DE102012012626B4 - Method for emphasizing and detecting repeated useful signal components within a source signal - Google Patents

Abstract

Method for emphasizing and detecting repeatedly occurring useful signal components within a source signal in which- a one-dimensional source signal q is present as a sequence of signal values, Absolute values are obtained, x(n) = |q(n)|, resulting in an absolute signal x,- that a version x1 of x shifted by a1 signal values is formed, x1(n) = x(n-a1),- that an intermediate signal y1 is formed by product formation point by point, y1(n) = x1(n)*x(n),- that a version x2 of the intermediate signal y1 shifted by a2 signal values is formed, x2(n) = y1(n-a2) ,- that a result signal y is formed by forming a minimum of x2 and y1 point by point, y(n) = Min(x2(n), y1(n)),- that candidate positions are determined from the result signal y by finding local maxima or local energy accumulations ,- that repetition positions in the source signal x are determined from each of the candidate positions, taking into account the distance values a1 and a2- that the repetition positions determined are output for further processing.

Description

The invention relates to a method for emphasizing and detecting useful signal components which occur repeatedly within a source signal - in a more or less similar form - according to the preamble of claim 1.

A fundamental problem in the field of digital signal processing is to detect all useful signal components potentially contained in a source signal.

The useful signal components of interest can be characterized using certain criteria such as duration, frequencies contained, bandwidths or volume curve. Alternatively, useful signal components of interest can be given using signal examples.

An important area of application for signal detection is speech signal processing, where the detection task can be to find specific words (keyword spotting) within a signal. Another area of application is acoustic environmental monitoring using microphones, where the occurrence of certain objects or circumstances is to be detected using characteristic noises, e.g. footsteps generated by walking movements, vehicle noises, engine noises, human speech, bird calls, machine damage.

In the field of radio signal analysis, the detection task can be to find specific radio transmissions. Considering multi-dimensional signals, there are numerous other applications. In this way, digital images can be understood as two-dimensional signals. A detection task then consists of finding objects within images. Examples are locating everyone in an image recorded by a surveillance camera, locating the license plate in an image of a motor vehicle, locating all buildings in an aerial photograph, etc.

A list of all objects detected in a source signal is usually expected as the solution to such detection tasks. In the following, these objects are also referred to as useful signals. It makes sense to specify each detected object in the form of its position within the source signal. The position can be specified in a one-dimensional signal by the start point (and possibly the end point), in a two-dimensional signal as a two-dimensional region - specifically as a rectangle - and for higher-dimensional signals accordingly by higher-dimensional sub-areas.

The application context underlying the method according to the invention involves a detection task in which the useful signals to be detected occur more than once in the source signal—in a more or less similar form—i.e. repeat themselves in the form of useful signal components, or signal components for short.

Examples are repeated calls for help in the case of voice detection, repeated calls of birds, signal noises or sequences of steps (one step is repeated) in the case of acoustic environmental monitoring, repeated transmissions of the same message for the radio example and aerial photographs showing a vehicle of a certain type several times for the two-dimensional example occurs.

If a signal component occurs twice (or more) times in a signal - in a more or less similar form - then a distance can be assigned to these occurrences by means of the difference in the respective starting points. In the same way, if a signal component occurs even more frequently, the distances of all possible pairs can be specified - for M occurrences, these are, for example, M*(M-1)/2 possible distances. Instead of distances one can also speak of offsets. For the case of higher dimensions one considers more generally offset vectors which give the distance in each dimension.

The method according to the invention relates in particular to such detection tasks in which certain distances between repeated useful signal components in the source signal are known in advance. This is advantageously used by the method according to the invention for the purpose of emphasizing and detecting these useful signal components, which are henceforth referred to as “repeated”, within the source signal.

If one knows, for example, the normalized speed of a specific machine and knows that certain machine defects always express themselves acoustically after one revolution in the same form, a (time) interval can be derived from this and used with the method according to the invention for finding machine defects. Another example is the detection of interference signals in the radio spectrum, which are based on characteristic overtone spectra, so that signal components of a specific interference are always repeated at a known frequency spacing.

State of the art

In the following, methods for emphasizing and detecting repeated useful signal components, also called repeated signal components, within source signals after the State of the art presented. Such methods are, for example, already from the DE 1 900 289 B famous.

Characteristic of a large number of methods for detecting repeated signal components according to the prior art is a procedure in which first repeated occurrences of a signal component are amplified (in the sense of worked out, emphasized), thus an intermediate signal is derived, and subsequently such an amplification step from the Intermediate signal significant signal values, ie indications of a signal repetition, are determined.

An important example of such methods are correlative methods, in particular autocorrelation methods. Autocorrelation, which can be looked up in any standard work on signal processing, is based on multiplying a given useful signal with versions of the same useful signal that have been shifted by certain integral distance values a and summing up the resulting product signal, AKF[x] (a) = ... x(0)* x(a) + x(1)*x(1+a) + x(2)*x(2+a) +... , where the sum is determined over a suitable sub-range of the source signal x. In general, a sequence of distance values a, for example a equal to 0, 1, 2, . . . , n, is run through systematically for this purpose. In the resulting autocorrelation signal AKF[x] to x, consisting of the values AKF[x] (0), AKF[x] (1),... AKF[x] (n), indicate particularly pronounced values, recognizable in general by local maxima, to signal components repeated in the source signal x: if AKF[x](a) is particularly pronounced, there may be a repetition of a signal component at a distance a in the source signal x. This makes itself, for example, in the EP 1 543 665 B1 disclosed methods.

A well-known advantage of autocorrelation is that it is very robust in the sense that even extremely weak repetitions can be detected (e.g. compared to a noise level). Efficient methods for determining correlations are, for example, from DE 199 18 610 A1 famous. In addition, correlation functions can generally be calculated very quickly using fast Fourier transformations. Transferred to the task on which the method according to the invention is based, in which the distance values a1, a2, , i.e. AKF[x] (a1), AKF[x] (a2),.... The assessment of the distinctiveness of these values can then be used to assess whether signal components are repeated in the distances a1, a2 in the source signal. A disadvantage of the prior art is that the signal components that may be present must be localized in some other way.

In the DE 1 900 289 B For the detection of periodically (ie continuously) recurring signal components with a fixed, previously known distance a, signal samples are continuously taken at specific intervals and in large numbers and accumulated. Thus, after DE 1 900 289 B be regarded as a variant of a correlative method.

From the DE 35 06 449 C1 a method is also known which implements an efficient form of a correlator by utilizing the intervals between periodic signal pulses.
A large number of other variants of correlative methods are known. Thus, the autocorrelation can be applied to the Fourier transform, also called spectral signal, Dx of a source signal x in order to detect repetitions within the spectral signal Dx. Examples of such spectral repetitions are overtones or harmonics of a fundamental, for example occurring in speech or music signals. Further examples are so-called multi-carrier signals in the field of radio signals.

