N _m samples of the carrier signal for a message bit. The detected bit will be 0 if the correlation message Cor _m is negative, it will be 1 otherwise.

It therefore appears that in order to detect the position of the tattoo identifier, i.e. the moment the message has been inserted, it is necessary to perform an intercorrelation for each sample of the carrier signal until the identifier is found. For important values of N , the search for the identifier is therefore long and complex while the detection of the transmitted bit is done quickly.

By way of example, for a value of N equal to 4096 and a sampling frequency of 44100 Hz, the number of multiplications to be performed per second will be 18010 ⁶ multiplications per second, which is much greater than the capacity of the processors Trade.

It is therefore important to reduce this computational complexity for searching for the tattoo identifier in the carrier signal, in particular for implementation in mobile terminals.

Moreover, some techniques of fast and efficient correlation calculation are described in the state of the art.

The document A. Alrutz, MR Schroeder "A fast Hadamard transform method for the evaluation of measurement using pseudo random test signals". Proceeding of the 11th lecture on acoustics, Paris, 1983 , describes a method that efficiently calculates intercorrelations. This method applies in the case where a random sequence has been inserted periodically.

The property that is exploited to calculate the correlation is the renewed and cyclic application of the random sequence.

This method therefore does not apply to the case where it is desired to detect the beginning of a tattoo frame that represents the tattoo identifier in an audio tattoo system.

Indeed, a single identifier being inserted at a time, but in several places of the carrier signal, for example every 10 to 50 seconds, and in a location favorable to the detection, the application of the method of Hadamard cited in the This document would generate in the computation of the correlation vector, spurious folding terms all the more important that the offset between the analysis block of the tattooed signal and the inserted sequence is large. These folding terms come from the product of the signal vector tattooed by the lower diagonal of the matrix of the circularly shifted random sequence. The elements of this diagonal are equal to the folded random sequence, taken in a reverse mirror order, hence the term "folding terms". These parasitic terms, which do not exist when the random sequence is periodic as used in the document of Alrutz and Schroeder, would decrease the value of the correlation peak in the case of a non-periodic random sequence thus affecting the reliability of the detection of the identifier.

Other tattoo insertion and detection techniques that improve the speed of the correlation calculation during detection exist. In the patent application WO 2004/010376 , the tattoo is inserted into the Fourier domain, the frequency mark is added to the signal after a shift linked one-to-one with the inserted identifier. The tattooed signal in the time domain is obtained by inverse Fourier transform. In detection each frame is passed in the frequency domain by Fourier transform. Several frames of Fourier coefficients are accumulated, the result of the accumulation being correlated with the frequency mark. The maximum of the correlation gives the value of the offset therefore the content of the message of the frame. The correlation maximum is calculated by inverse Fourier transform of the product of the Fourier transform of the signal accumulated by the conjugate of the Fourier transform of the frequency mark. This method requires many passages from a frequency domain to a transformed domain of the frequency domain and vice versa, the insertion taking place in the frequency domain.

The present invention provides a solution that does not have these disadvantages by providing a method of detecting a tattoo identifier that minimizes the number of operations to be performed and the complexity of calculations while allowing a reliable decision.

The method according to the invention offers the possibility of using fast correlation calculation algorithms that do not require periodic insertion of synchronization marks.

The method according to the invention can be implemented in real time on low capacity processors and in the case of continuous reception of a tattooed signal.

• préparation d'un bloc courant de N échantillons à partir du signal tatoué;
• calcul de corrélation par une méthode de transformation rapide entre les échantillons du bloc courant et ceux de la séquence aléatoire obtenue à partir de la au moins une valeur d'initialisation;
• comparaison à un seuil prédéterminé des valeurs de corrélation issues du calcul de corrélation des Ns premiers échantillons;
• itération des étapes précédentes avec un bloc courant décalé par rapport au bloc précédent de Ns échantillons tant que le maximum des dites valeurs de corrélation n'est pas supérieure au seuil prédéterminé, le maximum des valeurs de corrélation supérieur au seuil prédéterminé indiquant le début de l'identificateur de tatouage

For this purpose, the invention provides a method of detecting a tattoo identifier in an audio signal, the tattoo identifier having been inserted into the audio signal from a random sequence of a length of N samples, N being a multiple of Ns, obtained by at least one initialization value. The method is such that it comprises the following steps:

• preparing a current block of N samples from the tattooed signal;
• correlation calculation by a method of rapid transformation between the samples of the current block and those of the random sequence obtained from the at least one initialization value;
• comparing with a predetermined threshold the correlation values derived from the correlation calculation of the first Ns samples;
• iteration of the preceding steps with a current block shifted with respect to the preceding block of Ns samples as long as the maximum of said correlation values is not greater than the predetermined threshold, the maximum of the correlation values greater than the predetermined threshold indicating the beginning of the tattoo identifier

Thus, this method of block recovery, together with the implementation of a rapid transformation method, decreases the number of correlation calculations and the complexity of this calculation with respect to the sliding-window correlation calculation performed sample by sample.

In addition, such a method makes it possible to carry out a rapid scan of the tattooed signal even if only to detect the existence of a tattoo.

In a particular embodiment, the tattoo identifier indicates the position in the audio signal of an inserted message.

Thus, the detection of the tattoo identifier makes it possible to know where the detection of the inserted message can be carried out.

The current block preparation step comprises, in a preferred embodiment, an inverse filtering step of the tattooed signal and a Wiener filtering step using a filter calculated from a correlation function of a training sequence. the tattoo identifier.

Thus, taking into account certain characteristics of the sequence used to insert the tattoo identifier makes it possible to improve the detection of the tattoo identifier by accentuating any detection peaks.

According to one embodiment, the at least one initialization value is a key (K _a ) that makes it possible to generate the random sequence.

In another embodiment, the random sequence is obtained by two initialization values, the first being a first key (K _a ) making it possible to generate an intermediate random sequence, the second being a second key (K _b ) making it possible to encrypt the intermediate random sequence to obtain the random sequence.

The use of a second key therefore provides increased security.

• permutation des échantillons du bloc courant selon une première table de permutation;
• application de la transformée d'Hadamard aux échantillons ainsi permutés;
• permutation des échantillons du vecteur résultant de la transformée d'Hadamard selon une seconde table de permutation pour obtenir les valeurs de corrélations.

In a first embodiment, the fast transformation method is a method of the Hadamard type which comprises the following steps:

• permutation of the samples of the current block according to a first permutation table;
• application of the Hadamard transform to the samples thus permuted;
• permutation of the samples of the vector resulting from the Hadamard transform according to a second permutation table to obtain the correlation values.

The use of the Hadamard method in the process according to the invention is particularly well suited. The method thus produced makes it possible to dispense with the terms of folding.

Advantageously, the method comprises a preliminary initialization step in which the first and second permutation tables as well as the random sequence are calculated and stored.

In a particular embodiment, the calculation of the random sequence, of the first and of the second permutation table is optimized by a minimum number of bit operations performed on the bits of an integer representing in each case the shift register. generating the sequence or table.

• application d'une transformée de Fourier au bloc courant;
• calcul d'un produit sur les échantillons résultant de la transformée du bloc courant et ceux du complexe conjuguée de la séquence aléatoire ayant subi une transformée de Fourier;
• application d'une transformée inverse de Fourier pour obtenir les valeurs de corrélation.

