JP4360340B2 - Digital watermark detection apparatus, method, and program - Google Patents

Description

  The present invention relates to a digital watermark detection apparatus, method, and program, and in particular, a modification expressed by primary transformation such as enlargement / reduction, rotation, change in aspect ratio, affine transformation, etc., for a signal with a digital watermark embedded therein. The present invention relates to a digital watermark detection apparatus, method, and program for facilitating detection of a digital watermark even when added.

  Conventionally, by embedding digital watermarks into digital content, copyright information protection of digital content is performed, copyright information related to digital content is referred to, or analog media such as printed matter as digital content is used as an advertisement. On the other hand, services such as photographing with a digital camera or the like and reading information related to advertisement by reading a digital watermark have been realized.

  For example, when embedding a digital watermark, based on the embedding target component position information, spectrum spreading is performed to change the real and imaginary components of the complex matrix independently, and a watermark pattern is generated independently of the input image, and the actual image When the embedded image is generated by adding the pattern and the digital watermark is detected, the detection target sequence is generated based on the detection target component position information, the offset information is extracted, and the detection target sequence is corrected. There is a digital watermark method that performs spectral despreading later and detects a digital watermark embedded in the cut out pixel block (see, for example, Patent Document 1).

  Further, in the digital watermark, it is necessary that the digital signal representing the digital content is resistant to various modifications received after the digital watermark is embedded, that is, the digital watermark can be detected after being subjected to various modifications.

  For example, there is a method of estimating the degree of alteration applied to a digital signal by embedding a signal for geometric correction separately from the digital watermark information and detecting this (for example, , See Patent Document 2). However, in such a method, a signal for geometric correction is embedded in the digital signal in addition to the information of the digital watermark, which leads to deterioration of the quality of the digital signal embedded with the digital watermark and an increase in processing time of the digital watermark. There was a possibility.

  On the other hand, as a method of detecting a digital watermark from a digital signal in which a digital watermark is embedded (hereinafter referred to as a digital watermark embedded signal), a pattern having a common property is embedded as a digital watermark at a plurality of positions in the signal in advance. In addition, by obtaining the autocorrelation function of the signal embedded with the digital watermark at the time of detecting the digital watermark, it is expressed by the primary conversion such as enlargement / reduction, rotation, aspect ratio change, affine transformation added to the signal embedded with the digital watermark. There has been proposed a method of detecting a digital watermark that estimates the degree of modification to be performed, corrects the digital watermark-embedded signal that has been modified according to the estimated amount of modification, and facilitates detection of the digital watermark. It was.

  For example, a marker image having a pair of identical features is embedded in the original image to generate an image, and from the position of the peak point appearing in the autocorrelation function of the image rotated or enlarged / reduced to the marked image, There is a method for calculating the degree of rotation or enlargement / reduction applied to a signal embedded with a digital watermark (see, for example, Patent Document 3). Also, in this method, in order to find the peak point, the autocorrelation function value is filtered using a high-frequency pass filter (HPF) (hereinafter referred to as conventional method 3).

In addition, for example, a periodic pattern having repetition is used as the embedding pattern itself having information to be embedded as a digital watermark, and the enlargement / reduction ratio is calculated by observing the change of the period from the autocorrelation function at the time of detection. There is a method of performing digital watermark detection processing based on the enlarged / reduced rate (for example, see Patent Document 4).
Japanese Patent Application Laid-Open No. 2003-219148 “Digital watermark embedding method, digital watermark detection method, digital watermark embedding device, digital watermark detection device, storage medium storing a digital watermark embedding program, and storage medium storing a digital watermark detection program” Japanese Patent Application Laid-Open No. 11-355547 “Geometric Transformation Specific System” Japanese Patent Laid-Open No. 2001-209806 “Method and Computer Program for Detecting Image Rotation or Enlargement / Reduction” Japanese Patent Application Laid-Open No. 2002-111994 “Digital watermarking method and system having scaling resistance”

  In calculating the autocorrelation function as in the conventional method 3, when the size of the input signal, for example, the number of vertical and horizontal pixels of the image in the case of a two-dimensional image signal is large, the autocorrelation function for the entire signal is calculated. Since the calculation is a time-consuming process, it is necessary to calculate an autocorrelation function using a discrete Fourier transform or the like for a small area that can be processed.

  However, when the enlargement rate as a modification applied to the signal is large, the size of the pattern embedded in the signal becomes large compared to the region to be subjected to autocorrelation function calculation, and as a result, the autocorrelation function value There is a problem that the peak point does not appear in the area and the degree of modification cannot be estimated.

  The present invention has been made in view of the above points, and eliminates the need to calculate a wide area autocorrelation function both when the enlargement ratio as a modification given to the signal is small and when it is large, and calculates the autocorrelation function. To provide a digital watermark detection apparatus, method, and program that can easily detect a digital watermark by searching for a peak point appearing in an autocorrelation function value with high accuracy while suppressing a processing time required for Objective.

  FIG. 1 is a principle configuration diagram of the present invention.

The present invention (claim 1), the digital signal is a two-dimensional image signals, the predetermined pattern is embedded in advance as a predetermined electronic watermark so as not to be perceived pattern repeated human perception and expanded as modified or reduction, rotation, when the digital signal subjected to any of linear transformation changes the aspect ratio is input, an electronic watermark detection apparatus for estimating a transformation parameter representing the primary conversion applied to the digital signal,
A signal reduction means 10 for reducing the image size of a digital signal, which is a two-dimensional image signal to which an input modification has been applied, at a predetermined magnification;
Autocorrelation function calculation means for calculating an autocorrelation value of a digital signal which is a two-dimensional image signal reduced by the signal reduction means 10 and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
Search for an autocorrelation peak where the autocorrelation value is a peak in the autocorrelation function value array obtained by the autocorrelation function calculation means, and in the vicinity of the position where the autocorrelation peak value is stored in the autocorrelation function value array In the vicinity of the position where each value of the autocorrelation function value array within the predetermined range and the theoretical autocorrelation peak value appearing in the autocorrelation function value array due to the periodicity of the pattern repetition are stored. Autocorrelation peak point evaluation means for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within the range of
In accordance with the evaluation value obtained by the autocorrelation peak point evaluation unit, a detection target determination unit that determines an autocorrelation peak point candidate to be used as a detection target of embedded information;
The position where the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined by the detection target determining means is stored in the autocorrelation function value array, and in the autocorrelation function value array when no alteration is made Conversion parameter calculation means 20 for calculating a conversion parameter for performing a primary conversion applied to the digital signal based on a deviation from the position stored in the autocorrelation function value array .

In the present invention (Claim 2), a predetermined pattern is embedded in advance as a digital watermark so as not to be perceived by human perception repeatedly on a digital signal which is a two-dimensional image signal, and is expanded as a modification. Alternatively, when a digital signal subjected to primary conversion of any one of reduction, rotation, and aspect ratio change is input, a digital watermark detection apparatus that estimates a conversion parameter representing the primary conversion applied to the digital signal,
A signal reduction means for reducing the image size of the digital signal, which is a two-dimensional image signal to which the input modification has been applied, at a predetermined magnification;
Autocorrelation function calculating means for calculating an autocorrelation value of a digital signal which is a two-dimensional image signal reduced by the signal reduction means, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
Search for an autocorrelation peak where the autocorrelation value is a peak in the autocorrelation function value array obtained by the autocorrelation function calculation means, and in the vicinity of the position where the autocorrelation peak value is stored in the autocorrelation function value array In the vicinity of the position where each value of the autocorrelation function value array within the predetermined range and the theoretical autocorrelation peak value appearing in the autocorrelation function value array due to the periodicity of the pattern repetition are stored. Autocorrelation peak point evaluation means for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within the range of
In accordance with the evaluation value obtained by the autocorrelation peak point evaluation unit, a detection target determination unit that determines an autocorrelation peak point candidate to be used as a detection target of embedded information;
The position where the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined by the detection target determining means is stored in the autocorrelation function value array, and in the autocorrelation function value array when no alteration is made A conversion parameter calculating means for calculating a conversion parameter for performing a primary conversion applied to the digital signal from a deviation from a position where the autocorrelation peak appearing in is stored in the autocorrelation function value array;
Geometrical correction is performed by applying an inverse transformation to the conversion parameter to the modified digital signal, and the geometrically corrected digital signal is translated by a predetermined amount of movement within a predetermined region. And a parallel movement amount estimating means that attempts to detect a digital watermark from the digital signal and sets the movement amount when the digital watermark is detected as a parallel movement amount added to the digital signal.

The present invention (Claim 3) is the digital watermark detection apparatus according to claim 1 or 2 ,
Average and variance of autocorrelation values obtained by autocorrelation function calculation means for each digital signal that is a two-dimensional image signal that has been reduced and modified at each magnification when there are a plurality of predetermined magnifications. Or it further has an autocorrelation function value normalizing means for normalizing the autocorrelation value so that the standard deviation becomes the same .

The present invention (Claim 4) is the digital watermark detection apparatus according to any one of Claims 1 to 3 ,
When there are a plurality of different predetermined magnifications , an average of all evaluation values obtained by the autocorrelation peak point evaluation means for each digital signal that is a two-dimensional image signal that has been modified and reduced at each magnification. And an autocorrelation peak point evaluation value normalizing means for normalizing the evaluation value so that the standard deviation becomes the same .