In some cases, the Fourier transform of a source signal can also be used by itself to detect the occurrence - in a more or less similar form - of repeated signal components. This is preferably done in such cases when a particularly large number of repetitions and associated distance values a1, a2,... of a signal component are known and the distance values are also integer multiples of one another, for example a2=k*a1, where k is a positive whole number, it being particularly preferred if all distance values match. In contrast to (auto) correlation, which is particularly suitable for detecting repeated signal components, the Fourier transformation is particularly suitable for detecting periodic repetitions, ie repetitions that are always repeated after a certain distance value. If there are a particularly large number of repetitions of a signal component in a source signal analyzed by means of the Fourier transformation, this can be read off at specific points of the Fourier transformation using methods according to the prior art.

Numerous combinations of correlative and Fourier transformation methods are also known, for example the so-called spectral correlation or cyclic autocorrelation [Gardner 1986]. Furthermore, the application of correlative Techniques based on the so-called windowed Fourier transform, for example with the aim of recognizing repeated components of animal voices in the field of bioacoustic pattern recognition [Bardeli et al. 2008].

A method for visualizing repeated signal components that is widespread in the field of audio signal processing is based on the use of so-called similarity matrices. For this purpose, all positions of a source signal x = x (1) , x (2) ,..., x (n) are compared in pairs, resulting in a (symmetrical) matrix S whose position (i,j) has a similarity value S (i,j) between x(i) and x(j).

Of course, self-similarity matrices can be obtained not only directly for source signals x, but also in a particularly advantageous manner for any secondary representations of the source signal derived from the source signals by means of feature extraction. It is characteristic of self-similarity matrices S that repeated signal parts are expressed in the form of clearly prominent secondary diagonals. Such an off-diagonal can be used to determine the beginning and end of each repetition of a signal part or signal component [Müller 2007].

In relation to the task on which the method according to the invention is based, self-similarity matrices can be used, with corresponding secondary diagonals with distances corresponding to a1, a2, . . a signal component, need to be examined.

Preferably in such cases in which signal examples of the repeated signal component are known, methods of pattern recognition (pattern matching) known from the prior art can be used to detect repeatedly occurring signal components. A large class of such methods is known which locates all occurrences of a given portion of a signal (called a pattern) in a given target signal or set of target signals and reports all corresponding occurrences. The localization is carried out by specifying position information for each place of discovery. It is also said that the pattern is contained in a target signal at one point as a partial signal. Variants of this task allow certain, pre-configurable deviations between pattern and source or only ask for the best possible source with regard to a given quality measure. An overview of such prior art methods can be found in [Ouyang et al. 2012].

Another class of pattern recognition methods specializes in providing a particularly robust detection performance, especially in the event that the pattern and location can differ significantly from one another due to the application. One class of such methods uses a so-called generative approach to generate a representative model in a pre-processing step based on a representative set of existing patterns of the signal components of interest. Such a model is then matched to a source signal, with repeated occurrences of the pattern being reported as those locations that best match the pre-built model. A very common example in audio signal processing are the Hidden Markov Models (HMMs) [Rabiner 1989].

A method according to the prior art for specifically finding signal components occurring in a certain arrangement (called a constellation) in a signal is described in [Clausen et al. 2003]. The corresponding technique is called constellation search. When applied to useful signals, this method is based on generating signal features of different types from each of the signals involved (source signals) in a pre-processing step according to a specified rule, which are stored in so-called inverted lists. With the help of these lists and suitable intersection formation, an existing known signal component, from which a set of features was also generated by means of feature extraction, called query constellation, can then be efficiently localized within the signals involved, with a hit point being returned for each localization. A hit position clearly describes how the query constellation is to be shifted so that it matches the localized position in the relevant source signal.

Transferred to the present application scenario of the method according to the invention, the known distances a1, a2, . . . of signal components expected in the useful signal can be used to generate a suitable query constellation. Unfortunately, the procedure just outlined for using the constellation search cannot solve the present task with sufficient accuracy. However, certain principles on which the search for a constellation is based, in a modified form and in a form suitably combined with completely different characterizing properties, can contribute to the solution of the object of the method according to the invention.

Disadvantages of the Prior Art

Although the prior art methods perform well in many cases, there are several situations where their performance is limited. Such situations are in particular the superimposition of one or more useful signals (or useful signal components) or the superimposition of a useful signal (or useful signal components) with strong interference or noise components.

A fundamental disadvantage of correlative methods is that although they can be used to detect the distances between the positions of the signal components that are characteristic of repeated signal components, no direct statement can be made about the positions of the repeated signal components within the source signal. Such a localization of the signal components must therefore be carried out in some other way. Another disadvantage is that a multiple presence of repeated signal components cannot be distinguished from a single presence. A major disadvantage of purely correlative methods is that although repeated signal components are amplified, other signal components, such as those not repeated at the specified interval, are generally retained. While such components are often attenuated compared to the repeated signal components, they can still be reported as a miss in further processing.
A disadvantage of using the Fourier transformation to detect repeated signal components is the lack of possibility of directly determining the positions of the detected signal components and the implicit requirement that multiple, preferably periodic, repetitions are present. Thus, the Fourier transform is generally not particularly well suited to detecting only a few, say 2, 3 or 4 repeats present, especially if the corresponding distance values are not some integer multiple of each other.

Compared to pure correlation methods, self-similarity matrices are more suitable for determining the positions of individual repeated signal components. If, as is known, signal components are repeated several times, as in the task on which the method according to the invention is based, then when using self-similarity matrices for each repetition the presence of a pronounced secondary diagonal at the specified distance can be seen. In the case of weak repetitions, however, this process can be unreliable and thus lead to missing detections. Furthermore, the use of self-similarity matrices on individual signal values is not robust, since these are often mixed with noise, so that ultimately no clear (diagonal) structures are formed in the self-similarity matrix. For this reason, in most cases it is necessary to use pre-processing of the source signal to extract signal features that are adapted to the task. These features usually combine several adjacent values of the source signal into a single signal feature, which is then intended to provide a more meaningful comparison value within the similarity matrix. In some applications, however, this is not possible or no meaningful signal characteristics are known. A major disadvantage of the methods of pattern recognition mentioned according to the prior art is that the search is primarily based on examples, with one or more patterns that are as representative as possible of a signal component to be detected having to be present. The more robust the detection performance is desired, the more numerous and the more representative the available samples must be. A large number of patterns and a high level of robustness are often associated with reduced performance.

In principle, the constellation search method allows certain previously known properties, here distance values, relating to expected repetitions of signal components within a source signal to be modeled. However, the constellation search can only be applied to signal-based data in a practicable form, i.e. sufficiently efficiently, after pre-processing, in which a source signal is typically subjected to feature extraction according to certain criteria. The determination of hits takes place in a set-theoretical manner by forming intersections and combining so-called inverted lists. Furthermore, to reduce erroneous detections, the constellation search requires a certain size of the query constellation. Therefore, in the present application, where only a few, in particular only two, distance values are available for the possible generation of a query constellation, unfortunately a method for detecting repeated signal components within a source signal cannot be derived directly from the constellation search.

The method according to the invention goes one step further in principle than the prior art presented, in that it combines certain components of methods for constellation search in a modified form with parts of the autocorrelation, but instead of set operations (constellation search) and product sums (autocorrelation) the complementary, nonlinear arithmetic operations of product and minimum formation within an iterative process.