In a second embodiment, the recovery method of the invention makes it possible to use a Fourier-type method as a method of rapid transformation. This method consists of the following steps:

• applying a Fourier transform to the current block;
• calculating a product on the samples resulting from the transform of the current block and those of the conjugate complex of the Fourier transformed random sequence;
• application of a Fourier inverse transform to obtain the correlation values.

Advantageously, the method comprises an initialization step in which the Fourier transform of the random sequence and its conjugate complex are calculated and stored.

To improve the detection, the method further comprises a step of correcting the correlation values by eliminating parasitic terms.

• segmentation en sous-séquences du bloc courant;
• segmentation en sous-séquences de la séquence aléatoire, la deuxième moitié de la sous-séquence aléatoire étant mise à zéro;
• application d'une transformée de Fourier aux sous-séquences du bloc courant et aux sous-séquences de la séquence aléatoire;
• calcul de corrélations partielles par le produit sur les échantillons résultant de la transformée des sous-séquences du bloc courant et ceux du complexe conjugué des sous-séquences de la séquence aléatoire;
• sommation des corrélations partielles calculées;
• application d'une transformée de Fourier inverse à la somme issue de l'étape de sommation pour obtenir les valeurs de corrélation.

In an alternative embodiment using the Fourier transform method, the method comprises the following steps:

• segmentation into subsequences of the current block;
• segmentation into subsequences of the random sequence, the second half of the random subsequence being set to zero;
• applying a Fourier transform to the subsequences of the current block and the subsequences of the random sequence;
• calculating partial correlations by the product on the samples resulting from the transformation of the sub-sequences of the current block and those of the conjugate complex of the subsequences of the random sequence;
• summation of calculated partial correlations;
• applying an inverse Fourier transform to the sum resulting from the summing step to obtain the correlation values.

This variant makes it possible to dispense with the use of a large Fourier transform when the tattoo identifier has a large size.

• segmentation en sous-séquences du bloc courant;
• segmentation en sous-séquences de la séquence aléatoire, la première moitié de la sous-séquence aléatoire étant mise à zéro;
• application d'une transformée de Fourier aux sous-séquences du bloc courant et aux sous-séquences de la séquence aléatoire;
• calcul de corrélations partielles par le produit sur le complexe conjugué des échantillons résultant de la transformée des sous-séquences du bloc courant et ceux des échantillons résultant de la transformée des sous-séquences de la séquence aléatoire;
• sommation des corrélations partielles calculées;
• application d'une transformée de Fourier inverse à la somme issue de l'étape de sommation pour obtenir les valeurs de corrélation.

In a second variant, the method comprises the following steps:

• segmentation into subsequences of the current block;
• segmentation into subsequences of the random sequence, the first half of the random subsequence being set to zero;
• applying a Fourier transform to the subsequences of the current block and the subsequences of the random sequence;
• calculating partial correlations by the product on the conjugate complex of the samples resulting from the transformation of the subsequences of the current block and those of the samples resulting from the transformation of the subsequences of the random sequence;
• summation of calculated partial correlations;
• applying an inverse Fourier transform to the sum resulting from the summing step to obtain the correlation values.

• des moyens de préparation d'un bloc courant de N échantillons à partir du signal tatoué;
• des moyens de calcul de corrélation par une méthode de transformation rapide entre les échantillons du bloc courant et ceux de la séquence aléatoire obtenue à partir de la au moins une valeur d'initialisation;
• des moyens de comparaison à un seuil prédéterminé des valeurs de corrélation issues du calcul de corrélation des Ns premiers échantillons;
• des moyens d'obtention d'un nouveau bloc courant par décalage de Ns échantillons par rapport au bloc précédent mis en oeuvre tant qu'une des dites valeurs de corrélation n'est pas supérieure au seuil prédéterminé.

The invention also relates to a device for detecting a tattoo identifier in an audio signal, the tattoo identifier having been inserted in the audio signal from a random sequence of a length of N samples, N being multiple Ns, obtained by at least one initialization value. The device is such that it comprises:

• means for preparing a current block of N samples from the tattooed signal;
• correlation calculation means by a method of rapid transformation between the samples of the current block and those of the random sequence obtained from the at least one initialization value;
• means for comparing a predetermined threshold of the correlation values derived from the correlation calculation of the first Ns samples;
• means for obtaining a new current block by shifting Ns samples with respect to the preceding block implemented as long as one of said correlation values is not greater than the predetermined threshold.

The device thus described is able to implement the method according to the invention.

The invention finally relates to a computer program comprising code instructions for implementing the steps of the method according to the invention, when said program is executed by a processor.

The device and the computer program have the same advantages as the method they implement.

• la figure 1 représente un schéma bloc d'une méthode d'insertion de l'état de l'art;
• la figure 2 représente un schéma bloc d'une méthode de détection de l'état de l'art;
• la figure 3 illustre sous forme d'organigramme, les principales étapes d'un procédé de détection d'identificateur de tatouage conforme à l'invention;
• les figures 4a et 4b illustrent la méthode par recouvrement selon l'invention;
• la figure 5 représente les principaux modules d'un dispositif d'insertion d'identificateur de tatouage;
• la figure 6 représente les principaux modules d'un dispositif de détection d'identificateur de tatouage selon l'invention;
• la figure 7 illustre sous forme d'organigramme, une méthode de cryptage qui peut être utilisée par l'invention;
• la figure 8 illustre de façon détaillée, sous forme d'organigramme, l'étape de calcul de corrélation dans un premier mode de réalisation de l'invention;
• la figure 9 illustre de façon détaillée, sous forme d'organigramme, l'étape de calcul de corrélation dans un second mode de réalisation de l'invention;
• la figure 10 illustre une variante de réalisation du second mode de réalisation selon l'invention; et
• la figure 11 illustre un dispositif mettant en oeuvre le procédé selon l'invention.

Other features and advantages of the invention will appear more clearly on reading the following description, given solely by way of nonlimiting example, and with reference to the appended drawings, in which:

• FIG. 1 represents a block diagram of an insertion method of the state of the art;
• FIG. 2 represents a block diagram of a method of detection of the state of the art;
• Figure 3 illustrates in flowchart form, the main steps of a tattoo identifier detection method according to the invention;
• Figures 4a and 4b illustrate the overlay method according to the invention;
• Figure 5 shows the main modules of a tattoo identifier insertion device;
• FIG. 6 represents the main modules of a tattoo identifier detection device according to the invention;
• Figure 7 illustrates in flowchart form, an encryption method that can be used by the invention;
• FIG. 8 illustrates in detail, in flowchart form, the correlation calculation step in a first embodiment of the invention;
• FIG. 9 illustrates in detail, in flowchart form, the correlation calculation step in a second embodiment of the invention;
• FIG. 10 illustrates an alternative embodiment of the second embodiment according to the invention; and
• FIG. 11 illustrates a device implementing the method according to the invention.

Figure 3 illustrates the main steps of a tattoo identifier detection method according to the invention.

The step E31 is an initialization step which can make it possible to carry out calculations beforehand and to memorize them, for example the generation of the random sequence.

Step E32 consists in preparing a first current block of the filtered tattooed signal comprising an N number of samples.

Before obtaining the current block, a reading of a block of Ns samples of the tattooed signal, N being multiple of Ns, is performed. This block of Ns samples s _w ( n ), n = 0, .., N _s -1 is that which is available for reading on the storage disk or received directly by continuous flow.