Further, according to the present invention (Claim 5) , the detection target determining means of the digital watermark detection apparatus according to any one of Claims 1 to 4,
When the predetermined magnification is plural,
Priority peak point candidate number storage means for storing the number L of peak point candidates to be prioritized;
Embedded information detection upper limit number storage means for storing the number M of peak points to be output as detection targets;
For each digital signal that is a modified two-dimensional image signal reduced at each magnification, select L evaluation values in descending order of evaluation values obtained by the autocorrelation peak point evaluation means, and select It is larger than any of the evaluation values obtained by the autocorrelation peak point evaluation means for each digital signal that is a two-dimensional image signal that has been modified with a reduction scaled at each magnification. Autocorrelation peak point evaluation means for adding a predetermined value p ₀ ;
Evaluation values added and not added by the autocorrelation peak point evaluation value increasing means obtained for each digital signal which is a two-dimensional image signal that has been reduced and modified at each magnification. The M evaluation values for the embedded information detection upper limit count stored in the embedded information detection upper limit count storage means are selected in order from the largest evaluation value, and the peak point candidates corresponding to the selected M evaluation values are detected. Embedded information detection target peak point candidate selection means to output as,
Thus, embedded information detection target peak point candidates are determined.

  FIG. 2 is a diagram for explaining the principle of the present invention.

In the present invention (Claim 6), a predetermined pattern is embedded in advance as a digital watermark so as not to be perceived by human perception repeatedly on a digital signal which is a two-dimensional image signal, and expanded as a modification. or reduction, rotation, when the digital signal subjected to any of linear transformation changes the aspect ratio is input, the digital watermark detection apparatus for estimating a transformation parameter representing the primary conversion applied to the digital signal, a digital watermark A detection method,
A signal reduction step (steps 1 and 2) for reducing the image size of the digital signal , which is a two-dimensional image signal to which the input modification has been applied, at a predetermined magnification ;
An autocorrelation function calculating step of calculating an autocorrelation value of a digital signal, which is a two-dimensional image signal reduced in the signal reduction step, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
In the autocorrelation function value array obtained in the autocorrelation function calculation step, search for an autocorrelation peak in which the autocorrelation value peaks, and in the vicinity of the position where the autocorrelation peak value is stored in the autocorrelation function value array In the vicinity of the position where each value of the autocorrelation function value array within the predetermined range and the theoretical autocorrelation peak value appearing in the autocorrelation function value array due to the periodicity of the pattern repetition are stored. An autocorrelation peak point evaluation step for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within the range of
In accordance with the evaluation value obtained in the autocorrelation peak point evaluation step, a detection target determination step for determining autocorrelation peak point candidates to be used for detection of embedded information;
The position in which the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined in the detection target determination step is stored in the autocorrelation function value array, and in the autocorrelation function value array if no modification has been made Conversion parameter calculation step (steps 3 and 4) for calculating a conversion parameter for performing a primary conversion applied to the digital signal from a deviation from the position where the autocorrelation peak appearing in is stored in the autocorrelation function value array And do.

According to the present invention (Claim 7) , a predetermined pattern is embedded in advance as a digital watermark so as not to be perceived by human perception repeatedly on a digital signal which is a two-dimensional image signal, and is expanded as a modification. Alternatively, when a digital signal subjected to primary conversion of any one of reduction, rotation, and aspect ratio change is input, a digital watermark in a digital watermark detection apparatus that estimates a conversion parameter representing the primary conversion added to the digital signal A detection method,
A signal reduction step of reducing the image size of the digital signal, which is a two-dimensional image signal to which the input modification has been applied, by a predetermined magnification;
An autocorrelation function calculating step of calculating an autocorrelation value of a digital signal, which is a two-dimensional image signal reduced in the signal reduction step, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
In the autocorrelation function value array obtained in the autocorrelation function calculation step, search for an autocorrelation peak in which the autocorrelation value peaks, and in the vicinity of the position where the autocorrelation peak value is stored in the autocorrelation function value array In the vicinity of the position where each value of the autocorrelation function value array within the predetermined range and the theoretical autocorrelation peak value appearing in the autocorrelation function value array due to the periodicity of the pattern repetition are stored. An autocorrelation peak point evaluation step for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within the range of
In accordance with the evaluation value obtained in the autocorrelation peak point evaluation step, a detection target determination step for determining autocorrelation peak point candidates to be used for detection of embedded information;
The position in which the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined in the detection target determination step is stored in the autocorrelation function value array, and in the autocorrelation function value array if no modification has been made A conversion parameter calculating step for calculating a conversion parameter for performing a primary conversion applied to the digital signal from a shift from a position where the autocorrelation peak appearing in the position stored in the autocorrelation function value array;
Geometrical correction is performed by applying an inverse transformation to the conversion parameter to the modified digital signal, and the geometrically corrected digital signal is translated by a predetermined amount of movement within a predetermined region. An attempt is made to detect a digital watermark from a digital signal, and a parallel movement amount estimation step is performed in which a movement amount when the digital watermark is detected is added to the digital signal to obtain a parallel movement amount.

The present invention (Claim 8) is the electronic watermark detection method according to claim 6 or 7 ,
If the predetermined magnification is plural, each digital signal is a two-dimensional image signal applied the modified reduced at each magnification, the average and variance of the values of the autocorrelation obtained by the autocorrelation function calculating step Alternatively , an autocorrelation function value normalizing step for normalizing the autocorrelation value so that the standard deviation is the same is further performed.

Further, according to the present invention (Claim 9), in the digital watermark detection method according to any one of Claims 6 to 8, when there are a plurality of predetermined magnifications, a modification reduced at each magnification is added. Autocorrelation peak point evaluation value that normalizes the evaluation value so that the average and standard deviation of all evaluation values obtained by the autocorrelation peak point evaluation means are the same for each digital signal that is a two-dimensional image signal A normalization step is further performed.

Further, according to the present invention (Claim 10) , in the detection target determining step according to any one of Claims 7 to 10, when the predetermined magnification is plural,
For each digital signal that is a modified two-dimensional image signal reduced at each magnification, L evaluation points are selected in descending order of evaluation values obtained in the autocorrelation peak point evaluation step. A predetermined value that is larger than any of the evaluation values obtained in the autocorrelation peak point evaluation step for each digital signal that is a two-dimensional image signal that has been modified by reducing each magnification to the L evaluation values. Autocorrelation peak point evaluation value increasing step of adding the value p ₀ of
Evaluation values added and not added by the autocorrelation peak evaluation value increasing step obtained for each digital signal that is a two-dimensional image signal that has been reduced and modified at each magnification. In order from the largest evaluation value, M evaluation values of the embedded information detection upper limit count stored in the embedded information detection upper limit count storage means are selected, and peak point candidates corresponding to the selected M evaluation values are detected. An embedding information detection target peak point candidate selection step to output as a target is performed.

According to the present invention (claim 11), a predetermined pattern is embedded in advance as a digital watermark so as not to be perceived by human perception repeatedly on a digital signal which is a two-dimensional image signal, and expanded as a modification. or reduction, rotation, when the digital signal subjected to any of linear transformation changes the aspect ratio is inputted, a digital watermark detection program for estimating a transformation parameter representing the primary conversion applied to the digital signal,
A digital watermark detection program for causing a computer to function as each means of the digital watermark detection apparatus according to any one of claims 1 to 5.

According to the digital watermark detection apparatus of claims 1 and 2 and the digital watermark detection method of claims 6 and 7 , the auto-correlation function is calculated after reducing the input digital signal with modification. Therefore, a necessary autocorrelation peak can be obtained even by calculating the autocorrelation function in a small region, and the processing time required for the autocorrelation function calculation can be suppressed, thereby enabling efficient digital watermark detection.

Further, according to claim 1, 2 of the digital watermark detection apparatus and claim 6, 7 digital watermark detection method of the present invention, for performing the peak point evaluation for each signal that is reduced to a plurality of different magnifications, the signal Even when the degree of modification applied is assumed to be in a wide range from a small enlargement ratio to a large enlargement ratio, it is possible to search for autocorrelation peak points with high accuracy and to detect digital watermarks with high accuracy. Become.

  According to the digital watermark detection apparatus of claims 3 and 4 of the present invention and the digital watermark detection method of claims 8 and 9, by normalizing the autocorrelation function value and peak point evaluation value for each of a plurality of ranges, In comparison of peak point evaluation values across the range, it appeared in different ranges depending on the noise added to the signal and the autocorrelation of the original signal, from the evaluation value of the peak point candidate of the range where the correct peak point was obtained Detection failure due to high evaluation value of an erroneous peak point candidate can be suppressed, autocorrelation peak points can be searched with high accuracy, and highly accurate digital watermark detection can be performed.

  According to the digital watermark detection apparatus of claim 5 and the digital watermark detection method of claim 10, the peak point evaluation value of each range has an evaluation value larger than the evaluation value of the correct peak point candidate of the correct range. Even when there are many wrong peak point candidates in the wrong range, priority is given to the peak point candidates at the top of each range, so that the correct peak point candidates in the correct range can be targeted for embedded information detection. In addition, it is possible to suppress the failure of detection, and it is possible to limit the upper limit number of peak point candidates to be detected as a whole to a predetermined number, thereby limiting the upper limit of the entire detection processing time. . In addition, an embedded information detection method for determining which peak point candidate is to be preferentially detected according to conditions for alterations and noise added to signals that can be estimated in individual digital watermark detection, It is possible to easily change without changing the configuration, and it is possible to efficiently detect the digital watermark.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  In the digital watermark detection apparatus of the present invention, by calculating an autocorrelation function for each signal reduced at a plurality of reduction ratios with respect to the input signal to be detected as a digital watermark, a large expansion can be performed as a modification to the signal. Even in the case where it is made, it is essential to be able to find an autocorrelation peak by calculating an autocorrelation function value in a small region.