In addition, existing prior knowledge is used through the systematic inclusion of known properties of the expected signal repetitions in the form of distance values, but without the need, which is typical for methods of pattern recognition, of having to describe the signal components of interest by means of one or more signal examples. As a result, the stated disadvantages of the prior art can be avoided to a large extent.

object of the invention

In the light of the prior art presented, the object of the method according to the invention is to provide a method which enables useful signal components which occur repeatedly within a source signal to be suitably emphasized and subsequently detected. For this purpose, use is made of the fact that distances between repeatedly occurring useful signal components are known in advance.

solution of the task

The object is solved by the features of a method according to claim 1. Advantageous developments and refinements of the method according to the invention result from the dependent claims.

Description of the procedure

The invention makes use of properties of methods according to the preamble of claim 1.

For better orientation, but not restrictively, a scheme of the method according to the invention described below is shown in 1 shown. The method according to the invention processes a digital source signal q and two previously known distance values a1 and a2, which are given in the form of integer values. Given a1 and a2 as real numbers, appropriate rounding to the nearest integer is performed. In the following description it is assumed that a sequence of m signal values is given. This is not a limitation, however, since the method according to the invention can be carried out for signals of any length. The signal values can be real or complex numbers. In the first step, an absolute signal x (2) is formed from the source signal q (1) by forming absolute amounts point by point, ie x(1) = |q(1)|,...,x(m) = |q(m). )|. In the subsequent step (3), a shifted absolute signal x1(4) is formed from the absolute signal x(2) using the first distance value a1. Here, the signal value of x1 at point n is formed from the signal value of x at the point shifted by a1 by the rule x1(n)=x(n−a1). For values of n less than or equal to a1, i.e. for n-a1 less than zero, x1(n-a1)=0 is set without restriction and as a special case. Since the signal length m is significantly greater than a1 in preferred application scenarios, this special case affects relatively few values. Otherwise, the treatment of this special case is of a purely technical nature, serves solely to simplify the following description, more precisely to enable the use of uniform signal lengths m, and has no further significance for the principle of the method according to the invention; the signal values treated in this way can be neglected without affecting the result signal (8) finally determined.

In a subsequent step, an intermediate signal y1 (5) is formed from the absolute signal (2) and the shifted absolute signal (4) using point-by-point multiplication. In formulas, the values of the intermediate signal are formed by y1(1)=x1(1)*x(1),..., y1(m)=x1(m)*x(m).

In a further step (6), a shifted intermediate signal x2 (7) is then formed from the intermediate signal (5) using the second predetermined distance value a2, the values of the shifted intermediate signal x2 (7) being derived from the values of the intermediate signal y1 (5th ) by the rule x2(n)=y1(n-a2). As in the case of the determination of the shifted absolute signal, x2(n)=0 is selected as a special case and without restriction if n-a2 is less than or equal to zero in the latter rule. Everything mentioned above applies accordingly to this special case.

A result signal y (8) is then formed from the intermediate signal y1 (5) and the shifted intermediate signal x2 (7) by forming a minimum point by point, in formulas y(1)=Min(x2(1),y1(1)),.. .,y(m)=Min(x2(m),y1(m)), where the Min function returns the minimum value of two values given as an argument, as is usual in the prior art.

Within the result signal y (8), in a further step (9), a finite sequence of candidate positions (10) denoted by k(1), k(2),... is determined, with each of the candidate positions being an integer from the designated interval from 1 to m. Each candidate position can thus be interpreted as position information within the result signal. The candidate positions k(1), k(2), den, ie y(k(1)), y(k(2)),... are particularly large. The candidate positions k(i) are preferably formed here as positions of local maxima of y. Local maxima can be defined with respect to a locality parameter L in such a way that k(i) is called a local maximum if and only if y(k(i)) > y(r) for all r such that the absolute value of the difference k(i )-r is less than L, in formulas if |k(i) - r| < L. Alternatively, k(i) can be called a local maximum if the left-sided sequence y(k(i)-L),...,y(k(i)-1),y(k(i)) is a strictly increasing sequence of numbers and at the same time the right-hand sequence y(k(i),y(k(i)+1),...,y(k(i)+L) is a strictly increasing sequence of numbers The locality parameter is chosen appropriately depending on the area of application. In addition to the methods mentioned for determining candidate positions, numerous other methods according to the prior art can also be used for determining such candidate positions. Methods for signal segmentation, in particular for so-called spectral signal segmentation, should be particularly emphasized. Embodiments of such methods determine candidate positions by determining signal sections y(s),y(s+1),...,y(s+N) of length N+1 in which the sum y(s)+y(s +1)+...+y(s+N) of the signal values is particularly large, in particular larger than in neighboring signal sections Other dimensions, such as the sum of the squares of the absolute values, are also possible for the sum of the values. Another class of state-of-the-art methods for determining candidate positions are methods for so-called signal peak or peak extraction, such as those that calculate local maxima in so-called novelty curves [Foote 2000], which are extracted from the resultant signal using suitable pre-processing according to the state of the art.

Finally, in a last step (11), a set of repetition positions w(1), w(2),... (12) is determined from the candidate positions k(1), k(2),... (10). , where the known distance values a1, a2 can advantageously be included. If the distance values are the same, ie a1=a2, a preferred choice is w(n)=k(n)-a1 for all values n (corresponding to w(n)=k(n)-a2). If a1 and a2 do not match, and if the candidate positions include values k(i) and k(j) for which k(i)=k(j)+a1 applies exactly or approximately, then according to claim 2 there is one preferred choice for repeat positions associated with k(i) and k(j) in the three positions k(j), k(j)-a2 and k(j)-a2-a1. In the case of all other determined candidate positions k(i), the four positions k(i), k(i)-a1, k(i)-a2 and k(i)-a1-a2 can preferably be output as associated repetition positions .

In an alternative embodiment of the invention, which can be advantageous for certain applications, the steps of point-by-point minimum formation and point-by-point product formation are interchanged. This means that, firstly, an intermediate signal y1 (5) is formed from the absolute signal x (2) and the shifted absolute signal x1 (4) by forming a minimum point by point, in formulas y1(n) = Min(x(n), x1(n) ) for each value of n. Second, a result signal y (8) is formed from the intermediate signal y1 (5) and the shifted intermediate signal x2 (7) by pointwise product formation, in formulas y(n)=y1(n)*x2(n ) for each value of n. If the distance values a1 and a2 are the same, i.e. a1=a2, according to claim 5 a preferred choice is w(n)=k(n)-a1 for all values n (corresponding to w(n )=k(n)-a2). If a1 and a2 do not match, and if the candidate positions include values k(i) and k(j) for which k(i)=k(j)+a2 applies exactly or approximately, then according to claim 4 there is one preferred choice for repeat positions associated with k(i) and k(j) in the three positions k(i), k(i)-a2 and k(i)-a2-a1. In the case of all other determined candidate positions k(i), the four positions k(i), k(i)-a1, k(i)-a2 and k(i)-a1-a2 can preferably be output as associated repetition positions .