Cette étape de filtrage correspond à l'étape inverse du filtrage par le modèle psycho-acoustique effectué lors de l'insertion.This block is whitened by inverse filtering followed by adaptive filtering to give a signal vector $x_w^{\mathit{sub}}\left(\mathrm{not}\right),\mathrm{not}=0,...,{\mathrm{NOT}}_s-1.$

This filtering step corresponds to the inverse step of the filtering by the psycho-acoustic model carried out during the insertion.

This signal vector is put following the previously received vectors to constitute the current block of N samples x _w ( n ), n =, 0, ..., N- 1.

Step E32 is followed by step E33 in which a correlation calculation by a fast transformation method is performed on the samples of the current block and the samples of the random sequence generated from the same (s). Initialization value (s) used during insertion. This random sequence generation can be carried out just before calculating the correlation or at the initialization stage. It is then in the latter case stored in the initialization step and read during the calculation step.

In step E34, a test is performed to find out whether the maximum of the correlation values from the correlation calculation of step E33 of the first Ns samples is greater than a threshold S.

In the case where the maximum of the correlation values is greater than the threshold S, then the tattoo identifier is detected. This value of the correlation maximum greater than the threshold S then indicates the beginning of the tattoo identifier.

The detection device can then from this tattoo identifier determine the position in the audio signal of a message that would have been inserted. This message is placed just after the last sample of the tattoo identifier represented by the random sequence of length N.

The detection of the inserted message is carried out in a simple manner as mentioned with reference to FIG. 2, by a single correlation calculation. Once the message is detected, the detector will return to the search mode for the tattoo synchronization or identifier to locate the next message. In the case where only tattoo identifiers have been inserted at defined points of the signal, the tattooed signal is scanned as in the previous case continuously. When an identifier is detected, a flag is set to 1 and the detector returns to polling mode.

In the case where in step E34, none of the correlation values of the first Ns samples exceed the threshold S, then a new current block is taken into account by shifting Ns samples E35.

Figures 4a and 4b schematically illustrate the method of the invention.

The tattooed signal represented at 40 is continuously received. The tattoo identifier resulting from a random sequence of length N consists of R blocks of Ns samples, ie N = R * Ns samples. This tattoo identifier is shown in gray under the reference 41 in FIGS. 4a and 4b. A block analysis of N samples is performed. This current block x _w (n) for n ranging from 0 to N-1 is referenced in FIG. 42. A correlation calculation is performed between this current block and the random sequence w (n) generated as at the time of insertion. . The correlation values obtained for the first Ns are then compared as shown at 43 in FIGS. 4a and 4b. If the maximum of the correlation values does not exceed the threshold S, then the current block to be analyzed is shifted by an Ns number of samples as illustrated in FIGS. 4a and 4b.

Il est inséré à la place des Ns dernières composantes de droite comme représenté sur les figures 4a et 4b.When a new block of the tattooed signal of Ns samples is available by reading on storage disk (as for example CD, DVD, memory removable) or received in continuous flow, it undergoes a step of inverse filtering and Wiener filtering to give a signal vector ${\tilde{x}}_w^{\mathit{sub}}\left(\mathrm{not}\right),\mathrm{not}=0,...,{\mathrm{NOT}}_s-1.$

It is inserted in place of the last Ns right components as shown in Figures 4a and 4b.

de signal filtré est placé dans sa partie droite. C'est ce qui est indiqué sur la figure 4b où le bloc analysé en 42 devient le bloc courant. On constate sur la figure 4b que l'identificateur inséré représenté en grisé en 41 s'est décalé de N_s échantillons vers la gauche par rapport au bloc précédent représenté en 42 sur la figure 4a (et qui était le bloc courant de l'étape précédente).To move to the new current block, the previous block is shifted to the left of Ns samples and the block ${\tilde{x}}_w^{\mathit{sub}}\left(\mathrm{not}\right),\mathrm{not}=0,...,{\mathrm{NOT}}_s-1.$

Filtered signal is placed in its right part. This is indicated in FIG. 4b where the block analyzed at 42 becomes the current block. It can be seen in FIG. 4b that the inserted identifier represented in gray at 41 shifted N _s samples to the left with respect to the previous block represented at 42 in FIG. 4a (and which was the current block of the step previous).

In this FIG. 4b, it can be seen that the correlation calculation gives a value greater than the threshold on the first Ns samples. This therefore indicates the position k _s of the beginning of the tattoo identifier. We can then deduce the end of the tattoo identifier knowing its length N to possibly detect a message that would have been inserted after the identifier.

With this overlay analysis method, an example of calculating the correlation is given by the following expression: $$\begin{array}{cc}\mathit{Horn}\left(k\right)={\displaystyle \underset{\mathrm{not}=0}{\overset{\mathrm{NOT}-1}{\Sigma }}}w\left(\mathrm{not}+k\right)x_w\left(\mathrm{not}\right)& \mathrm{k}\mathrm{=}\mathrm{0}\mathrm{,}\mathrm{...}\mathrm{,}\mathrm{NOT}\mathrm{-}\mathrm{1}\end{array}$$

The preceding equation can then be written in the following matrix form: $$\left[\begin{array}{c}\mathit{<figure-callout\; id="0"\; label="Horn"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">Horn</figure-callout>}\left(0\right)\\ \mathit{<figure-callout\; id="1"\; label="Horn"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">Horn</figure-callout>}\left(1\right)\\ \mathit{<figure-callout\; id="2"\; label="Horn"\; filenames="imgf0003.png"\; state="\{\{state\}\}">Horn</figure-callout>}\left(2\right)\\ \mathit{<figure-callout\; id="3"\; label="Horn"\; filenames="imgf0003.png"\; state="\{\{state\}\}">Horn</figure-callout>}\left(3\right)\\ \mathrm{.}\\ \mathit{Horn}\left(\mathrm{not}-2\right)\\ \mathit{Horn}\left(\mathrm{not}-1\right)\end{array}\right]=\left[\begin{array}{ccccc}w\left(0\right)& \mathrm{<figure-callout\; id="1"\; label="w"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">w</figure-callout>}\left(1\right)& \mathrm{.}& w\left(\mathrm{NOT}-2\right)& w\left(\mathrm{NOT}-1\right)\\ w\left(1\right)& w\left(2\right)& & w\left(\mathrm{NOT}-1\right)& w\left(0\right)\\ w\left(2\right)& \mathrm{<figure-callout\; id="3"\; label="w"\; filenames="imgf0003.png"\; state="\{\{state\}\}">w</figure-callout>}\left(3\right)& \mathrm{.}& w\left(0\right)& w\left(1\right)\\ w\left(3\right)& w\left(2\right)& & w\left(1\right)& \mathrm{<figure-callout\; id="2"\; label="w"\; filenames="imgf0003.png"\; state="\{\{state\}\}">w</figure-callout>}\left(2\right)\\ \mathrm{.}& \mathrm{.}& \mathrm{.}& \mathrm{.}& \mathrm{.}\\ w\left(\mathrm{NOT}-2\right)& w\left(\mathrm{NOT}-1\right)& \mathrm{.}& w\left(\mathrm{NOT}-4\right)& w\left(\mathrm{NOT}-3\right)\\ w\left(\mathrm{NOT}-1\right)& \mathrm{<figure-callout\; id="0"\; label="w"\; filenames="imgf0001.png,imgf0002.png"\; state="\{\{state\}\}">w</figure-callout>}\left(0\right)& \mathrm{.}& w\left(2\right)& w\left(\mathrm{NOT}-2\right)\end{array}\right]\left[\begin{array}{c}{x}_w\left(0\right)\\ x_w\left(1\right)\\ x_w\left(2\right)\\ {\mathrm{<figure-callout\; id="3"\; label="x\; w"\; filenames="imgf0003.png"\; state="\{\{state\}\}">x\; </figure-callout>}}_w\left(3\right)\\ \mathrm{.}\\ x_w\left(\mathrm{NOT}-2\right)\\ x_w\left(\mathrm{NOT}-1\right)\end{array}\right]$$