  Hereinafter, in the first to third embodiments, the digital watermark embedding / detection method as the background of the present invention will be described in detail, and then the digital watermark detection apparatus of the present invention will be described in the fourth embodiment. Is described.

  First, terms used in the present invention will be defined.

  In the present invention, information to be embedded and detected as an electronic watermark is referred to as a “signal”. In the following embodiments, processing for mainly two-dimensional image signals will be described as an example, but video and audio signals may be used. It goes without saying that the same method can be applied to any n-dimensional signal.

  Next, the “digital watermark embedding detection communication model” used in the present invention will be described.

  FIG. 3 is a diagram showing a digital watermark embedding / detection model as a background of the present invention.

  Embedded information 102 is embedded in the pre-embedding signal 101 given to the system using the digital watermark embedding apparatus 200. The embedded information 102 may be any information as long as it is detected and used later. For example, the identifier of the content, the copyright information, the URL, the identifier of the user who purchased the content, the name, the address, the content It may be an identifier of a transaction that sold the product, an identifier of a device such as a digital camera that has generated the content, information on the device, a flag for controlling whether or not the content can be copied, a signal for detecting alteration of the content, and the like. Further, in a system that aims at obtaining the size of the signal modification 107 described later by digital watermark detection, the embedded information 102 may be a predetermined fixed value. When the embedded information 102 is a fixed value, a value stored in advance in the digital watermark embedding apparatus 200 may be used without being explicitly input to the digital watermark embedding apparatus 200. Needless to say.

  The embedded signal 104 output from the digital watermark embedding apparatus 200 is added with noise such as additive noise 106 and signal modification 107 through the communication path 105.

  Here, the additive noise 106 is, for example, a noise component that is added or subtracted to a value obtained by an element divided in terms of time, space, and frequency, such as a pixel value or a DCT coefficient in an image signal. For example, it may be noise added to a signal or encoding noise of a signal such as MPEG or JPEG.

  The signal modification 107 is modification that changes the temporal, spatial, and frequency positions of the signal, such as the shape of the subject in the image signal, and includes, for example, signal translation, enlargement / reduction, Modifications such as rotation, change in aspect ratio, primary transformation such as affine transformation, and partial cropping may be used.

  Additive noise 106, signal modification 107, and may or may not be intended by a person. The order in which the additive noise 106 and the signal modification 107 are added does not matter.

  Further, changes to signals other than those described above may be added on the communication path 105.

  The embedded signal 108 modified through the communication path 105 is input to the digital watermark detection apparatus 400, and the same detection information 110 as the embedded information 102 is output.

[First Embodiment]
The following is an example of a method for embedding a digital watermark detected by the present invention.

  FIG. 4 is a configuration example of the digital watermark embedding apparatus according to the first embodiment.

  The digital watermark embedding device 200 includes a watermark pattern generation unit 201 and a watermark pattern addition unit 202.

  The digital watermark is embedded according to the following procedure.

  1) The watermark pattern generation unit 201 generates a watermark pattern 203 to be embedded as a digital watermark into a signal based on the embedded information 102 input to the digital watermark embedding device 200.

  As a method of generating the watermark pattern 203, for example, a discrete inverse Fourier is used for a method in which the watermark coefficient matrix is changed based on the embedded information 102 as described in the first embodiment of Patent Document 1 described above. It may be a signal pattern obtained by performing conversion. An example of a watermark pattern generated when a two-dimensional image signal is targeted is shown as a watermark pattern 301 in FIG. The method of generating the watermark pattern 203 is not limited to this, and is configured so that the presence of the watermark pattern 203 is not easily perceived by human perception in an embedded signal 104 obtained as a result of adding a watermark pattern described later. Any pattern may be used as long as it is a pattern, for example, a pattern configured based on another existing digital watermark embedding method.

  2) The watermark pattern addition unit 202 repeatedly adds the watermark pattern 203 generated by the watermark pattern generation unit 201 to the pre-embedding signal 101 input to the digital watermark embedding device 200, and adds the embedded signal 104. obtain. For example, when a two-dimensional image signal is targeted, the watermark pattern 301 is repeatedly added to the pre-embedding signal 101 that is the input image signal in a tiled manner as in the example shown in FIG.

  As a method of adding the watermark pattern 203, for example, as described in the first embodiment of Patent Document 1, the pattern is emphasized and added as necessary according to a given intensity parameter. May be. Further, according to the characteristics of the signal of the part to which the pre-embedding signal 101 is added, for example, the pattern strength is reduced at a part where the added pattern is conspicuous, and the pattern strength is increased at a part where the pattern is not conspicuous. The watermark pattern may be added by changing the intensity of the pattern.

  Next, digital watermark detection will be described.

  FIG. 6 shows a configuration example of the digital watermark detection apparatus in the first embodiment.

  The digital watermark detection apparatus 400 performs a primary conversion such as translation, enlargement / reduction, rotation, aspect ratio change, affine transformation, partial clipping, etc., on an electronic watermark embedded signal obtained by using the above-described digital watermark embedding apparatus 200. This is an apparatus for detecting an embedded digital watermark from a signal subjected to processing such as lossy compression.

  The digital watermark detection apparatus 400 shown in the figure includes a primary conversion parameter estimation unit 401, a geometric transformation correction unit 402, a parallel movement amount estimation unit 403, and an embedded information detection unit 404.

  The digital watermark detection process is largely performed through the following steps.

  FIG. 7 is a flowchart of digital watermark detection processing according to the first embodiment.

Step 101) Primary transformation parameter estimation:
The primary conversion parameter estimation unit 401 calculates an autocorrelation value 405 of the input signal, searches for a point where the autocorrelation value takes a peak, estimates a primary conversion parameter given to the signal, and estimates an estimated primary A conversion parameter 406 is obtained.

Step 102) Geometric correction by linear transformation:
In the geometric transformation correction unit 402, the input signal 405 is subjected to inverse transformation with respect to the primary transformation parameter 406 estimated in step 101 to geometrically correct the signal.

Step 103) Parallel displacement estimation:
The parallel movement amount estimation unit 403 estimates the parallel movement amount given to the signal by performing a full search of the parallel movement amount with respect to the geometric transformation corrected signal 407 corrected in step 102, and the estimated parallel movement amount. 408 is obtained.

Step 104) Detection of embedded information:
The embedded information detection unit 404 detects the remaining embedded information from the geometric transformation corrected signal 407 corrected in step 102 based on the estimated parallel movement amount 408 estimated in step 103, and obtains watermark detection information 409.

  The detection of embedding information, for example, generates a geometric transformation detection target coefficient matrix as described in Patent Document 1 described above, and is used in digital watermark embedding corresponding to the detection target coefficient matrix and symbols of the embedding information. It may be made by obtaining a correlation value with the symbol series.

  Next, the process of the primary conversion parameter estimation unit 401 will be described.

  FIG. 8 is a configuration example of the primary conversion parameter estimation unit in the first embodiment.

  The primary transformation parameter estimation unit 401 shown in the figure includes an autocorrelation function value array generation unit 501, a fixed filter search unit 502, a 90-degree pair search unit 503, a detailed signal filter search unit 504, and a conversion parameter estimation unit 505. The

  An example of the procedure for estimating the primary conversion parameter by the primary conversion parameter estimation unit 401 is shown below.

  FIG. 9 is a flowchart of the primary conversion parameter estimation procedure according to the first embodiment.

  Step 201) First, an autocorrelation function value array generation unit 501 of the input signal 506 is calculated with respect to the input signal 506 input by using the autocorrelation function value array generation unit 501.

  Here, the autocorrelation function value array is obtained by calculating an autocorrelation function of an input signal given as n-dimensional discrete data and storing the obtained n-dimensional data in the array. An example of the autocorrelation function value array 508 is shown in FIG. An autocorrelation function value array obtained for a two-dimensional image signal (FIG. 10 (a)) that has been rotated, enlarged, or reduced as a modification, and the value of the autocorrelation function value expressed as the luminance of the image Is FIG. 10 (b). Here, it can be seen from the periodicity of the pattern previously embedded in the input signal that lattice-like peak points appear in the autocorrelation function value array.

  The calculation method of the autocorrelation function value array 508 is described in, for example, “Keiji Watanabe“ Image Engineering Handbook ”Asakura Shoten (1986) p. 240”. Known procedures such as despread Fourier transform of a power spectrum obtained by discrete Fourier transform of an input signal have been known.

  In this embodiment, the purpose is to detect an autocorrelation peak of a pattern composed of embedded information embedded in a signal. When the array is calculated, the autocorrelation peak for the embedded pattern may not appear sufficiently clearly. Therefore, in order to efficiently detect the target autocorrelation peak, a filter corresponding to the frequency characteristic of the embedding pattern may be applied to the input signal 506 before calculating the autocorrelation function value array. For example, when the embedded information is embedded in the high frequency region of the signal, a high frequency pass filter may be applied and further clipping processing may be performed. For example, a Laplacian filter shown in FIG. 11 may be used as the high-frequency pass filter.

  Step 202) The top N autocorrelation peak points are searched for the autocorrelation function value array 508 obtained in Step 201 using the fixed filter search unit 502.

  The fixed filter search unit 502 uses the autocorrelation function value array 508 obtained in step 201 for the impulse response space of a filter having a frequency distribution similar to the frequency distribution of the pattern formed by the embedded information embedded in the signal. The autocorrelation function value array 508 is filtered using a fixed filter that imitates the shape in the region, and the top N pieces are selected in order from the point where the magnitude of the filtered signal is the strongest in the above array And output as a peak point candidate 509.