If more than two distance values are known, i.e. generally K distance values a1,...,aK, a sequence of K intermediate signals y1,...,y(K+1) can be formed according to claim 7, wherein in each of K steps one of the intermediate signals y(i+1) is formed by one of the operations product formation or minimum formation from the preceding intermediate signal yi and the shifted intermediate signal yi(n-ai). This procedure is detailed in 2 shown. For reasons of simplification, but not limitation, the absolute signal x(2) is converted into a first intermediate signal y1(2a) without any further changes. This is immediately converted (3) into a shifted intermediate signal x1 (4) using the first distance value a1. A second intermediate signal y2 (5) is then determined with prior selection of one of the operations of forming a minimum or forming a product. If the formation of the minimum operation is selected, the second intermediate signal is determined as the minimum y2(n)=Min(y1(n), x1(n)) for all values of n. If the formation of the product operation is selected, the second intermediate signal is determined as the product y2 (n)=y1(n)*x1(n) for all values of n. In general, an intermediate signal yi is determined in the ith step after selecting one of the operations minimum formation or product formation. If the operation minimum formation is selected, the i-th intermediate signal is determined as the minimum yi(n)=Min(y(i-1) (n), x(i-1) (n)) for all values of n If the product formation operation is selected, the second intermediate signal is determined as the product yi(n)=y(i-1) (n)*x(i-1) (n) for all values of n. In 2 is chosen as an embodiment K=4, resulting in a sequence of five intermediate signals. It is characteristic of the method according to the invention that in this procedure at least one of the intermediate signals is formed by product formation and at least one of the intermediate signals is formed by minimum formation. Subsequent to the step of determining candidate positions, repeat positions can preferably be determined as follows: In the event that all distance values match, ie a1=a2=...=aK, if K=2*M is an even number, for each candidate position k as a repetition position the number k-a1*M is a preferred choice. If K=2*M+1 is an odd number, one of the numbers k-a1*M or k-a1*(M+1) is preferable for each candidate position k as a repetition position. In the event that the distance values a1, a2,..., aK do not match, such repetition positions are preferably to be determined for a candidate position k that result from k by subtracting one or more correction factors, each of which is the sum of one or more of the Distance positions is revealed. The exact selection of the correction factors depends on the selected spacing positions and can be derived in an obvious manner in individual cases, taking into account the disclosed peculiarities of the method according to the invention.

Instead of one-dimensional signals, as described above, the method according to the invention can also be applied to multidimensional signals with a positive whole number D of dimensions. Instead of a sequence of signal values given for a one-dimensional signal x = x(1), x(2) , .... , x (n), where the position specification j, j is an integer, the signal value x(j ) is specified at position j, a D-dimensional signal occurs in a D-dimensional sequence of signal values x (n1, ..., nD) , each of which is specified by specifying position information n1, ..., consisting of D integers. nD is specified. In an application example with D=2, one obtains (two-dimensional) images and x(n1, n2) denotes the signal value at the x-coordinate n1 and the y-coordinate n2 of the image. In an application example, x(n1, n2) can be a grayscale value, alternatively a color value. In an application example with D=3, an image sequence or also a film can be described, with the argument n3 representing a time specification in a signal value x(n1, n2, n3) in addition to the last-mentioned application example. A D-dimensional signal x can be converted into a signal shifted by a in the j-th dimension by specifying a distance value a and a dimension component j, where j is an integer greater than or equal to 1 and less than or equal to D. In this shifted signal, the value x(n1,...,nj-a,...,nD) now takes the place of a signal value x(n1,...,nj,...,nD). The method according to the invention can thus be extended to the case of multidimensional signals, the method according to claim 7 being expanded in such a way that in addition to distance values a1, ..., aK, a specification list d1, ..., dK of dimensions to be selected is now provided, with each i the value di is greater than or equal to 1 and less than or equal to D is chosen. The intermediate signal y(i+1) is now formed from the intermediate signal yi and a version of the intermediate signal yi shifted by ai in the di-th dimension. In formulas, y(i+1) becomes in the case of a product formation by y(i+1) (n1,...,nD) = Min (yi (n1,..., nj,..., nD) , yi (n1, .... nj-ai,..., nD) ) and in case of a product formation by y(i+1) (n1,...,nD) = yi (n1, ..., nj , ...,nD) *yi (n1,...,nj-ai,...,nD) formed. The decision as to whether a product is formed or a minimum is formed in the i-th step is made in advance on the basis of the given task. In principle, all 2^K conceivable combinations of product and minimum formation are possible. In the process according to the invention, however, particular preference is given to sequences of product and minimum formation which contain at least one product formation and at least one minimum formation. As in the one-dimensional case, candidate positions in the result signal can be found using state-of-the-art methods for detecting local maxima or energy accumulations. The determination of repeat positions can be made in an obvious manner by a person skilled in the art considering the principles disclosed in claims 2, 4, 5, 6 and 8, considering that these principles are applied independently for each of the D dimensions. Another obvious variant of the method according to the invention consists in considering vector-valued distance values (A1, . . . , AD) instead of a distance value a present as a number when D-dimensional signals are present. Intermediate signals y(i+1) are thus formed from previous intermediate signals yi and (A1,...,AD)-shifted versions, with each of the values y(i+1) (n1, ..., nj , .. ., nD) by multiplying yi (n1,...,nj,...,nD) *yi (n1-A1,...,nD-AD) or, in the case of minimum formation, by Min(yi( n1,...,nj,...,nD) , yi(n1-A1,...,nD-AD)), claim 15.

In numerous technical applications, the source signals are so-called time signals q=q(1),...,q(m), where q(1),...,q(m) is interpreted as a time sequence of signal values. However, the method according to the invention is independent of such an interpretation and can be applied to any signals specified as a sequence or any other function. A special case that is important for many technical applications is when the source signal is a spectral signal. Instead of a chronological sequence of signal values, a sequence of frequency values. 3 illustrates this for the case that is particularly preferred in applications, in which a temporal source signal q(1) is obtained by using a spectral transformation D according to the prior art, preferably a discrete Fourier transformation (DFT), also preferably executed as a fast Fourier transformation (FFT), in one first step into a Fourier transform Dq = Dq(1),...,Dq(m) (2), which in turn is converted into an absolute signal x=x(1),..., using methods according to the state of the art. x(m) (2a), with formation of the absolute value x(i)=|Dq(i) |, which can be further processed with the method according to the invention, so that finally a result signal y=y(1) ,..., y(m) (8) arises.