La matrice w _c est une matrice circulante construite à partir de sa première ligne w^T, T étant l'opérateur de transposition.In matrix notation: $$\widetilde{\mathrm{VS}}\mathit{gold}={\tilde{w}}_{\mathrm{vs}}{\tilde{x}}_w$$

The matrix w _c is a circulating matrix constructed from its first line w ^T , where T is the transposition operator.

Since the tattoo identifier is not inserted periodically, the lower half of the matrix w _c will give rise to parasite aliasing terms in the matrix product. To avoid this, the method according to the invention applies an overlap of N _s = N l R samples and calculates the maximum of the correlation coefficient on the N _s first values of the correlation vector. VS or as described above with reference to Figures 3, 4a and 4b.

If the sequence were inserted periodically, at any place where we would place ourselves in a tattooed signal block, we would have an entire random sequence of N points (actually one of the lines of the matrix w _c ). The optimal offset would be given by the row of the matrix which gives the maximum of the scalar product w _c x _w, equal to N in the ideal case. In our case, only one sequence of N samples is inserted at a time. For example, if we assume that the sequence starts at sample n = -1 in the tattooed signal, the maximum correlation will be found for the second row of the matrix w _c . There will be a parasite term given by the w (0) at the end of the line. For n = -2 the max is given by the third line. There are two parasitic terms w (1) and w (0) (reverse sequence) and so on. The use of the overlay makes it possible to limit the influence of parasitic terms by taking only the first Ns of the correlation.

In an example where R = 8192/1424 = 8, instead of having a correlation maximum of N we will have a maximum of N * (8192-1) / 8192 = N * 0.9998, if we have an offset of 1 and N * (8192-1024) / 8192 = N * 0.875 in the most unfavorable case of an offset of N _s = 1024.

To perform the scalar product efficiently, the sequence of the first row of the matrix w _c can be taken equal to a PN sequence of maximum length in which case we will have: $$\mathrm{NOT}=2^{\mathit{ms}}-1$$

The usual values of Ms are 12 and 13 which gives synchronization sequences of size 4095 and 8191.

With reference to FIG. 5 , we will now describe a device for inserting a tattoo identifier. This device comprises a module for obtaining a random sequence w (n) from at least one initialization value.

Thus, the random sequence can be obtained from a single initialization value, for example, a first key K _a . This first key K _a will allow in a first embodiment to directly generate the random sequence w (n) used to generate the tattoo identifier. In this case, the encryption module 53 is not present.

In a second embodiment, the encryption module 53 is present. The random sequence w (n) used to generate the tattoo identifier is obtained from two initialization values. The first is for example a first key K _a which will make it possible to generate an intermediate random sequence w _s (n), the second is a second key K _b which will make it possible to encrypt this intermediate random sequence to obtain the random sequence w (n ).

The random sequence w (n) obtained by the module 51 will serve as in the insertion device of Figure 1 to generate the tattoo identifier by the module 54. A psycho-acoustic model 55 is also taken into account as well as a gain 56 to be added at 57 to the audio signal s (n) and form the tattooed signal s _w (n).

FIG. 6 represents a device for detecting a tattoo identifier according to the invention. This detection device just like the device insertion includes a module for obtaining the random sequence that has been used to generate the tattoo identifier.

In the same way as for the insertion device, this random sequence obtaining module 61 can in a first embodiment comprise only a module for generating the random sequence w (n) from the first key K _has or comprise both the module 62 for generating an intermediate random sequence w _s (n) from the first key K _a and an encryption module 63 of this intermediate sequence from the second key K _b for get the random sequence w (n).

Thus, the detection device can detect the tattoo identifier only if it knows the one or two keys used to obtain the random sequence.

An example of a possible application in the embodiment with a single key is for example to associate with each audio content, a user or operator number as a tattoo identifier. The key will be deduced from this number through a correspondence table. Thus, in the context of secure online sales of audio content, the tattoo identifier makes it possible to detect transfers or abusive exchanges of content and to trace the origin of the problem. Indeed, in the presence of pirated content found on a website, the entity holding the user keys can test the detection of the tattoo identifier for each of the keys. The one that will detect the tattoo identifier will be the one that belongs to the user who is at the origin of the abuse. The detection being very fast, the search can be carried out quickly for a large number of keys.

In the case where the tattoo identifier is not used alone but is followed by a message, the key of the identifier makes it possible to affirm that the signal is tattooed when the presence of the identifier is detected. In this case, the possibility of detecting the message can be conditioned by the knowledge of the same key as that of the identifier. However, the identifier key and the message key may be different.

The embodiment with two keys, opens an even wider scope and provides additional security.

The key K _b may represent an order number on a user or dealer and the key K _may be specific to the owner of the work or a company operating for several operators. In this case the leakage of pirated content can be located as follows: given a pirated content, the trusted third party holder of the key K _has will search for all keys K _b, which gives a positive response detection and who will be at the origin of the leak?

In an access control application in a reader, the key K _a represents a user number and the key K _b , the number of the work. It is then possible to test if the work number of the right of access license corresponds to that which has actually been inserted by tattooing. If the test result is negative, then the content is hacked and the drive is blocked.

We can also notice that one of the keys can be public, for example K _b while keeping K _a secret. In this case, if the key K _b represents a content number, it makes it possible to give access to this content. The association of audio sequences with keys K _b may be known to all, for example via a database available on the internet.

The detection device also comprises a filter module referenced at 64 for whitening and equalizing the tattooed signal during reception and thus forming a signal vector x _w ^sub (n).

The signa s _w is first filtered by an inverse filter A ( z ) having coefficients only in the numerator. This filter is calculated from the correlation coefficients of the psychoacoustic mask of the signal by the Levinson-Durbin algorithm. Then from the signal s _wz well filtered, we will calculate a Wiener F _w (z) filter that will try to model w by linear filtering s _wz . In fact the minimization between the mark w and its estimated from the signal s _wz leads to the resolution of the following equations: $$R_{\mathit{wx}}{F}_w=R_{\mathit{dx}}$$

sera donné par: $$x_w^{\mathit{sub}}\left(n\right)={\displaystyle \sum _{j=-M_k}^{M_k}}{F}_w\left(j\right)s_{\mathit{wz}}\left(n-j\right)$$

Where R _wx is a Toëplitz matrix formed from the autocorrelation function of the signal s _wz and R _dx a correlation vector calculated from a training sequence. As a training sequence, the sequence w can be taken as an example from different starting or offset samples. The correlation function is then calculated for offsets ranging from 0 to M _k , where M _k is the number of filter coefficients. The resulting filtered signal $x_w^{\mathit{sub}}\left(\mathrm{not}\right)$

will be given by: $$x_w^{\mathit{sub}}\left(\mathrm{not}\right)={\displaystyle \underset{j=-M_k}{\overset{M_k}{\Sigma }}}{F}_w\left(j\right)s_{\mathit{wz}}\left(\mathrm{not}-j\right)$$

The use of a Wiener filter calculated from a training sequence makes it possible to improve the detection of the tattoo identifier with respect to a method using only inverse filtering.