  When the pattern constituted by the embedded information has a shape based on the sinc function as shown in an example of the description of the detailed filter search unit 504 described later, the shape based on the sinc function is imitated as a filter shape, You may use the filter of the shape shown in FIG.

  Here, imitation means that the fixed filter is configured so that the rough shape is similar to the ideal filter shape. That is, the ideal filter shape varies depending on the position on the autocorrelation function value array, as performed by the detailed filter search unit 504 described later, but in order to substitute them with a common fixed shape filter, For example, this indicates that a fixed filter having a large center value and a small value around one pixel apart from the ideal sinc function shape shown in FIG.

  How far the peak point appearing in the autocorrelation function value array is from the origin depends on the amount of enlargement / reduction in the modification 107 added to the signal and the size of the embedded pattern (for example, the watermark pattern 301 in FIG. 5). It depends on Sato. For example, in the case of a pattern repeatedly embedded in a two-dimensional image signal, the length of one side of the pattern is L, and the modification 107 added to the signal ranges from 0.5 times reduction to 1.5 times enlargement. If included, it is only necessary to search for a peak point in a range from a distance of 0.5 L to 1.5 L from the origin in the autocorrelation function value array.

  Further, since the autocorrelation function values at the origin symmetry position in the autocorrelation function value array are common from the definition of the autocorrelation function value, it is sufficient to search only two quadrants in the autocorrelation function value array.

  For example, the peak point search range in the fixed filter search unit 502 may be performed only within the frame surrounded by the thick line in FIG.

  The filtering process using the fixed filter can be performed at a higher speed than the process using the detailed filter described later, and it is possible to avoid performing unnecessary arithmetic processing in subsequent processes with a high processing load. It is possible to estimate a primary conversion parameter.

  Step 203) Using the 90-degree pair search unit 503 for the N peak point candidates 509 obtained in Step 202, vectors for each point from the origin are orthogonal to each other, and the lengths of the vectors match. Such a pair of points is searched and selected as a basis vector candidate 510 represented by the basis vector of the primary transformation added to the signal. However, according to the amount of modification given to the assumed signal, the above-described orthogonal relationship and vector length matching may be determined with a margin.

  The concept of the basis vector candidate search is shown in FIG. 14A and 14B, it is assumed that the entire rectangular area represents the autocorrelation function value array 508 and the origin is at the intersection of the arrows. Each point in FIG. 14A is a peak point candidate obtained in step 202. The result of selecting the basis vector candidate 510 from these peak point candidates is shown in FIG. 14B, and here, two sets of basis vector candidates are obtained.

  A specific implementation example of the 90-degree pair search unit 503 will be described later.

  Here, an example is shown in which the pattern constituted by the embedding information is embedded so as to form a square lattice pattern in the signal, but the embedding is performed in a different shape (for example, a rectangle or a parallelogram). In this case, in the above procedure, a point having a predetermined angle determined by the embedding shape instead of the orthogonal relationship and a ratio of a predetermined vector length determined by the embedding shape instead of matching the vector length is selected. It may be.

  Step 204) Using the detailed filter search unit 504, the candidates are narrowed down in more detail for the basis vector candidates 510 obtained in Step 203. Specifically, a filter having a shape in the spatial region of an impulse response having a frequency distribution similar to the frequency distribution of the pattern embedded in the signal in advance for the autocorrelation function value around each point in the basis vector candidate Is used to calculate an evaluation value corresponding to each point. For each base vector candidate, the sum of the above two evaluation values included in the base vector candidate is obtained, and the upper M sets of base vector candidates having a large sum are selected and set as an estimated base vector 511.

  A specific implementation example of the detailed filter search 504 will be described later.

Step 205) Using the conversion parameter estimation unit 505, the modified primary conversion parameter added to the signal is calculated based on the estimated basis vector 511 obtained in Step 204. Specifically, the base vector position (FIG. 15A) when it is assumed that the signal has not been modified is the base vector obtained in step 204 (for example, FIG. 15B) or its origin. In this example, the basis vectors obtained in step 204 are (x ₁ , y ₁ ), ( x ₂ , y ₂ ), the estimated primary transformation parameter matrix is

の４種類となる。

There are four types.

  Here, if the embedded information detection unit 404 is configured to be able to detect embedded information even if the signal is a mirror image and a 90-degree rotated signal, only one of the above matrices is estimated and linearly converted. You may make it output as a parameter.

  Here, an example is shown in which the pattern constituted by the embedding information is embedded so as to form a square lattice shape in the signal. However, the embedding information is arranged in different shapes (for example, a rectangle or a parallelogram). In this case, the basis vector position when it is assumed that the signal has not been modified is different from that shown in FIG. It goes without saying that the calculation formula changes to other forms.

  The estimated primary conversion parameter 507 determined above is output as the output of the primary conversion parameter estimation unit 401.

  Next, an implementation example of the 90-degree pair search unit 503 will be described.

  Specifically, the 90-degree pair search unit 503 can be realized by performing the following procedure.

  FIG. 16 is a flowchart of an implementation example of the 90-degree pair search unit in the first embodiment.

  Step 1201) A set λ for accumulating basis vector candidates, which is an output value of the 90-degree pair search unit 503, is an empty set, and η is a set equal to a set ξ of peak point candidates 509 obtained by the fixed filter search unit 502 To do.

  Step 1202) One peak point candidate x ″ is arbitrarily extracted from the set η, and x ″ just extracted is removed from η.

  Step 1203) A position z obtained by rotating x ″ by 90 degrees counterclockwise is obtained.

  Step 1204) Let ζ be a set of elements obtained by removing x ″ from ξ.

  Step 1205) One peak point candidate y ″ is arbitrarily extracted from ζ, and y ″ just extracted is removed from ζ.

  Step 1206) It is determined whether or not y ″ exists in the vicinity of z. The criterion for determining whether or not y ″ is near is the primary conversion until a modification such as an aspect ratio change is performed as modification 107 for the signal. The determination criterion is whether or not the parameter exists within a wide area depending on whether the parameter can be estimated.

  Specific criteria for determining whether or not it is close will be described later.

  Step 1207) If y ″ is determined to be in the vicinity of z, a set of (x ″, y ″) is added to λ.

  Step 1208) It is determined whether or not ζ is empty. If it is not empty, steps 1205 to 1208 are repeated until empty.

  If ζ becomes empty, the process proceeds to step 1209.

  Step 1209) It is determined whether or not η is empty. If it is not empty, Steps 1202 to 1209 are repeated until η becomes empty.

  If η becomes empty, the process proceeds to step 1210.

  Step 1210) The basis vector candidates 510 accumulated so far in λ are output and the process ends.

Here, if the fixed filter search unit 502 is configured such that the peak point candidates 509 are limited to the range of y> 0 in the autocorrelation function array 508, η is in the range of x> 0 and y> 0. It is sufficient that the selection is limited to η ← {(x, y) | (x, y) εξ and x> 0 and y> 0} instead of η ← ξ in step 1201.
If it is said, it is efficient. Needless to say, the range to be selected is limited in accordance with the existence range of the peak point candidate 509 and the rotation direction in step 1203.

  It is assumed that the above sets λ, η, ζ, ξ are stored in storage means (not shown) such as a memory.

  The neighborhood determination method in step 1206 will be described below.

  FIG. 17 shows an example of the definition of the neighborhood in the 90-degree pair search in the first embodiment.

  In FIG. 17A, when the modification 1301 that does not include the aspect ratio change is performed as the modification 107 to the signal, the basis vectors represented by the peaks of the autocorrelation function value array 508 are represented by x and y. Further, as a modification 107 to the signal, when a modification 1302 in which an aspect ratio change is further added before and after the modification 1301, the basis vectors represented by the peaks of the autocorrelation function value array 508 are represented by x ″ and y ″. . Further, a position where x ″ is rotated by 90 degrees is set as z.

  Here, since the peak point actually observed as the peak point candidate 509 is x ″ and y ″ when the modification 1302 is added, in step 1206, z and If y ″ is within the range of possible positional relationships, it can be determined that it is in the vicinity.

  In the modification 1302, the aspect ratio changes performed before and after the modification 1301 are respectively represented by A and B below.

とすると、ｙ”はｚを用いて次のように表される。

Then, y ″ is expressed as follows using z.

α，βが１−ａ≦α≦１＋ａ，１−ｂ≦β≦１＋ｂ（但し、０＜ａ＜１，０＜ｂ＜１）を動くとき、ｚ＝（ｘ_０，ｙ_０）に対してｙ”＝（ｘ_１，ｙ_１）が動く範囲は図１７（ｂ）の双曲線と直線に囲まれた領域になる。

When α and β move 1−a ≦ α ≦ 1 + a, 1−b ≦ β ≦ 1 + b (where 0 <a <1, 0 <b <1), z = (x ₀ , y ₀ ) The range in which y ″ = (x ₁ , y ₁ ) moves is a region surrounded by a hyperbola and a straight line in FIG.

  Therefore, in step 1206, it is sufficient to determine that y ″ is in the vicinity when y ″ exists in the region of FIG.

In order to simplify the determination of the area, another graphic area that roughly includes the area shown in FIG. 17B, for example, a trapezoid area, a sector area, or the like is defined, and the vicinity is determined within that area. It may be. Furthermore, it goes without saying that the processing may be performed by modifying the expression so that the determination can be performed only by integer arithmetic in order to process the operation at high speed. For example, if the determination is made with an area cut out of the fan shape surrounded by the curve in FIG. 17C, the determination is made based on the positive / negative of the following formulas F ₁ , F ₂ , F ₃ , and F _4. Also good.