In a particularly advantageous embodiment for certain applications according to claim 10, see 3 , a second result signal y′(12) is determined from the result signal y (8) before the step of determining candidate positions, from which the candidate positions can then be determined as explained above. To determine the second result signal, a spectrogram (sometimes also called a sonagram or windowed Fourier transform) S =S(1),...,S(r) (10) formed. Here, window functions according to the prior art can be selected, preferably rectangular, Hann, Hamming, or triangular windows. Furthermore, according to the state of the art and according to the given application, a window width F and an increment S must be selected. The entry S(1) of the spectrogram then results from the result signal y using the vector S(1)=D(y(1) ,...,y(1+F-1)), where D is the discrete Fourier transformation and y(1),..., y(1+F-1) a section of the result signal y of the length F corresponding to the window width. The entry S(2) of the spectrogram then results from the result signal y using the vector S( 2) = D(y(1+S),..., y(1+S+F-1)), where y(1+S),..., y(1+S+F-1) is a section of the result signal y of the length F corresponding to the window width, shifted by the increment S. Correspondingly, each of the F-dimensional vectors S(i) results in S(i) = D(y(1+i* (S-1)) , ..., y(1+ i*(S-1)+ F-1) In a further step (11), the one-dimensional second result signal y′=y′(1),...,y′(r) is formed from the vector-valued spectrogram (10), the value y being preferred '(i) as energy or absolute value - determined mathematically by any length value using known vector norms In a further preferred embodiment, the value y'(i) is determined as a similarity or also as a distance value of neighboring vectors of S, preferably S(i) and S(i-1), or alternatively S(i) and S(i+1), Preferred similarity values are dot products, normalized dot products, angular values and Euclidean distances.

An essential capability of the method according to the invention is the emphasizing of useful signal components repeated in the source signal within the result signal determined by means of the method according to the invention or within a second result signal. The emphasized useful signal components are recognizable in the result signal as particularly strong signal values of the result signal that emerge at specific positions. Such positions are usually given by particularly large individual signal values or by a particularly large signal energy (sum or square sum of the absolute signal values) within an interval of signal values that is short compared to the signal length. The positions of such signal values can be determined using methods known from the prior art for finding local maxima or local energy accumulations and are further processed in the method according to the invention in the form of the candidate positions mentioned above. Numerous other methods for determining candidate positions are known from various areas of signal processing and pattern recognition. Particularly preferred methods are methods for so-called spectral segmentation, which are known from the field of application of radio signal processing, for example from DE 42 43 113 C2 and the DE 32 20 073 C1 . Further preferred methods are known from the field of audio signal processing, for example general methods for the detection of pulse-like signals [Kaupinnen 2002] or signals for the detection of onset times in music signals [Bello et al 2005]. Such methods can be modified with prior art methods so that they can be used to find candidate positions within the method of the invention.

Instead of the two-place mathematical operations of product and minimum formation, alternative two-place operations can also be used. In this case, instead of forming the product, those operations can preferably be used which, for two input numerical values, determine a result numerical value which is only small in terms of magnitude if both of the input numerical values are small in terms of magnitude and otherwise large in terms of magnitude. Preferred examples of such two-digit operations are the summation of two numbers and the maximum of two numbers. Furthermore, instead of forming the minimum, those operations can preferably be used which, for two input numerical values, determine a result numerical value which is only large in terms of order of magnitude if both of the input numerical values are of the same order of magnitude are large and otherwise small in magnitude.

It is also possible to use only approximately known distance values a. Thus, in the step of determining each intermediate signal, instead of determining a single intermediate signal with respect to a predetermined distance value a, a set of intermediate signals can now be formed for a set of distance values similar to a. Of these, only that intermediate signal which has particularly pronounced signal components can preferably be used in the further processing. In this case, similar distance values to a are those of a similar order of magnitude, for example values deviating only by small whole numbers.

Advantages of the Invention

Advantages of the method according to the invention are an extremely high level of robustness in the sense that—in a more or less similar form—repeatedly occurring useful signal components can still be detected and localized even if they are superimposed by very strong noise and/or by others of interest, signal components are superimposed. Furthermore, the method according to the invention can be applied universally to, in principle, all types of signals and to multidimensional signals. Of course, the method according to the invention can obviously also be applied to feature representations derived from signals, in particular feature sequences, since these in turn can be interpreted as signals.

Furthermore, the method can be implemented very efficiently and is therefore also very suitable for hardware-related realizations, for example on GPU-based (Graphics Processing Unit) architectures. In this respect, the invention also includes a device for carrying out the method according to the invention or its implementation variants. Also encompassed by the invention is a software program product which e.g. B. programmed accordingly a digital signal processor (DSP) or controls its function.

Significant advantages of the method according to the invention compared to the prior art are set out below. Compared to correlation-based methods, in particular autocorrelation methods, the method according to the invention offers the advantage that not only can it be used to detect the presence of a repeated signal component with predetermined distance values, but also—compared to the prior art—the positions of the repeatedly occurring signal components within the source signal can be determined be able.

It is also advantageous that multiple, more precisely all, occurrences of repeated signal components with the specified distance values can also be determined together with their positions. To avoid erroneous detection reports, it is also very advantageous that the method according to the invention, in contrast to purely correlative methods, specifically weakens or completely suppresses signal components that are not of interest - through the formation of the minimum - but at the same time the advantageous property of correlation-based methods, namely repetitions of interest in predetermined ones To reinforce distances, owns.

Compared to methods according to the prior art, which combine several versions of a source signal shifted by given distance values, the method according to the invention is characterized in that in each step iteratively combines a previously generated intermediate signal with the one of the given distance values shifted, Intermediate signal combined, whereby the signal level of the combined signals can only be compared and whereby signal components that are not of interest can only be meaningfully suppressed.

Compared to the use of self-similarity matrices, the method according to the invention dispenses with the consideration of several, namely n, different partial results, in this case secondary diagonals, if several, for example n, distance values a1, . . . are known, since the method according to the invention generates a single result signal using all distance values , which can ultimately be used solely to determine the repeated signal components sought. This principle avoids missing detections, which can result when using self-similarity matrices because some, in the limit case even only one, of the concise secondary diagonals were not detected. In addition, the generally additionally required step of feature extraction is not required in the method according to the invention.

Compared to the constellation search, the method according to the invention offers on the one hand a possibility of determining repeated signal components directly and without a feature extraction step from the source signal. Furthermore, this is also and especially possible with a high detection quality when only a few characteristic distance values are known.

application areas

The method according to the invention offers a wide range of possible applications in the field of signal-based pattern recognition in all areas of application in which the detection of useful signals within source signals plays a role and the useful signals, also known as signal components or useful signal components, occur repeatedly in some form. Particularly noteworthy here are one-dimensional and two-dimensional useful signals and source signals. In the case of the one-dimensional source signals, there are preferably possible applications in the field of audio signal processing, in particular the processing of acoustic signals, the processing of speech signals, bioacoustic pattern recognition and music signal processing. Repeatedly present useful signals in these areas of application are, for example, but not exclusively, machine noises during regular operation and/or in the event of defects, footstep noises, knocking noises, human calls, bioacoustic noises (animal voices, calls, movement noises) or noises occurring in the musical rhythm. After the application of a spectral transformation, for example Fourier transformation, a large number of other repeated useful signals are often found, for example harmonic components of music, harmonic components of human speech, which are particularly found in tonal speech sounds, or other useful signals that have harmonic components, also known as overtones. Such components are namely characterized in that the frequencies 2*F, 3*F, 4*F, F, 2*F, 3*F,... are. In the method according to the invention, such a scenario can also be understood such that distance values a1=F, a2=F, a3=F, . . . are known. Typically, however, the set of distance values used will be small, e.g., less than 10, and finite. A temporal variability of the fundamental frequency F is also typical, such as in the case of harmonic signals from the areas of speech, music and bioacoustics. In order to use the method according to the invention, it is therefore particularly appropriate in this case to consider short time excerpts. In a special embodiment, a meaningful basic frequency F, ie here a meaningful distance value a=F, is suitably estimated before the application of the method according to the invention, if this is not known in advance. In this case, a rough value range for F will typically be known, within which the estimation can take place with the aid of the source signal.