Indeed, the training sequence makes it possible to take into account certain characteristics of the sequence w and thus to accentuate the possible peaks of detection.

The correlation detection module 65 comprises means for preparing a current block of N samples of the filtered tattooed signal, correlation calculation means between the samples of the current block and the random sequence w (n) and comparison means. correlation values obtained at a threshold S.

These means allow the implementation of the method described in Figure 3 according to the invention.

The encryption performed both in the module 53 and the module 63 on a random sequence w _s (n) can be performed by random permutation of the samples of the sequence w _s (n) or by an encryption method such as that described for example in the document B. Schneier "Applied Cryptography, Protocols and Sources in C". John Wiley & Sons, Inc., 1996 p397 and 398 .

An example of permutation encryption is illustrated by the algorithm shown in FIG .

The heart of the algorithm is to calculate a permutation matrix by means of a key of 16 bytes or 128 bits.

A TabS array is first initialized in step E71 at values from 0 to N-1. A 256-byte size TabK table is initialized at step E72 by repetition, as many as many times as necessary, the contents of the Key key table containing the 16 bytes of the key. The CurPos integer is initialized to 0.

In steps E73 to E75, for each index i from 0 to N-1, a CurPos index is generated randomly by addition of CurPos, TabS of index i and TabK of index i modulo 256, the TabK table being of size 256, the whole being taken modulo N. Then, in steps E77 to E79, the permutation table that will be used to scramble or encrypt the random sequence is obtained by permuting the two values it contains to the index curpos and to the index i with the aid of the intermediate variable TmpVal.

So we end up with a permutation table that contains for i = 0, ..., N -1, the rank of the sample to be swapped. To scramble the random sequence, it is now sufficient to read the permutation table TabS (i) for i = 0, N-1 and to swap the content of w _s of index i with the content of w _s of index TabS ( i). The encrypted sequence, calculated once and for all, can be stored in memory.

A method of generating a PN sequence is for example a method as described in the document B. Schneier. "Applied cryptography, protocols and sources in C". John Wiley & Sons, Inc., 1996, pages 372 to 375 . The loopback points of the register are given by the generator polynomial of the sequence.

By way of example, for the case Ms = 12, one of the primitive generator polynomials is given by poly (x) = 1 + x + x ^ 4 + x ^ 6 + x ^ 12 or poly = ca0 in hexadecimal or else 1100 1010 0000 in binary. The loopback points then correspond to the bits equal to 1 in the binary decomposition of poly.

One way to generate a PN sequence optimally with a minimum of bit operations is now described. The coefficients of the high power generating polynomial are in correspondence with the low order bits of the shift register, the output bit being in the high order.

   int state;
   N = 1 << Ms) - 1
   int poly = 0xca0;
   state = 1;
   w[0] = -2 * (state >> (Ms - 1) & 1) + 1;
   for(i = 1; i < N; i++)
    {
    state = ((state << 1) & N) | (xor(poly & state)&1);
    w[i] = -2 * (state >> (Ms - 1) & 1) + 1;
   }

The descriptive code for this method is given below:

 int state;
   N = 1 << Ms) - 1
   int poly = 0xca0;
   state = 1; 
   w [0] = -2 * (state >> (Ms - 1) & 1) + 1;
   for (i = 1; i <N; i ++)
    {
    state = ((state << 1) & N) | (xor (poly & state) &1);
    w [i] = -2 * (state >> (Ms - 1) & 1) + 1;
   } 

The integer state contains the current state of the shift register, each cell of the register being represented by a bit location of the integer state. It is initialized to 1. To calculate the state state at the next instant, the state content is shifted to the left of the bit position, then a mask is made by doing a binary operation "and" of the result with N the integer containing M _s bits of low weight to 1. It remains only to insert in the least significant bit position, by addition by means of a binary operation "or", the operation of " or exclusive '(xor) bits of the bitwise product of the register state with the polynomial generator (poly & state).

The function xor (n) calculates one or exclusive bits of the integer n as an argument.

The method described above makes it possible to obtain the sequence from an implementation of the shift register by a single line of code requiring only a few bit operations: an "offset" operation, three operations "and", an operation "or" and an "exclusive" operation regardless of the generator polynomial. Indeed, these operations are performed on the bits of an integer representing in each case the generation shift register of the sequence.

In the state of the art, for example the Schneier reference cited previously on page 375, the implementation of the Fobinacci shift register requires more binary operations: seven "offset" operations, two "and" operations, one operation "or" and five operations "or exclusive".

This PN sequence thus generated can be used for example in the first embodiment described with reference to FIG. 8.

FIG. 8 illustrates in detail, in flowchart form, the correlation calculation step E33 of FIG. 3 in the case of a first embodiment of the invention. This first embodiment uses the fast transform method of Hadamard as a method of rapid transformation.

Thus, at the end of step E32, a current block x _w ( n ) of N samples has been prepared.

The next operation E81 is to swap the vector x _w ( n ), that is to say the re index according to a first permutation Pe. At the end of this permutation step, we obtain a vector xe ( i ), i = 0, ..., N -1.

   xe[0] = 0;
   for (i = 0; i < N; i++)
     xe[TabPe[i]] = xw[i];
    }

An example code implementing this function is:

 xe [0] = 0;
   for (i = 0; i <N; i ++)
     xe [TabPe [i]] = xw [i];
    } 

The Hadamard transform is then applied to step E82 to the vector xe (i), i = 0, ..., N -1 . An explanation of Hadamard's method can be found in the paper A. Alrutz, MR Schroeder "A Fast Hadamard Transforms Method for the Evaluation of Measurement Using Pseudo Random Test Signals", Proceedings of the 11th Conference on Acoustics, Paris, 1983 .

 void FastHadamard(double *x, long N)
     {
     long i, il, j, k, k1, k2;
     double temp;
     Kp = (1<<N)-1;
     k1 = Kp;
     for (k = 0; k < N; k++)
        {
       k2 = k1 >> 1;
       for (j = 0; j < k2; j++)
           {
           for (i = j; i < Kp; i = i + k1)
               {
               il = i + k2;
               temp = x[i] + x [i1] ;
                xe[i1] = xe[i] - xe[i1];
               xe[i] = temp;
               }
           }
       kl = k1 >> 1;
       }
 }

An example of the procedure of the Hadamard transform is presented below:

 void FastHadamard (double * x, long N)
     {
     long i, he, j, k, k1, k2;
     double temp;
     Kp = (1 << N) -1;
     k1 = Kp;
     for (k = 0; k <N; k ++)
        {
       k2 = k1 >>1;
       for (j = 0; j <k2; j ++)
           {
           for (i = j; i <Kp; i = i + k1)
               {
               he = i + k2;
               temp = x [i] + x [i1];
                xe [i1] = xe [i] - xe [i1];
               xe [i] = temp;
               }
           }
       k1 = k1 >>1;
       }
 } 

Step E82 is followed by step E83 in which a second permutation Ps is applied to the vector resulting from the Hadamard transform, in order to obtain the correlation values corresponding to the order of the initial sampling times.

    for (i = 0; i < N-1; i++)
     {
      cor[i] = xe[TabPs[i]];
     }
    cor[N-1] = 0;

An example code implementing this function is:

 for (i = 0; i <N-1; i ++)
     {
      cor [i] = xe [TabPs [i]];
     }
    cor [N-1] = 0; 

At the end of step E83, correlation values Cor (k) are obtained which can be used directly in step E34 described with reference to FIG.