^{_{F 1 = b m 2 y 0}} x- (b m -b n) 2 x 0 y
^{_{F 2 = b m 2 y 0}} x- (b m + b n) 2 x 0 y
^{_{F 3 = (a m + a}} n) 2 (x 0 2 + y 0 2) -a m 2 (x 2 + y 2)
^{_{F 4 = (a m -a n}} ) 2 (x 0 2 + y 0 2) -a m 2 (x 2 + y 2)
However, α = 1 / α ', 1-a' and ≦ α '≦ 1 + a' , a '= a n / a m, b-b n / b m, a n, a m, b n, b m are It is an integer.

  Next, an implementation example of the detailed filter search unit 504 will be described.

  Since the autocorrelation function value array 503 is represented as an inverse Fourier transform of the power spectrum with respect to the original signal, the frequency distribution of the autocorrelation function value is a watermark pattern embedded in the signal (for example, the watermark pattern 301 in FIG. 5). It is obvious that the frequency band shape is similar to that of. In particular, in the case of a watermark pattern embedded by repeated tiling, a periodic impulse-like peak appears in the autocorrelation function value.

  It is generally known that the Fourier transform of a pattern repeated in the spatial domain has an impulse-like frequency distribution that is sparsely repeated in the frequency domain, and the embedded pattern (d) obtained by repeatedly tiling the embedded pattern in FIG. The frequency distribution is as shown in FIG.

Considering that the observed autocorrelation function value array 508 has undergone modification 107 to the signal, the base vector position (x ₀ , y ₀ ) without modification is moved to (x, y). Considering the modification amount B ^{(x, y)} in the frequency domain corresponding to the modification amount A ^{(x, y)} , the shape obtained by converting the original frequency domain shape (c) with B ^{(x, y)} and converting it to a power spectrum It can be seen that (g) is the frequency distribution of the autocorrelation function value (h). Therefore, a peak as shown in (h) of FIG. 18 appears in the autocorrelation function value array 508 depending on the magnitude of the signal value obtained by filtering the autocorrelation function value array 508 using the bandpass filter having the shape (g). It can be judged whether it is appearing.

  Therefore, for each of the basis vector candidates 510 input to the detailed filter search unit 504, a bandpass filter having a shape (g) in the frequency domain transformed corresponding to the position of the basis vector candidate is configured. After filtering the power spectrum obtained in the calculation process of the autocorrelation function value array generation unit 501, inverse Fourier transform is performed, and the evaluation value of each base vector candidate 510 is obtained with the signal intensity at the base vector position.

  Next, a specific operation of the parallel movement amount estimation in step 103 of FIG. 7 will be described.

  The estimation of the parallel movement amount in the parallel movement amount estimation unit 403 can be configured as follows.

  For example, consider the case of watermark embedding for a two-dimensional image signal. Since the watermark pattern is repeatedly embedded in the signal as shown in FIG. 5, if the size of the watermark pattern 301 is n × n, the parallel movement from (0, 0) to (n−1, n−1). The detection of digital watermark is attempted for all of the amounts, and the amount of translation when detected can be set as the amount of translation given as the modification 107.

  Here, the determination as to whether or not the digital watermark has been detected by the trial of the digital watermark may be determined based on whether or not the value indicating the accuracy of the detection result exceeds a predetermined threshold. Further, it may be determined that the detection can be performed with the amount of parallel movement when the value representing the accuracy of the detection result is maximized. For example, when digital watermark detection is performed using a symbol detection unit as described in the second embodiment of Patent Document 1, the correlation between the symbol sequence and the detection target sequence in Patent Document 1 is described above. A value may be used as a value representing the accuracy of the detection result.

  FIG. 19 is a flowchart of the operation of the parallel movement amount estimation unit in the first embodiment.

  Steps 1501, 1502, 1503) The variable max and the offset positions x and y for storing the maximum value of the detection result accuracy are initialized to zero. However, if the detection result accuracy can be a negative number, the minimum value that can be obtained the detection result accuracy is set to max.

  Step 1504) Watermark detection is performed when it is assumed that (x, y) translation is performed as the modification 107 for the signal, and a value representing the accuracy of the detection result is substituted into p.

  Step 1505) It is determined whether p is larger than max. If not larger, the process proceeds to Step 1508. If larger, the process proceeds to step 1506.

  Step 1506) p is substituted into max, and the parallel movement amounts x and y at that time are substituted into mx and my, respectively, and stored.

  Step 1507) Further, it is determined whether or not max exceeds a predetermined threshold value, and if it exceeds, the process proceeds to Step 1513. If not, the process proceeds to step 1508.

  Step 1508) Add displacement dx to lateral offset x. When the change movement amount is estimated in pixel units, dx may be 1.

  Step 1509) If x is less than n, return to Step 1504 and repeat Step 1504 to Step 1508 until x reaches n.

  Step 1510) Add dy to the vertical offset y. When estimating the amount of translation in pixel units, dy may be 1.

  Step 1511) If y is less than n, the process returns to Step 1503, and Steps 1503 to 1510 are repeated until y reaches n.

  When y reaches n, the parallel movement amounts mx and my giving the maximum detection result accuracy value in step 1512 are output as the estimated parallel movement amount 408.

  In step 1507, when max exceeds the threshold value, the estimated parallel movement amount may be determined without searching for all the parallel movement amounts by the procedure of steps 1513 to 1521.

  Steps 1513 and 1514) -k is substituted for the displacement amounts xx and yy, respectively.

  Step 1515) Watermark detection is performed when it is assumed that (x + xx, y + yy) has been translated as the modification 107 for the signal, and a value representing the accuracy of the detection result is substituted into p.

  Step 1516) It is determined whether or not p is larger than max. If larger, the process proceeds to step 1517.

  Step 1517) Substituting p for max, and substituting the parallel movement amounts x + xx and y + yy for mx and my, respectively, and storing them.

  Step 1518) The displacement dx is added to the lateral displacement xx. When estimating the amount of translation in pixel units, dx may be 1.

  Step 1519) If xx is less than or equal to k, the process returns to Step 1515, and Steps 1515 to 1518 are repeated until xx reaches k.

  Step 1520) Add dy to the longitudinal displacement yy. When estimating the amount of translation in pixel units, dy may be 1.

  Step 1521) If yy is less than or equal to k, return to Step 1514 and repeat Step 1514 to Step 1520 until yy reaches k.

  When yy reaches k, detection is performed in the vicinity of the offset position (x, y) exceeding the threshold value in steps 1513 to 1521, and the estimated parallel movement is determined with the maximum amount of parallel movement among them. By setting the amount, the estimated parallel movement amount can be determined efficiently.

  Further, as described in Patent Document 1 described above, when the digital watermark information is divided into a plurality of symbols and a signal is embedded, the above-described parallel movement amount search is performed on one of the plurality of symbols. And the resulting parallel movement amount is output as the estimated parallel movement amount 408, and the detection processing of the remaining symbols is performed by the embedded information detection unit 404, thereby efficiently detecting the digital watermark. Can do.

[Second Embodiment]
In the present embodiment, a configuration example for more efficiently processing the detailed filter search unit 504 in the first embodiment will be described.

  The definition of terms and the configuration other than the detailed filter search unit 503 are the same as those in the first embodiment.

  In the realization method of the detailed filter search unit 504 in the first embodiment described above, since it is necessary to perform inverse Fourier transform processing on each of the basis vector candidates, the processing takes a very long time, and the pattern A filter having a repetitive shape of an impulse-like peak with a frequency distribution determined corresponding to repetition has a problem that it is difficult to configure with high accuracy due to a rounding error problem of the discrete Fourier transform.

  Therefore, the detailed filter search unit 504 may be more efficiently realized by performing filtering on a predetermined range around candidate points in the spatial domain instead of performing filtering in the frequency domain.

  A detailed filter search for a two-dimensional signal will be described below as an example.

  When the embedding pattern has a frequency distribution as shown in FIG. 18A, individual peaks appearing in the autocorrelation function value (FIG. 20A) after only rotation is applied are shown in FIG. It has a frequency distribution like The impulse response of a low-pass filter passing through a rectangular area is the sinc function

となることから、ピークの空間領域での形状は、図２０（ｃ）のようなｓｉｎｃ関数の形状となる。但し、図２０（ｃ）は簡略化のため一次元で表現している。

Therefore, the shape of the peak in the space region is a sinc function shape as shown in FIG. However, FIG. 20C is expressed in one dimension for simplification.

  Further, the shape of the autocorrelation peak appearing in the autocorrelation function value array 508 depends on how much the watermark pattern embedded in the signal (for example, the watermark pattern 301 in FIG. 5) is enlarged or reduced by the modification 107 to the signal. It becomes gentle and sharp. This is because the frequency band in which the watermark information is embedded is shifted by scaling. This is shown in FIGS. 20 (d) to 20 (f).

  Each peak appearing as a basis vector in the autocorrelation peak (FIG. 20D) after being reduced by 0.5 times and rotated has a frequency distribution as shown in FIG. 20E, and as a result, As shown in FIG. 20 (f), the shape of the peak in the spatial region becomes a shape of a sinc function that is more sensitive than when it is 1.0 times. In addition, the shapes in these two-dimensional spaces are rotated by the same amount as the rotation applied to the signal.

  Therefore, the theoretical peak shape can be expressed by the shape of a sinc function determined for each modification amount corresponding to the position where the autocorrelation peak appears.