In the case of two-dimensional signals, the method according to the invention can be used particularly preferably in the field of image processing and pattern recognition in image processing. Possible uses are given in particular where certain, exemplary known objects, object constellations, objects, etc. are searched for as sub-images of a given image, with the exemplary known objects being characterized in that certain of their sub-components have characteristic distances from one another, i.e. provide distance values, preferably at least two distance values that can be used by means of the method according to the invention. Examples of such objects and characteristic distance values are buildings or building floor plans with given distances between the parallel outer walls, aerial photos of streets with distances between lanes and/or lane boundaries, also vehicles with distances resulting from length and diameter, further images, for example copyrighted images, in which repeated sub-components, for example parallel polylines, occur that have characteristic distances from each other. In the latter application, the method according to the invention can be used, for example, to derive an efficient method for determining the use of a given image, for example as a sub-image of another image, with an application-relevant area of application being the determination of copyright infringements.

There are also possible uses for the method according to the invention in the field of radio reconnaissance, as well as in the surveillance and monitoring of radio traffic. In particular, there is application potential in the area of unauthorized, or "non-cooperative" reception. There are application scenarios where specific useful signals are to be detected within a recorded source signal, in this case a radio signal, but the exact time position and/or the exact transmission frequency of which are not known. However, in application scenarios that are particularly preferred here, information is known about repetitions occurring in these useful signals; more precisely, signal components are repeated at intervals of certain interval values and/or in frequency at intervals of certain frequency intervals. Examples are burst signals, such as ARQ signals, or multi-carrier signals. The repetitions can be exact or only approximately exact or even similar. The method according to the invention can be used particularly advantageously here if useful signals are to be detected which are very weak, for example only slightly stronger than the background noise. Furthermore, such useful signals can be emphasized particularly suitably by the method according to the invention those in which one or more of the repeated components are partially obscured by other signals in time and frequency.

The method according to the invention is typically also well suited for detecting interference signals, i.e. signals that were generated unintentionally or not cooperatively and are expressed in a frequency signal by unwanted signal components, and can also disadvantageously superimpose other useful signals. Interfering signals are often present in a frequency signal in the form of signal components repeated at certain intervals. Here, too, there are often previously known distance values, for example distances that are multiples of the wavelength or the reciprocal wavelength of the mains voltage frequency. Furthermore, there are often distance values that are characteristic of specific electrical devices or machines. The method according to the invention can thus be used particularly advantageously for detecting the resulting interference signals using such previously known distance values.

exemplary embodiments

The invention is based on one in the 4 and 5 illustrated preferred embodiment described. In 4 shows (i) a source signal with three signal components (useful signal components) at the positions marked s1, s2 and s3. In this case, s1 has the distance a from s2, and s2 also has the distance a from s3. The distance values a1=a and a2=a are thus supplied to the method according to the invention. According to a realistic application scenario, the source signal is typically selected in such a way that it is not exactly equal to zero at any point. The horizontal signal areas shown in (i) (in front of s1, between s1 and s2, between s2 and s3 and behind s3) thus consist of non-zero values. This representation takes into account the scenario that is characteristic of applications, in which a signal obtained or recorded in some way, even at points where no useful signals are present, consists of at least some type of noise background made up of signal values that are small in the order of magnitude. The absolute signal is then formed from the source signal, which in the present application example, however, corresponds to the source signal and is therefore not shown separately. In 4 (i) shows the absolute signal at the same time. The shifted absolute signal x1 shown in (ii) is then formed by shifting by the distance value a1=a. The intermediate signal y1 shown in (iii) is then formed by point-by-point product formation, which for each n becomes y1(n) = x(n)*x1(n), equivalent to y1(n)=x(n)*x( well), results. Due to the shift by the distance value a, the shifted absolute signal (ii) contains the signal components at positions s1+a, s2+a and s3+a. Since signal components occur in the absolute signal (i) and in the shifted absolute signal (ii) at the points s1+a=s2 and s2+a=s3, the intermediate signal (iii) contains particularly large signal components at these two points, denoted in (iii). with S1 and S2. Since there is also a signal component in the absolute signal (i) at point a1 and in the shifted absolute signal (ii) at point s3+a and none of the signals is equal to zero in any area, the intermediate signal (iii) is due to the product formation signal components, denoted by G1 and G2, can also be found in these positions. However, these components are not as pronounced as components S1 and S2. It can thus be seen how the product formation step emphasizes signal components which occur at the same place in both signals (i) and (ii) over those which only occur in one of the two signals.

5 shows the effect of the subsequent step of minimizing the intermediate signal with the shifted intermediate signal. For a better illustration of how the minimum formation works, see (i) of 5 an intermediate signal y (designation from 1 ) shown that with the in 4 (i) shown absolute signal. In addition to an illustration of the functional principle of minimum formation, a comparison with the functional principle of product formation is thus made possible. The signal in 5 (i) shows signal components (a useful signal repeated three times) at positions s1, s2 and s3, with a distance value of a between s1 and s2 and between s2 and s3. In (ii) again a version of the signal from (i) shifted by a distance value a is shown, the shifted intermediate signal x2 (designation from 1 ). The latter shows distinct signal components at the positions s1+a, s2+a and s3+a. In (iii) the result signal determined from the product formation can be seen, which is calculated mathematically at each point n as y(n) = Min(y1(n),x2(n)). Since the intermediate signal y1 (i) and the shifted intermediate signal x2 (ii) each have pronounced signal components at the points s2=s1+a and s3=s2+a, the result signal y (iii) has pronounced signal components at these points even after the minimum formation , labeled S1 and S2. For the points s1 and s3+a, at which only one of the two signals y1 (i) and x2 (ii) has a pronounced signal component, the formation of the minimum causes a suppression of signal components, so that at these points in the result signal y (iii) there are no pronounced signal components occur.

The step of product formation advantageously brings about an amplification of signal components repeated at a distance a in a source signal. The components G1 and G2 resulting in product formation are also referred to as ghost components, since they can represent additional signal components in the signal resulting after product formation that are not present in the source signal. In particular, two ghost components result for each signal component present in a source signal which is not repeated at the predetermined distance a, ie which does not represent a useful signal to be amplified or detected by the method according to the invention. An essential insight on which the method according to the invention is based is that the step of minimum formation has the opposite effect of product formation by suppressing those signal components which are not repeated at the distance a. It is therefore particularly advantageous to combine the two steps of product formation and minimum formation, as disclosed in the method according to the invention.