In an improved embodiment, a correction step E84 is implemented.

• des termes parasites à l'origine du repliement apparaissent à la fin de la matrice carrée. Il s'agit de w(0) sur la deuxième ligne, de w(0) et w(1) sur la troisième ligne, de w(0) w(1) et w(2) sur la quatrième ligne. La contribution C ^-(k) de ces termes parasites est à retrancher de Cor(k).
• des termes utiles manquent. En effet la séquence de la deuxième ligne commence à w(1), il manque donc la contribution de w(0). Celle de la troisième ligne commence à w(2) , il manque donc la contribution de w(0) et de w(1). La référence de temps étant prise à n=0, ces valeurs sont à pondérer avec le signal tatoué antérieur à la trame courante x_w (-1), x_w (-2), x_w (-3), ... et la contribution C⁺ (k) de ces termes manquants est à ajouter à Cor(k).

Indeed, by examining equation (4), we find that:

• parasitic terms at the origin of folding appear at the end of the square matrix. This is w (0) on the second line, w (0) and w (1) on the third line, w (0) w (1) and w (2) on the fourth line. The contribution C ^- ( k ) of these parasitic terms is to be subtracted from Cor ( k ).
• useful terms are missing. Indeed the sequence of the second line starts at w (1), so the contribution of w (0) is missing. That of the third line starts at w (2), so the contribution of w (0) and w (1) is missing. Since the time reference is taken at n = 0, these values are to be weighted with the tattooed signal prior to the current frame x _w (-1), x _w (-2), x _w (-3), ... and the contribution C ⁺ ( k ) of these missing terms is to be added to Cor ( k ) .

The contribution C ^- ( k ) is equal to: $${\mathrm{VS}}^{-}\left(k\right)={\displaystyle \underset{j=0}{\overset{k-1}{\Sigma }}}w\left(j\right)x_w\left(\mathrm{NOT}-k+j\right)$$

That of C ⁺ ( k ) is equal to: $${\mathrm{VS}}^{+}\left(k\right)={\displaystyle \underset{j=0}{\overset{k-1}{\Sigma }}}w\left(j\right)x_w\left(-k+j\right)$$

The corrected correlation value will be given as a function of the uncorrected intercorrelation Cor ^nc (k) by: $$\begin{array}{cc}\mathit{Horn}\left(k\right)={\mathit{Horn}}^{\mathrm{nc}}\left(k\right)+{\mathrm{VS}}^{+}\left(k\right)-{\mathrm{VS}}^{-}\left(k\right)& k=0,...,{\mathrm{NOT}}_s-1\end{array}$$

As we are only covering situation the N _s first cross-correlation coefficients are taken into account and are therefore correct.

The correction introduced by equation (9) is therefore applied to step E84. At the end of this step, corrected correlation values Cor (k) are compared with a threshold at step E34 as described with reference to FIG.

The permutations Pe and Ps are defined by respective tables TabPe and TabPs which are calculated at the initialization step E31 of FIG.

 N = (1 << Ms) - 1;
 state = 1;
 
 TabPe[0] state;
 
 for(i = 1; i < N; i++)
   {
   state = ((state << 1) & N) | (xor(poly & state) &1);
   TabPe[i] = state;
 }

The permutation table TabPe to be assigned to the values of x _w of the input signal vector is calculated for example by the code below according to the generator polynomial named "poly", the function "xor" and Ms the degree of the generator polynomial. This permutation table is then stored in memory.

 N = (1 << Ms) - 1;
 state = 1;
 
 TabPe [0] state;
 
 for (i = 1; i <N; i ++)
   {
   state = ((state << 1) & N) | (xor (poly & state) &1);
   TabPe [i] = state;
 } 

The state of the state register will take all values between 1 and N-1 but in an order that will be characterized by the permutation table. The state state is first initialized to 1. To calculate the state state at the next instant, the state bits at the current time are shifted to the left of a position and masked by the integer having M _s bits of low weight to 1. We add to the result of this operation, in the least significant bit position, the result of the exclusive or the product of the polynomial by the state (poly & state) masked by the integer 1. Here again the permutation sequence is calculated using the minimum number of bit operations as previously described for generation of the PN sequence.

 N = (1 << Ms) - 1;
 int state = 1 << (Ms - 1);// Initialisation du registre à décalage
 
 TabPs[0] = state;
 
 for(i = 1; i < N; i++)
   {
   state = (state >> 1) ^(poly * (state & 1));
   TabPs[i] =state;
 }

TabPs, the table of the permutation Ps to be assigned on the values of x e at the output of the Hadamard transform is calculated by the code shown below as a function of the generator polynomial "poly" and Ms the degree of the generator polynomial. This permutation table is then stored in memory.

 N = (1 << Ms) - 1;
 int state = 1 << (Ms - 1); // Initializing the shift register
 
 TabPs [0] = state;
 
 for (i = 1; i <N; i ++) 
   {
   state = (state >> 1) ^ (poly * (state &1));
   TabPs [i] = state;
 } 

The state variable of the state register is first initialized to 2 ^Ms -1. Then, to calculate the state state at the next instant, the state bits at the current time are first shifted one position to the right. The state at the next instant is the result of the operation "or exclusive" bit by bit of the result of the previous operation with the product of the polynomial by the state (poly & state) masked by the integer 1. Here again the permutation sequence is calculated using the minimum number of bit operations.

With reference to FIG. 9 , we will now describe a second embodiment for the correlation calculation step E33. In this embodiment, the fast transformation method is a Fourier transform method.

Thus, at the end of step E32, a current block x _w ( n ) of N samples was prepared.

The following operation E91 consists of calculating the Fourier transform of the vector x _w ( n ) to get the transform X _w .

La transformée de Fourier de Cor(k) est donnée par: $$\mathit{FFT}\left(\bar{C}\mathit{or}\right)={\bar{W}}^{*}{\bar{X}}_w$$

où W * et X _w sont respectivement les transformées de Fourier de w et de x _w . En conséquence: $$\bar{C}\mathit{or}={\mathit{FFT}}^{-1}\left(\bar{W}*{\bar{X}}_w\right)$$

In this embodiment, the calculation of the correlation is performed according to the following equation (10): $$\begin{array}{c}\mathit{Horn}\left(k\right)={\displaystyle \underset{\mathrm{not}=0}{\overset{\mathrm{NOT}-1}{\Sigma }}}w\left(\mathrm{not}\right)x_w\left(\mathrm{not}+k\right)\end{array}$$

The Fourier transform of Cor (k) is given by: $$\mathit{FFT}\left(\widetilde{\mathrm{VS}}\mathit{gold}\right)={\tilde{W}}^{*}{\tilde{X}}_w$$

or W * and X _w are respectively the Fourier transforms of w and of x _w . Consequently: $$\widetilde{\mathrm{VS}}\mathit{gold}={\mathit{FFT}}^{-1}\left(\tilde{W}*{\tilde{X}}_w\right)$$

The transform of the random sequence w , whether from a sequence generated with a single key or a sequence generated by a first key and encrypted by a second is either performed during step E91, or performed at the initialization step and then stored as for obtaining the random sequence.