  The spatial domain shape for the frequency distribution in FIG. 18A can be easily and quickly obtained using a sinc function without inverse Fourier transform by combining a rectangular domain pass low-pass filter. That is, a filter having a frequency distribution of an embedding pattern (FIG. 18A) is considered as a parallel type and a series type filter of a plurality of rectangular area pass type low pass filters. Further, the impulse response of the parallel type and the series type filter is Therefore, the theoretical peak shape of interest can be obtained only by calculating the sum and product of the sinc function shapes, which are the impulse responses of the respective rectangular area passing low-pass filters. Can be sought.

  Furthermore, the target theoretical peak shape can be obtained by performing coordinate conversion for rotation and enlargement / reduction in the spatial domain according to the rotation amount and enlargement / reduction amount corresponding to the target peak point position.

  Since the range of the sinc function used in the calculation is limited by the target frequency range, the sinc function value can also be expressed in the form of a look-up table, and in this case, higher speed processing is possible.

  Using the theoretical peak shape obtained above, for each point of the basis vector candidate 510, the theoretical spatial shape corresponding to the candidate point position and the autocorrelation function value of a predetermined region around each point By performing the filtering for obtaining the correlation value, it is possible to efficiently obtain the evaluation value only by the correlation calculation in the spatial domain without requiring the inverse Fourier transform for each of the basis vector candidates. For example, the correlation value with the peak shape 1702 obtained from the theoretical frequency distribution determined according to the candidate point position is obtained for the autocorrelation function value array 1701 in FIG. FIG. 21 shows an example of a one-dimensional signal, but it goes without saying that the same applies to an n-dimensional signal.

  Furthermore, since the autocorrelation function value series 508 is expressed as a set of discrete points, the basis vector candidate 510 is also given as a point above the lattice point, but the true basis vector position is at an intermediate position of the lattice point. In some cases, a small amount of deviation is considered for each point of the basis vector candidate 510, and when the amount of deviation is moved, a deviation amount that maximizes the correlation value between the peak shape and the autocorrelation function value array is obtained. A base vector with higher accuracy may be obtained by redefining it as a base vector candidate point. For example, with respect to the autocorrelation function value array 1704 in FIG. 21B, a correlation value with the theoretical peak shape 1705 having a peak at a position slightly shifted leftward from the base vector candidate point is obtained.

  In the case of a two-dimensional signal, for each point of the basis vector candidate 510, for example, 1/10 subpixels of the lattice point interval a are sequentially displaced in the x direction and the y direction, and a peak point exists at each position. Then, the correlation between the theoretical peak shape on the assumption and the observed autocorrelation value is obtained by convolution operation on the lattice points around the original base vector candidate point, and the one with the largest correlation value is determined as the evaluation value. do it. Further, the position of the basis vector candidate may be corrected using the displacement amount at that time.

[Third Embodiment]
In the present embodiment, a configuration example for processing the primary conversion parameter estimation unit 401 in the first and second embodiments described above with higher accuracy will be described below.

  Definitions of terms, configurations other than the primary transformation parameter estimation unit 401, and configurations of parts of the primary transformation parameter estimation unit 401 that are not described below are the same as those in the first and second embodiments.

  In the primary conversion parameter estimation unit 401 of the first and second embodiments described above, the estimated base vector 511 output from the detailed filter search unit 504 erroneously includes a peak point obtained depending on the original base vector. It may be included as a basis vector. Here, the peak point obtained depending on the original basis vector indicates, for example, the peak point (b) represented as a linear combination of the correct basis vectors (a) in FIG. There is a possibility that a high evaluation value can be obtained by the detailed filter search unit even for a basis vector candidate constituted by such points, but a correct estimated primary transformation parameter cannot be obtained from such a basis vector candidate. There is a possibility that the detection of the digital watermark will fail. In the present embodiment, an example of removing such basis vector candidates will be described.

  FIG. 23 illustrates a configuration example of a primary conversion parameter estimation unit that performs dependent peak point removal in the third embodiment.

  The configuration shown in the figure is a configuration in which a dependent peak point removing unit 1912 is added to the primary conversion parameter estimating unit 401 in FIG. 8.

  The dependent peak point removing unit 1912 removes base vector candidates that become dependent peak points from the estimated basis vector 1911 output from the detailed filter searching unit 1904 (same as the detailed filter searching unit 504 in FIG. 8). The remaining one is output as an estimated basis vector 1913 from which the dependent peak points have been removed.

  Further, instead of removing the base vector that becomes the dependent peak point, the evaluation value obtained in the detailed filter search unit 1904 corresponding to the original base vector and the base vector that becomes the dependent peak point are obtained. Based on the obtained evaluation value, each evaluation value may be changed so that the evaluation value of the original base vector is necessarily larger than the evaluation value of the base vector serving as the dependent peak point. By changing the evaluation value in this manner, it is possible to prevent the detection of the watermark from failing by removing the correct base vector from the estimated base vector 1911 due to an incorrect estimated base vector obtained due to noise on the signal. can do.

  The conversion parameter estimation unit 1905 obtains an estimated primary conversion parameter 1907 based on the estimated base vector 1913 from which the dependent peak point output from the dependent peak point removing unit 1912 has been removed.

  FIG. 24 is a flowchart of the operation of the dependent peak point removing unit in the third embodiment.

  Step 2001) A set of pairs (x, y) of estimated basis vectors 1911 output from the detailed filter search unit 1904 is denoted by ξ.

  Step 2002) Let η be a set having the same elements as ξ.

  Step 2003) One estimated basis vector pair (x, y) is arbitrarily extracted from η, and (x, y) just extracted is removed from η.

  Step 2004) A set having all elements other than (x, y) in ξ is set as ζ, and a set ω for storing dependent peak points to be searched for is set as an empty set.

  Step 2005) One estimated basis vector pair (x ′, y ′) is arbitrarily extracted from ζ, and the currently extracted (x ′, y ′) is removed from ζ.

  Step 2006) When (x ′, y ′) is at a subordinate position of (x, y) and the evaluation value p [(x, y) obtained by the detailed filter search unit 1904 for (x, y) ] Is smaller than the evaluation value p [(x ′, y ′)] for (x ′, y ′).

  If the condition is not satisfied, the process proceeds to step 2008. If the condition is satisfied, the process proceeds to step 2007.

  Step 2007) (x ′, y ′) is added to the set ω in which the dependent peak points are accumulated, and the process proceeds to Step 2008.

  Step 2008) It is determined whether or not ζ is empty, and if not empty, the process returns to Step 2005, and Steps 2005 to 2008 are repeated until empty.

  If ζ becomes empty, the process proceeds to step 2009.

  Step 2009) The evaluation value of each estimated basis vector is changed as follows.

  The evaluation value p [(x, y)] of (x, y) is changed to the maximum evaluation value among the evaluation values of the found dependent peaks (that is, pairs of all estimated basis vectors included in ω).

  The evaluation value p [(x ′, y ′)] of the found dependent peak (that is, all estimated basis vector pairs included in ω) is changed to the original evaluation value before the change of (x, y). To do.

  That is, by changing the evaluation value of the original estimated basis vector to be always larger than the evaluation value of the estimated basis vector that is the dependent peak point, the dependent peak point may be prioritized and detected. There can be no.

  Step 2010) It is determined whether η is an empty set. If it is an empty set, the process proceeds to Step 2003, and Steps 2003 to 2010 are repeated until η is empty.

  The above sets η, ζ, ξ, ω are assumed to be stored in storage means (not shown) such as a memory.

  Here, in step 2009, an example is shown in which the evaluation value of the original estimated basis vector is exchanged with the evaluation value of the estimated basis vector that is the dependent peak point. However, any other modification method may be used as long as the modification method is such that the evaluation value of the estimated base vector serving as the dependent peak point is always larger than the evaluation value. For example, the evaluation value of the original estimated base vector is assumed to be 1.1 times the maximum evaluation value of the estimated base vector that is the dependent peak point, and the evaluation value of the estimated base vector that is the dependent peak point is It may not be changed.

  Further, as a determination condition in step 2006, the ratio of the evaluation value p [(x ′, y ′)] to (x ′, y ′) and the evaluation value p [(x, y)] to (x, y), or You may add to a condition that a difference is less than a predetermined threshold value. Usually, the dependent peak point and the original peak point are obtained as the same level of peaks in the autocorrelation function value array 1908. Therefore, if the evaluation values are extremely different, the basis vector of (x, y) is There is a possibility that it is an incorrect basis vector obtained by noise. For this reason, by adding the above conditions, if the evaluation value is too far away, the correct basis vector is erroneously removed without performing the dependent peak point removal processing on the corresponding basis vector. Can be prevented.

  Here, the explanation is made on the assumption that the larger the value of the evaluation value p is, the more likely that it is a basis vector, but conversely, the smaller the value of the evaluation value p is, the more likely that the basis vector is. Needless to say, when an evaluation index representing the above is used, the manner of increasing or decreasing the evaluation value is reversed.

  Finally, an example of the subordinate position is shown.

  FIG. 25 shows the dependent positions in the third embodiment. When the basis vector pair (A, B) is limited to the first and second quadrants, a pair that may be erroneously recognized when considering up to a lattice point that is twice the distance is shown in FIGS. k).

[Fourth Embodiment]
FIG. 26 shows a configuration example of the digital watermark detection apparatus in the fourth embodiment of the present invention.

  The digital watermark detection apparatus 2200 of the present invention shown in the figure includes a signal reduction unit 2201, an autocorrelation function calculation unit 2202, an autocorrelation function value normalization unit 2203, an autocorrelation peak point evaluation unit 2204, and an autocorrelation peak point evaluation value normalization. The image forming unit 2205, the detection target determining unit 2206, the geometric correcting unit 2207, and the embedded information detecting unit 2208.