6 illustrates the facts presented using an absolute signal for a source signal that was obtained from a power density spectrum of a radio signal. Repeated signal components, denoted by A, are visible in the absolute signal, which were generated by interference signals in the present application. Neighboring signal components A have a distance from one another of a=2430 Hertz. Furthermore shows 6 three other particularly pronounced signal components, denoted by B, which, however, have nothing to do with the interference signals of interest. The aim of the method according to the invention is now to emphasize and detect the signal components A using the known distance value a. 7 Figure 12 shows the effect of applying a product formation step once to determine an intermediate signal. In addition to an amplification of the components of the interference signal A of interest, the development of ghost components B′ relative to the signal components B of no interest can be clearly seen. 8th serves to illustrate the effect of a repeated application of the product formation step, in which a product signal of the intermediate signal from 7 is formed with a version of this intermediate signal that is shifted by a. It can be seen how, in addition to a further amplification of the signal components A of interest, further ghost components B'' arise.

Compared to the method of double product formation just presented 9 the result signal resulting from the method according to the invention. It can be clearly seen here how the signal components of the interference signal of interest are emphasized in the distance of the distance value a. In the original absolute signal ( 6 ) existing signal components B that are not of interest are in the signal in 9 strongly suppressed and not or only with great difficulty recognized. Ghost components are also not visible. It is obvious how from the result signal in 9 the positions of the signal components of interest can be determined in a simple manner and according to the prior art by methods for detecting local energy accumulations or local maxima.

A preferred embodiment of the method according to the invention (the process sequence itself is shown in 3 ) is in the 6 , 9 , 10 and 11 illustrated. the 6 shows, as already explained above, an absolute signal which was formed from a source signal by spectral transformation, here Fourier transformation, with subsequent formation of the absolute signal. It should also be mentioned that alternatively and according to the prior art, further steps of signal transformation, in particular the formation of signal squares, the formation of higher signal powers, the formation of logarithms and/or signal scaling can be carried out as point-by-point operations on the absolute signal before further processing with the method according to the invention takes place according to a preferred embodiment. The resulting signal is then used instead of the absolute signal (2a, 3 ) for further processing. 9 then shows, as already shown, the result signal y (8 in 3 ). 10 shows that in step 9 ( 3 ) spectrogram S = S(1),...,S(r) (10 in 3 ). The vectors S(1), S(2),... are sequentially in the columns of the in 10 matrix S represented by intensity values. Dark values mean pronounced (large) values. It can be clearly seen again how signal components develop at distance a. The window width selected to determine the windowed Fourier transformation corresponds to 1500 Hz in the example shown, the selected step size is 500 Hz. In addition to the recognizable distinctiveness of the signal components, it can be seen that each of the components has a horizontal, relatively uniform line pattern. This pattern, which is uniform within the components, causes a particularly strong expression of the signal components A of interest during the further implementation of the method according to a further preferred embodiment 11 that in step 11 (according to 3 ) determined second result signal y' (in 11 denoted by 1), which in the present application example benach by comparison barter vectors of S, more precisely by scalar products of energy normalized neighboring vectors. The value y'(i) in the application example contains the scalar product of energy-normalized versions of the vectors S(i) and S(i+1), also referred to as the cosine measure in the technical literature, but this does not exclude alternative comparison methods. Mere examples of other possible distance values are the mutual entropy, the Euclidean distance, the distance with respect to any vector norm or metric, or the Kullback-Leibler (KL) divergence.

11 also shows how the positions of the signal components (A) of interest can be determined from the second result signal (1) by means of a method for so-called spectral segmentation according to the prior art. This method works as follows: a first auxiliary signal (2), which represents a smoothed version of the result signal, is generated from the second result signal (1) by low-pass filtering. A second auxiliary signal (3) is then generated by low-pass filtering the second result signal (1) reduced by the first auxiliary signal (2). After determining the points of intersection of the second auxiliary signal (3) with the second result signal (1), the position of the minimum value of the second result signal (1) with respect to a locally limited neighborhood is determined for each such point of intersection. In the example at 11 the positions of these minimum values are marked with triangles. Neighboring triangles are used as interval boundaries to determine signal components that may be of interest. A left interval boundary is formed when the right-hand part of the second result signal (3) (ie the signal to the right of the determined interval boundary) is rising; in the example, such a left interval boundary is denoted by (4a). A right interval boundary is also formed when the left-hand part of the second result signal (3) (ie the signal to the left of the determined interval boundary) is rising; in the example, such a right interval boundary is denoted by (4b). A significance value is then formed from the energy difference (temporary sum of the signal differences) between the second result signal (1) and the second auxiliary signal (3) between all pairs of immediately adjacent left and right interval boundaries. In the example shown, the significance value determined in this way between the left interval limit (4a) and the right interval limit (4b) has the unitless value 3.25. Finally, for each determined interval, a detection message for a potentially interesting signal component is output precisely when the associated significance value exceeds a previously specified threshold value. In 11 six such finally reported intervals are marked with a diamond. 12 finally shows an application scenario of the method according to the invention for the case of multidimensional signals according to a further preferred embodiment, here for the exemplary case of two dimensions. The example shows a framed (two-dimensional) image in which four vehicles (1), (2), (3) and (4) are shown in top view. The figure shows (not part of the image content) the dimensions (a1) to (a6) of the vehicles, vehicles (1), (2) and (4) having the same length and width dimensions (a1), (a2) and (a3 ) own. The dimensions (a4), (a5) and (a6) of the vehicle (3) are different. The method according to the invention can now be used, for example, with prior knowledge of a vehicle of type (1) and its dimensions, to highlight the vehicle positions by the known vertical distances (a2) and (a3) of the multidimensional configuration according to claim 9 are supplied, where twice the vertical (or y) dimension is chosen as dimension values. As the example shows, if edges or lines and distances between such lines are present, it is advisable to use a method for emphasizing edges, such as a filtering method according to the prior art, before using the method according to the invention.

pamphlets

• [Müller 2007] Müller, M. (2007): Information Retrieval for Music and Motion. Springer, ISBN 978-3-540-74047-6.
• [Gardner 1986] Gardner, W. (1986): The Spectral Correlation Theory of Cyclostationary Time-Series, Signal Processing 11, pp. 13-36.
• [Bardeli et al. 2008] Bardeli R., Wolff D., Clausen M. (2008): Bird Song Recognition in Complex Audio Scenes, Proceedings of the International Expert meeting on IT-based Detection of Bioacoustical Patterns, BfN-Skripten 234: 93-101.
• [Ouyang et al. 2012] Ouyang W, Tombari F; Mattoccia, S. ; Di Stefano, L. ; Wai-Kuen Cham (2012): Performance Evaluation of Full Search Equivalent Pattern Matching Algorithms, IEEE Trans. on Pattern Matching and Machine Intelligence, vol. 34, no. 1.
• [Rabiner 1989] Rabiner, L. (1986): A Tutorial on Hidden Markov Models in Selected Applications in Speech Recognition, Proc. of the IEEE, vol. 77, no. 2.
• [Clausen et al. 2003] Clausen M., Körner H., Kurth F. (2003): An Efficient Indexing and Search Technique for Multimedia Databases, SIGIR 2003 Workshop on Multimedia Retrieval.
• [Kaupinnen 2002] Kauppinen, I. (2002): Methods for detecting impulsive noise in speech and audio signals. In the 14th International Conf. on Digital Signal Processing Proc. DSP 2002 (Cat. No.02TH8628), vol. 2, pp. 967-970.
• [Bello et al 2005] Bello, J.P.; Daudet L.; Abdallah, S.; Duxbury, C.; Davies M and Sandler M (2005) A tutorial on onset detection in music signals. IEEE Transactions on Speech and Audio Processing, 13(5):1035-1047.
• [Foote 2000] J Foote (2000): Automatic audio segmentation using a measure of audio novelty. In proc. IEEE Int. conf Multimedia and Expo (ICME2000), vol. I, New York, pp. 452-455.