In the case where the random sequence is encrypted, this sequence is not generated according to the PN sequence generation method described above but according to a random sequence generation based for example on a generation of random numbers.

During the initialization step E31, the conjugate complex W * of the Fourier transform W is also calculated and stored in memory.

In step E92, the product of W * X _w in the Fourier domain is performed. We obtain Y ( k ) = X _w ( k ) W * ( k ) for k ranging from 0 to N-1. Then, in step E93, an inverse Fourier transform of this product is applied to deduce the cross-correlation as described by equation (12).

The correlation values Cor (k) are thus obtained which are compared with a threshold for the first Ns samples as described in step E34 with reference to FIG.

As in the embodiment using the Hadamard transformation, parasitic terms appear, these terms being removable by applying the correction described by the above equations 7, 8 and 9.

This procedure requires a large FFT when the tattoo identifier has a large size, for example N = 4096 or N = 8192, common cases for digital audio signals.

To avoid this, a first variant embodiment of this second embodiment is now described with reference to FIG.

This solution consists of calculating the correlation for overlapping blocks by segmenting the calculation of the FFT into several smaller size FFTs, again in recovery. For example for a size of 8192, we can perform R = 8 FFT of size 1024 and sum the results of the partial intercorrelations.

et la séquence w(n) en sous-séquences w_i (n) telles que : $$\begin{array}{cc}{w}_i\left(n\right)=\{\begin{array}{cc}w\left(n+N_{s}i\right)& n=0,1,\cdots ,N_s-1\\ 0& n=N_s,\cdots ,2N_s-1\end{array}& i=0,1,\cdots ,R-1\end{array}$$

The proposed technique consists in subdividing the sequence x _w ( n ) into subsequences x _wi ( n ) which overlap as follows: $$x_{\mathit{wi}}\left(\mathrm{not}\right)=x_w\left(\mathrm{not}+{\mathrm{NOT}}_{s}i\right)\mathrm{not}=0,1,\cdots ,2{\mathrm{NOT}}_s,i=0,1,\cdots ,R-1$$

and the sequence w ( n ) in subsequences w _i ( n ) such that: $$\begin{array}{cc}{w}_i\left(\mathrm{not}\right)=\{\begin{array}{cc}w\left(\mathrm{not}+{\mathrm{NOT}}_{s}i\right)& \mathrm{not}=0,1,\cdots ,{\mathrm{NOT}}_s-1\\ 0& \mathrm{not}={\mathrm{NOT}}_s,\cdots ,2{\mathrm{NOT}}_s-1\end{array}& i=0,1,\cdots ,R-1\end{array}$$

We can notice that the sequence x _wi (n) i = 0, ..., R - 1 actually comprises N + N _s terms, the last subsequence being equal to x _{w R -1} ( n ) = x _w ( n + N _s ( R -1)) n = 0.1, ···, 2 N _s -1

For each pair of subsequences x _wi ( n ) and w _i ( n ) of length 2 N _s , the sequence Cor _i ( m ) representing the periodic correlation of x _wi ( n ) and w _i ( n ) is calculated. What is written with the original sequences x _w (n) and w ( n ): $$\begin{array}{cc}{\mathit{Horn}}_i={\displaystyle \underset{\mathrm{not}=0}{\overset{{\mathrm{NOT}}_s-1}{\Sigma }}}{x}_w\left(\mathrm{not}+i{\mathrm{NOT}}_s\right)w\left(\mathrm{not}+i{\mathrm{NOT}}_s+m\right)& m=0,1,\cdots ,{\mathrm{NOT}}_s\end{array}$$

Summing these partial correlations on the R sub-blocks, we obtain the total correlation between the two sequences. However, there is a more efficient method for calculating intercorrelations through the FFT.

For that we calculate X _wi ( k ) and W _i ( k ), the FFTs of the signals x _wi ( n ) and w _i (n) then the product: $$Y_i\left(k\right)={\tilde{W}}_i^{*}\left(k\right){\tilde{X}}_{\mathit{wi}}\left(k\right)$$

Summing up the contributions on the R subframes of Y _i in the frequency domain and performing an inverse Fourier transform, the desired cross-correlation is obtained: $$\widetilde{\mathrm{VS}}\mathit{gold}=\mathit{IFFT}\left({\displaystyle \underset{i=0}{\overset{R-1}{\Sigma }}}{\tilde{Y}}_i\right)$$

supposées avoir été calculées à l'initialisation par FFT de w _i défini par l'équation (15).Thus, with reference to FIG. 10 , in step E110 the sequences ${\tilde{W}}_i^{*}i=0,...,R-1$

assumed to have been calculated on initialization by FFT of w _i defined by equation (15).

à l'étape E140 et on somme le résultat dans le tableau de nombres complexes Z à l'étape E150. Lorsque les R contributions de ${\bar{Y}}_i=W_i^{*}{\bar{X}}_{\mathit{wi}}$

ont été sommées on sort de la boucle en réponse positive au test E160 et l'intercorrélation s'obtient en calculant la FFT inverse de Z à l'étape E170. La taille de la FFT inverse est de 2 N_s. Then, for each overlapping sub-block of i = 0 E120 at i = R-1 E160, the Fourier transform of x _wi ( n ) is calculated in step E130 and then the complex product is made. ${\tilde{Y}}_i=W_i^{*}{\tilde{X}}_{\mathit{wi}}$

in step E140 and sum the result in the array of complex numbers Z in step E150. When R contributions from ${\tilde{Y}}_i=W_i^{*}{\tilde{X}}_{\mathit{wi}}$

have been summed out of the loop in positive response to the E160 test and the cross-correlation is obtained by calculating the inverse FFT of Z in step E170. The size of the inverse FFT is 2 N _s.

We will now describe a second variant of the second embodiment of the correlation calculation step E33.

La transformée de Fourier de Cor sera alors donnée par la relation suivante: $$\mathit{FFT}\left(\bar{C}\mathit{or}\right)=\bar{W}{\bar{X}}_w^{*}$$

For this, the equation (19) is used for the calculation of the overlapping intercorrelation, the intercorrelation occurring this time as a function of the samples of the returned sequence of x _w (sign - of x _w ( nk )): $$\begin{array}{c}\mathit{Horn}\left(k\right)={\displaystyle \underset{\mathrm{not}=0}{\overset{\mathrm{NOT}-1}{\Sigma }}}w\left(\mathrm{not}\right)x_w\left(\mathrm{not}+k\right)\end{array}$$

The Fourier transform of VS gold will then be given by the following relation: $$\mathit{FFT}\left(\widetilde{\mathrm{VS}}\mathit{gold}\right)=\tilde{W}{\tilde{X}}_w^{*}$$

et un nouveau partionnement w dans lequel les zéros sont introduits au début de bloc et non plus à la fin comme dans l'équation (15): $$\begin{array}{cc}{w}_i\left(n\right)=\{\begin{array}{cc}0& n=0,1,\cdots ,N_s-1\\ w\left(n+N_{s}i\right)& n=N_s,\cdots ,2N_s-1\end{array}& i=0,1,\cdots ,R-1\end{array}$$