  The digital watermark detection process in this embodiment is largely performed through the following steps.

  FIG. 27 is a flowchart of digital watermark detection processing according to the fourth embodiment of the present invention.

Step 401) Signal reduction:
The signal reduction unit 2201 reduces the input signal 2209 input to the digital watermark detection apparatus 2200 to a plurality of sizes such as 1 time, 1/2 time, 1/4 time,.

  For example, when the image signal is a two-dimensional image signal, the image is reduced so that the image size is spatially reduced. The reduction of the image is expressed, for example, as the expansion and contraction, which is a geometric transformation described in the document “Koji Watanabe,“ Image Engineering Handbook ”, Asakura Shoten (1986). Interpolation processing such as a method may be performed.

  Here, in the configuration example of FIG. 26, an example is given in which the reduction ratio is a power of ½, but any reduction ratio may be used as long as the value corresponds to the correction in the subsequent geometric correction unit 2207. I do not care. For example, it may be configured to be 1 / 1.5 times, 1/3 times, or 1/5 times.

  Hereinafter, processing is performed for each reduced signal obtained by the signal reduction unit 2201, and the processing for each reduced signal is referred to as "range".

Step 402) Autocorrelation function calculation:
An autocorrelation function calculation unit 2202 calculates an autocorrelation function for each reduced signal obtained by the signal reduction unit 2201. As a method for calculating an autocorrelation function value, for example, a known procedure such as inverse discrete Fourier transform of a power spectrum obtained by performing discrete Fourier transform on an input signal is known. The autocorrelation function calculation unit 2202 corresponds to the autocorrelation function value array generation unit 501 in the first embodiment.

Step 403) Autocorrelation function value normalization:
The autocorrelation function value normalization unit 2203 normalizes the value for each range with respect to the autocorrelation function value obtained by the autocorrelation function calculation unit 2202. In this normalization, normalization is performed so that autocorrelation function values for each range have similar distributions. For example, in the case of a two-dimensional image signal, for example, an autocorrelation function value obtained using a discrete Fourier transform having a size of 512 × 512 is obtained in the form of an array having a size of 512 × 512. The following value conversion is performed so that the distribution of the values of each element of the array has an average value of 0 and a variance of 1.0 in each range.

但し、ｆ（ｘ，ｙ）は位置ｘ，ｙにおける自己相関関数値、ｍはｆ（ｘ，ｙ）の平均値、σはｆ（ｘ，ｙ）の標準偏差であり、ｆ’（ｘ，ｙ）は、位置ｘ，ｙにおける変換後の自己相関関数値とする。

Where f (x, y) is an autocorrelation function value at positions x and y, m is an average value of f (x, y), σ is a standard deviation of f (x, y), and f ′ (x, y y) is an autocorrelation function value after conversion at the positions x and y.

  By performing normalization of the autocorrelation function value, the evaluation value calculated later in the autocorrelation peak point evaluation unit 2204 based on the autocorrelation function value has a similar distribution among the ranges. It is possible to suppress the evaluation of erroneous peak points in other ranges.

  When obtaining the average m and standard deviation σ of the autocorrelation function value f (x, y), the value near the origin of the original autocorrelation function value may be ignored and the value calculated. That is,

ここで、上記の式では原点付近を除く領域として原点からの距離が所定の値ｒを越えている

Here, in the above formula, the distance from the origin exceeds the predetermined value r as an area excluding the vicinity of the origin.

であることを条件としたが、例えば、原点を囲む矩形領域を除く領域（│ｘ│＞ａ，│ｙ│＞ｂ）であることを条件としてもよいし、その他の定義であっても構わない。

However, for example, it may be a condition excluding a rectangular area surrounding the origin (| x |> a, | y |> b), or other definitions may be used. Absent.

  The value near the origin of the autocorrelation function is a correlation value with itself in the signal, and always takes a large value (correlation value 1.0) at the origin. For example, a pixel adjacent to a pixel that is a signal such as an image It is clear that the autocorrelation value near the origin also takes a relatively large value because the correlation of the origin is high, but the origin correlation value is meaningless in the peak point search to estimate the magnitude of the modification Therefore, by performing normalization in the area from which this is removed, the evaluation values can be more effectively aligned between ranges.

Step 404) Autocorrelation peak point evaluation:
The autocorrelation peak point evaluation unit 2204 uses the autocorrelation function value for each range normalized by the autocorrelation function value normalization unit 2203 to search for an autocorrelation peak point and to obtain an evaluation value for each peak point candidate. calculate.

  The search for the peak point and the calculation of the evaluation value for the peak point candidate may use, for example, a value obtained by applying HPF to the autocorrelation function value as described in Patent Document 3 described above. Evaluation values based on detailed filter search as shown in the first and second embodiments may be used.

  As an output of the autocorrelation peak point evaluation unit 2204, a set of one or a plurality of peak point candidates and an evaluation value corresponding thereto is output for each range. In addition, like the basis vector candidates obtained by the 90-degree pair search unit 503 in the first embodiment described above, a set of basis vectors whose peak point candidates correspond to the dimensions of the signal (for example, a two-dimensional image) In the case of a signal, it may be expressed as a pair of two basis vectors).

  The autocorrelation peak point evaluation unit 2204 corresponds to the fixed filter search unit 502, the 90 degree pair search unit 503, and the detailed filter search unit 504 of the primary transformation parameter estimation unit 401 in the first embodiment described above. However, the present invention is not limited to the same configuration as each of these parts. That is, the autocorrelation peak point evaluation unit 2204 may be configured in any way as long as it is configured to search for an autocorrelation peak point and calculate an evaluation value for each peak point candidate. It may be configured differently from the embodiment.

Step 405) Autocorrelation peak point evaluation value normalization:
The autocorrelation peak point evaluation value normalization unit 2205 normalizes each evaluation value obtained by the autocorrelation peak point evaluation unit 2204 for each range. This normalization is performed so that the evaluation values of the peak point candidates for each range have the same degree of distribution. For example, for n-number of ranges, the peak point candidate _p 1, 1 of each m sets each each _{range, ..., p 1, m,} ..., p n, 1, ..., p n, m is obtained In this case, normalization is performed so that, for example, the average is 0 and the standard deviation is 1.0 for each range.

但し、ｅ（ｐ_ｎ，ｍ）はピーク点候補ｐ_ｎ，ｍに対する評価値ｅ’（ｐ_ｎ，ｍ）はピーク点候補ｐ_ｎ，ｍに対する正規化後の評価値、Ｍ_ｎはｅ（ｐ_ｎ，１），…，ｅ（ｐ_ｎ，ｍ）の平均値、σ_ｎはｅ（ｐ_ｎ，１），…，ｅ（ｐ_ｎ，ｍ）の標準偏差とする。

_However, e _{(p n, m)} is the evaluation value e _{'(p n, m)} to the peak point candidates _{p n, m} evaluation value after normalization to the peak point candidate _{p n, m, M n} is e (p _{n, 1} ),..., e (p _{n, m} ), and σ _n is the standard deviation of e (p _{n, 1} ),..., e (p _{n, m} ).

  By normalizing the autocorrelation peak point evaluation value in this way, the evaluation value of each peak point can have the same distribution between each range, and the wrong peak point of other ranges Can be prevented from being highly evaluated.

Step 406) Determination of detection target:
In the detection target determining unit 2206, the peak point candidates to be detected by the embedded information detection unit 2208 are detected using the peak point evaluation values obtained by the autocorrelation peak point evaluation value normalizing unit 2205. select.

  The selection method may be the simplest method of selecting the top K (K ≧ 1) in order from the highest evaluation value of each peak point across the range.

  In addition, when a plurality of peak points at very close positions or the same position are listed as candidates, the duplication is removed, and a single peak point candidate is collectively collected, and the detection target is It is also possible to determine the embedded information, whereby it is possible to suppress the redundant detection of the embedded information with respect to the case where the result of the geometric correction described later is almost the same, and it is possible to detect without waste.

Step 407) Conversion parameter calculation and geometric correction:
The geometric correction unit 2207 calculates a conversion parameter corresponding to each peak point candidate of the detection target determined by the detection target determination unit 2206 and represents the amount of modification applied to the signal, and then applies the input signal 2209 to the input signal 2209. Then, the signal is corrected by performing the inverse transformation of the modification using the transformation parameter. This is because, as described in the second embodiment, the position of the point appearing as the peak point of the autocorrelation function value is determined according to the amount of modification such as enlargement / reduction or rotation added to the signal. We are using.

  At this time, inverse conversion is performed for each range in consideration of calculating the autocorrelation function after the input signal is reduced in the signal reduction unit 2201. For example, when a two-dimensional image signal is taken as an example, it is assumed that a pattern repeatedly embedded in a signal as a digital watermark has a size of 128 × 128 pixels. In the signal reduction unit 2201, when peak points appear at the positions (96, 0) and (0, 96) of the autocorrelation function value in the range reduced to ½, this is the original (that is, the signal is reduced). (If the autocorrelation function is not calculated), a peak point appears at the position (192, 0) (0, 192). As a result, the signal is modified by a factor of 1.5. It means that the expansion of was done. Therefore, in this case, the geometric correction unit 2207 reduces the input signal 2209 by 1 / 1.5. If a peak point appears at the same position of the autocorrelation function value in the range reduced to 1/4, a reduction of 1 / 3.0 is performed.

  Note that the detection target determination in the detection target determination unit 2206 and the geometric correction in the geometric correction unit 2207 may be reversed in order. That is, before performing detection target determination, geometric correction is performed on all peak point candidates obtained by the autocorrelation peak point evaluation value normalization unit 2205, and then the detection target is actually determined. May be.