Claims (17)

Method for emphasizing and detecting repeatedly occurring useful signal components within a source signal in which - a one-dimensional source signal q is available as a sequence of signal values, - two distance values a1 and a2 are known as estimates of the distances of repeatedly occurring useful signal components, comprising the steps - that absolute values are obtained point by point from the source signal q, x(n) = |q(n)|, resulting in an absolute signal x, - that a version x1 of x shifted by a1 signal values is formed, x1(n) = x(n-a1), - that an intermediate signal y1 is formed by product formation point by point, y1(n) = x1(n)*x(n), - that a version x2 of the intermediate signal y1 shifted by a2 signal values is formed, x2(n) = y1(n-a2), - that by forming a minimum of x2 and y1 point by point, a result signal y is formed, y(n) = Min(x2(n), y1(n)), - that candidate positions are determined from the result signal y by finding local maxima or local energy accumulations, - that repetition positions in the source signal x are determined from each of the candidate positions, taking into account the distance values a1 and a2 - That the repetition positions determined are output for further processing. procedure after claim 1 , in which, in the case of different distance values a1 and a2 for two specific candidate positions k1 and k2, for which k1=k2+a1 applies exactly or approximately, k2, k2-a2 and k2-a2-a1 are output as repetition positions. procedure after claim 1 , in which the pointwise product formation step is carried out before the pointwise minimum formation step. procedure after claim 3 , in which, in the case of different distance values a1 and a2 for two specific candidate positions k1 and k2, for which k1=k2+a2 applies exactly or approximately, k1, k1-a2 and k1-a2-a1 are output as repetition positions. Procedure according to one of Claims 1 - 4 , in which, in the case of the same distance values a1=a2, a repetition position w=k-a1 is determined for each specific candidate position with the value k. Procedure according to one of Claims 1 - 5 , in which, in the case of different distance values a1 and a2 for a candidate position k, k, k-a1, k-a2 and k-a2-a1 are determined as repetition positions. Procedure according to one of Claims 1 or 3 , for which any finite number K of distance values a1,...,aK are known, where K is greater than two, the absolute signal is denoted by y0 (y0=x) and instead of forming an intermediate signal, now a sequence of K +1 intermediate signals y1,...,y(K+1), where each intermediate signal y(i+1) is formed from the previous intermediate signal yi and a shifted version yi(n-di) thereof, either by minimizing, y( i+1) (n) = Min(yi(n), yi(n-ai)), or by product formation, y(i+1) (n) = yi(n)*yi(n-ai). and the result signal y is finally formed from yK, y(n)=yK(n), but at least one of the intermediate signals is formed by product formation and at least one of the intermediate signals is formed by minimum formation. procedure after claim 7 , in which all distance values a1,...,aK agree and, if K=2*M is an even number, the number k-a1*M is determined as repetition position for each candidate position k, and, if K=2*M +1 is an odd number, for each candidate position k, the number k-a1*M or the number k-a1*(M+1) is determined as the repetition position. Method according to one of the preceding claims, in which the source signal is an N-dimensional signal, q=q(n1,...,nD), thus also the absolute signal x=x(x1,...,xN), and additionally for the list of distance values a1,..., ak there is a specification list d1,..., dk of dimensions to be selected and each individual step of minimum or product formation now relates to one individual of the D dimensions as specified in the specification list, in the case of di=j and the formation of the minimum, y(i+1) (n1,...,nD) = Min(yi(n1,..., nj,..., nD) , yi(n1,. .., nj-ai,..., nD)) or in the case of product formation y(i+1) (n1,...,nD) = yi(n1, ..., nj , ..., nD) *yi(n1,...,nj-ai,...,nD). Procedure according to one of Claims 1 - 8th , in which the discrete Fourier transform Dq is first determined from the source signal q, the absolute signal x is determined from the absolute value of the Fourier transform, x = |Dq|, a spectrogram STFT[ from the result signal y before the step of finding local maxima using windowed Fourier transformation y] is obtained and a second result signal y' is obtained from the spectrogram by normalizing the columns point by point or by comparing adjacent columns in pairs, which is fed to further processing instead of y. Method according to one of the preceding claims, in which the candidate positions are determined from the result signal using methods for spectral segmentation. Method according to one of the preceding claims, in which a summation is carried out at one or more points instead of the product formation. Method according to one of the preceding claims, in which a maximum is formed at one or more points instead of the product being formed. Method according to one of the preceding claims, in which, in the step of determining an intermediate signal, instead of determining an individual intermediate signal with regard to a predetermined distance value a, a set of intermediate signals is now formed for a set of distance values similar to a, from which only that intermediate signal of the further Processing is supplied, which has particularly pronounced signal components. Method according to one of the preceding claims, in which the source signal is an N-dimensional signal, q=q(n1,...,nD), thus also the absolute signal x=x x1,...,xN), and a list of N-dimensional distance vectors a1,...,ak and the j-th shifted intermediate signal xj by shifting the j-th intermediate signal by the j-th distance vector aj=(aj1,...,ajD) as xj (n1, ..., nD) = yj (n1-aji,...,nD-ajD) is formed. Device for emphasizing and detecting repeatedly occurring useful signal components within a source signal full Means for reading in a in the form of a sequence of signal values present one-dimensional source signal q, Means for estimating two distance values a1 and a2 of the distances between repeatedly occurring useful signal components, means for generating pointwise absolute values from the source signal q, x(n) = |q(n)|, resulting in the generation of an absolute signal x, means for forming a version x1 of x shifted by a1 signal values, x1(n) = x(n-a1), Means for point-by-point product formation of an intermediate signal y1, y1(n) = x1(n)*x(n), Means for generating a version shifted by a2 signal values x2 of the intermediate signal y1, x2(n) = y1(n-a2), Means for point-by-point minimization of x2 and y1 Means for generating a result signal y, y(n) = Min(x2(n), y1(n)), Means for finding local maxima or local energy accumulations for determining candidate positions from the result signal y, Means for determining repetition positions in the source signal x from each of the candidate positions, taking into account the distance values a1 and a2 Means for outputting the determined repetition positions. Computer program or computer program product, with a program code for carrying out all steps of the method claim 1 , if the computer program runs on a computer or if the computer program product is operatively connected to a computer.

Images