To have a value of the correlation without folding with the intercorrelation using the returned sequence, we must again use the recovery of: $$x_{\mathrm{wi}}\left(\mathrm{not}\right)=x_w\left(\mathrm{not}-{\mathrm{NOT}}_s+{\mathrm{NOT}}_{s}i\right)\mathrm{not}=0,1,\cdot \cdot \cdot ,2{\mathrm{NOT}}_s-1,\mathit{i}=0,\cdot \cdot \cdot ,R-1$$

and a new partitioning w in which the zeros are introduced at the beginning of the block and no longer at the end as in equation (15): $$\begin{array}{cc}{w}_i\left(\mathrm{not}\right)=\{\begin{array}{cc}0& \mathrm{not}=0,1,\cdots ,{\mathrm{NOT}}_s-1\\ w\left(\mathrm{not}+{\mathrm{NOT}}_{s}i\right)& \mathrm{not}={\mathrm{NOT}}_s,\cdots ,2{\mathrm{NOT}}_s-1\end{array}& i=0,1,\cdots ,R-1\end{array}$$

Note that in the present case the sequence x _wi ( n ) i = 0, ..., R -1 used for partitioning comprises the samples x _w (- N _s +1), ..., x _w (- 1), x _w (0), x _w (- N _s -1).

The calculation of the correlation will be given by the same method as that described with reference to FIG. 10, except that w _i is given by equation (22) and the array of complex Y _i is given in step E 140 by: $$Y_i\left(k\right)={\tilde{W}}_i\left(k\right){\tilde{X}}_{\mathit{wi}}^{*}\left(k\right)$$

According to an embodiment chosen and shown in FIG. 11, a device embodying the invention is for example a microcomputer 210 which comprises, in a known manner, notably a processing unit 220 equipped with a microprocessor, a read-only memory type ROM 230, RAM RAM type 240. The microcomputer 210 may include in a conventional and non-exhaustive manner the following elements: a keyboard, a screen, a microphone, a speaker, a communication interface, a disk drive, a storage medium ...

The ROM 230 includes registers storing a computer program PG including program instructions adapted to implement a tattoo identifier detection method according to the invention as described with reference to Figure 3. This program PG is thus adapted to prepare a current block of N samples of a tattooed signal received at input 250, to perform a correlation calculation by a method of rapid transformation between the samples of the current block and those of the random sequence obtained from at least one initialization value received at input 260 and comparing the resulting correlation values to a threshold to derive the position of the output tattoo identifier 270.

During power-up, the PG program stored in the read-only memory 230 is transferred into the random access memory which will then contain the executable code of the invention as well as registers for storing the variables necessary for the implementation of the invention. .

More generally, a means of storage, readable by a computer or by a microprocessor, whether or not integrated into the device, possibly removable, stores a program implementing the tattoo identifier detection method according to the invention.

Claims (14)

A method of detecting a tattoo identifier in an audio signal, the tattoo identifier having been inserted in the audio signal from a random sequence of a length of N samples, N being a multiple of Ns, obtained by at minus an initialization value, the method being characterized in that it comprises the following steps: preparing a current block of N samples from the tattooed signal; correlation calculation by a method of rapid transformation between the samples of the current block and those of the random sequence obtained from the at least one initialization value; comparing with a predetermined threshold the correlation values resulting from the correlation calculation of the first Ns samples; iteration of the preceding steps with a current block shifted with respect to the preceding block of Ns samples as long as the maximum of said correlation values is not greater than the predetermined threshold, the maximum of the correlation values greater than the predetermined threshold indicating the beginning of the tattoo identifier Method according to claim 1, characterized in that the tattoo identifier indicates the position in the audio signal of an inserted message. A method according to any one of claims 1 or 2, characterized in that the current block preparation step comprises a step of inverse filtering the tattooed signal and a Wiener filtering step using a filter calculated from a function correlation of a training sequence of the tattoo identifier. Method according to any one of claims 1 to 3, characterized in that the at least one initialization value is a key (K _a ) which makes it possible to generate the random sequence. Method according to any one of claims 1 to 3, characterized in that the random sequence is obtained by two initialization values, the first being a first key (K _a ) making it possible to generate an intermediate random sequence, the second being a second key (K _b ) for encrypting the intermediate random sequence to obtain the random sequence. Process according to any one of Claims 1 to 4, characterized in that the fast transformation method is a method of the Hadamard type which comprises the following steps: permutation of the samples of the current block according to a first permutation table; - Application of the Hadamard transform to the samples thus permuted; permutation of the samples of the vector resulting from the Hadamard transform according to a second permutation table to obtain the correlation values. Method according to claim 6, characterized in that it comprises a preliminary initialization step in which the first and second permutation tables as well as the random sequence are calculated and stored. Method according to claim 7, characterized in that the calculation of the random sequence, the first and the second permutation table is optimized by a minimum number of bit operations performed on the bits of an integer representing in each case the generation shift register of the sequence or table. Method according to any one of claims 1 to 5, characterized in that the method of rapid transformation is by a method of Fourier transform type which comprises the following steps: - applying a Fourier transform to the current block; calculating a product on the samples resulting from the transform of the current block and those of the conjugate complex of the Fourier transformed random sequence; - Application of a Fourier inverse transform to obtain the correlation values. Method according to claim 9, characterized in that it comprises an initialization step in which the Fourier transform of the random sequence and its conjugate complex are calculated and stored. Method according to one of Claims 1 to 5, characterized in that the fast transformation method is a Fourier transform method which comprises the following steps: segmentation into subsequences of the current block; segmentation into subsequences of the random sequence, the second half of the random subsequence being set to zero; - Application of a Fourier transform to the sub-sequences of the current block and the subsequences of the random sequence; calculating partial correlations by the product on the samples resulting from the transformation of the sub-sequences of the current block and those of the conjugate complex of the subsequences of the random sequence; summation of the calculated partial correlations; applying an inverse Fourier transform to the sum resulting from the summing step in order to obtain the correlation values. Method according to one of Claims 1 to 5, characterized in that the fast transformation method is a Fourier transform method which comprises the following steps: segmentation into subsequences of the current block; segmentation into subsequences of the random sequence, the first half of the random subsequence being set to zero; - Application of a Fourier transform to the sub-sequences of the current block and the subsequences of the random sequence; calculating partial correlations by the product on the conjugate complex of the samples resulting from the transformation of the subsequences of the current block and those of the samples resulting from the subsequences of the random sequence; summation of the calculated partial correlations; applying an inverse Fourier transform to the sum resulting from the summing step in order to obtain the correlation values. Device for detecting a tattoo identifier in an audio signal, the tattoo identifier having been inserted in the audio signal from a random sequence of a length of N samples, N being a multiple of Ns, obtained by at minus an initialization value, characterized in that it comprises: means for preparing a current block of N samples from the tattooed signal; correlation calculation means by a method of rapid transformation between the samples of the current block and those of the random sequence obtained from the at least one initialization value; means for comparing, at a predetermined threshold, the correlation values derived from the correlation calculation of the first Ns samples; means for obtaining a new current block by shifting Ns samples relative to the preceding block implemented as long as the maximum of said correlation values is not greater than the predetermined threshold. Computer program comprising code instructions for carrying out the steps of a conforming tattoo identifier detection method to one of claims 1 to 12, when said program is executed by a processor.

Images