Step 408) Embedded information detection:
An embedded information detection unit 2208 detects embedded information based on the corrected signal obtained by the geometric correction unit 2207. The embedded information detection method may be any method depending on the digital watermark embedding method. For example, the digital watermark detection method according to Patent Document 1 described above may be used.

  According to the detected information, it is judged to what extent the detection has been performed, and the detection result with the highest detection accuracy or the detection result with the detection accuracy exceeding a predetermined threshold is output. It may be. For example, when digital watermark detection is performed using a symbol detector as described in the second embodiment of Patent Document 1, the correlation value between the symbol sequence and the detection target sequence in Patent Document 1 is described. May be used as a value representing the detection accuracy.

  Further, in order to perform processing efficiently, detection is performed on the target to be detected in order from a corrected signal of a candidate having a large evaluation value according to the evaluation value obtained by the autocorrelation peak point evaluation value normalization unit 2205. When the detection accuracy exceeds a predetermined threshold, the remaining detection processing may be stopped and the detection result of detection exceeding the threshold may be output. As a result, the embedded information detection process can be efficiently performed even when a plurality of candidates are obtained, and the autocorrelation peak point evaluation value of the correct peak point is the wrong range and the peak point candidate. Even in the case where the evaluation is slightly lower than the evaluation value, the embedded information detection can finally be succeeded, and the embedded information detection with high reliability can be performed.

  Next, the detection target determination process in step 406 and the efficiency of embedded information detection in step 408 will be described.

  When performing embedded information detection processing multiple times for peak point candidates in multiple ranges, detection failure due to selecting an incorrect peak between ranges that can occur even if evaluation value normalization is performed An efficient detection target determination method that limits the upper limit of the time required for detection processing while compensating will be described.

FIG. 28 is a flowchart of detection object determination in the fourth embodiment of the present invention, and FIG. 29 shows an example of detection object determination in the fourth embodiment of the present invention (step 2301).
-Number of peak point candidates to be prioritized in each range L
-The upper limit number M of embedding information detection processing as a whole
Decide. These may be fixed values set in the digital watermark detection apparatus 2200, or may be input at the start of the digital watermark detection process. FIG. 29 shows examples of peak point candidates and evaluation values for each range, and detection targets determined from these, and here, L = 2 and M = 8.

  For the input autocorrelation peak point candidate, the following processing is performed for each range. Here, i is a range to be processed.

Step 2302) Let ζ _{i be} a set of autocorrelation peak point candidates in range i, and let p (x) be a peak point evaluation value for x in x∈ζ _i .

Step 2303) from the large ones of the autocorrelation peak point evaluation value p (x) in order for X∈zeta _i for the upper L peaks point candidate, adding a predetermined numerical value p ₀ autocorrelation peak point evaluation value. In other words, the p (x) ← p (x ) + p 0. Determined in advance to take a large value the value of the time p ₀ is than either of the autocorrelation peak point evaluation value. For example, in range 1 of FIG. 29, 10000 is added as p0 for peak point candidates (u, v) and (x, y) whose evaluation values are large as 241, 212, and the respective evaluation values are changed to 10021, 100212. is doing. FIG. 29 also shows the above-described removal of overlap. For example, in range 1, one of the overlapping peak point candidates (x, y) is removed.

  Step 2304) It is determined whether processing has been performed for all ranges i. If there is an unprocessed range, i is replaced with i and steps 2302 to 2303 are repeated.

  Step 2305) Let η be the set of peak point candidates for all ranges.

  Step 2306) The top M peak point candidates are extracted in descending order of the autocorrelation peak point evaluation value p (x) for x∈η, and the set is set as ζ. For example, in FIG. 29, eight peak point candidates (n, o), (u, v), (x, y),..., (C, d) are selected.

  Step 2307) ζ is output as the determined detection target.

  Thus, by changing the top L peak point candidates so as to increase the peak point evaluation value, the top L peak point candidates can be preferentially set as targets for embedded information detection. As a result, even if there are many wrong peak point candidates in the wrong range, the peak point evaluation value in each range has a larger evaluation value than the correct peak point candidate evaluation value in the correct range. It is possible to prioritize the peak point candidates at the top and make them the targets for embedded information detection. As a result, it is possible to suppress the detection failure due to an erroneous peak point candidate, and the upper limit number of peak point candidates to be detected as a whole can be limited to a predetermined number. The upper limit can be limited.

Moreover, according to said structure, for example,
(A) If L = 0 and M = 1, embedded information detection is performed for one peak point candidate having the highest evaluation through all ranges.

  (B) If L = 0 and M = 3, embedded information detection is performed for three peak point candidates from the ones that are highly evaluated through all ranges.

  (C) If L = 1 and M = the number of ranges, one peak point candidate with the highest evaluation is selected for each range, and embedded information detection is performed for each peak point candidate for each range.

  (D) If L = 2 and M = 10, embedding information detection is performed for 10 peak point candidates as a whole by giving priority to the two peak point candidates from the ones with high evaluation for each range.

  Thus, various embedded information detection methods can be determined, and the detection operation can be easily changed without changing the configuration of the apparatus.

In particular, the signal 101 to be embedded with the digital watermark includes, for example, a signal with strong autocorrelation such as a repetitive pattern or a straight line in the case of an image, or is compressed and encoded at a high compression rate. If the autocorrelation peak you want to find is buried in the autocorrelation peak (false peak) of the signal,
(I) When many false peaks are included in the same range;
(Ii) There are few false peaks in the correct peak range, but many false peaks are included in other ranges;
Therefore, it is important to detect the embedded information in consideration of which is more likely depending on the condition. For example, the method (b) is suitable for the case (i), and the method (c) is suitable for the case (ii). It is possible to detect a digital watermark.

  In addition, the operation | movement of the structure shown in FIG.6, FIG.8, FIG.23, FIG.26 in said 1st-4th embodiment is constructed | assembled as a program, and it installs and performs it on the computer utilized as a digital watermark detection apparatus. Or can be distributed via a network.

  It is also possible to store the constructed program in a portable storage medium such as a hard disk device, a flexible disk, or a CD-ROM, and install or distribute the program in a computer.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made within the scope of the claims.

  The present invention is applicable to digital watermark detection technology.

It is a principle block diagram of this invention. It is a figure for demonstrating the principle of this invention. It is a digital watermark embedding / detection model as the background of the present invention. It is an example of composition of a digital watermark embedding device in a 1st embodiment. It is an example of tiling of a watermark pattern in the first embodiment. It is a block diagram of the digital watermark detection apparatus in 1st Embodiment. 3 is a flowchart of digital watermark detection processing in the first embodiment. It is an example of a structure of the primary conversion parameter estimation part in 1st Embodiment. It is a flowchart of the estimation procedure of the primary conversion parameter in 1st Embodiment. It is an example of the autocorrelation function value array in 1st Embodiment. It is an example of the Laplacian filter in 1st Embodiment. It is an example of the shape of the fixed filter in 1st Embodiment. It is an example of the search range of the autocorrelation peak point in 1st Embodiment. It is an example of a 90 degree pair (base vector) candidate search in 1st Embodiment. It is an example of the conversion parameter calculation from the basis vector in 1st Embodiment. It is a flowchart of the 90-degree pair search part implementation example in 1st Embodiment. It is an example of the definition of the vicinity in the 90 degree pair search in 1st Embodiment. It is explanatory drawing of the frequency distribution of the autocorrelation function value in 1st Embodiment. It is a flowchart of operation | movement of the parallel displacement estimation part in 1st Embodiment. It is a frequency distribution of the autocorrelation peak point shape in 2nd Embodiment. It is a figure for demonstrating evaluation value calculation of the base vector candidate in the sub pixel unit in 2nd Embodiment. It is an example of the recognition error of the vector subordinate to the basis vector in 3rd Embodiment. It is an example of a structure of the primary transformation parameter estimation part which performs the dependent peak point removal in 3rd Embodiment. It is a flowchart of operation | movement of the dependent peak point removal part in 3rd Embodiment. It is an example of the subordinate position in 3rd Embodiment. It is a structural example of the digital watermark detection apparatus in the 4th Embodiment of this invention. It is a flowchart of the digital watermark detection process in the 4th Embodiment of this invention. It is a flowchart of the detection target determination process in the 4th Embodiment of this invention. It is an example of detection object determination in the 4th Embodiment of this invention.

Explanation of symbols

10 signal reduction means 20 conversion parameter calculation means 101 pre-embedding signal 102 embedding information 104 embedded signal 105 communication path 106 additive noise 107 signal modification 108 modified embedded signal 110 detection information 200 digital watermark embedding apparatus 201 watermark pattern generation Unit 202 watermark pattern addition unit 203 watermark pattern 301 watermark pattern 400 digital watermark detection apparatus 401 primary transformation parameter estimation unit 402 geometric transformation correction unit 403 parallel movement amount estimation unit 404 embedded information detection unit 405 input signal 406 estimated primary transformation parameter 407 geometric transformation Corrected signal 408 Estimated parallel movement amount 409 Watermark detection information 501 Autocorrelation function value array generation unit 502 Fixed filter search unit, fixed filter search unit 503 90 degree pair search unit, 90 Degree pair search unit 504 detailed filter search unit, detailed filter search unit 505 conversion parameter calculation unit, conversion parameter estimation unit 506 input signal 507 estimated primary conversion parameter 508 autocorrelation function value array 509 peak point candidate 510 basis vector candidate 511 estimated basis vector 1301 Modification 1302 that does not include aspect ratio modification Modification 1301 that modifies aspect ratio before and after 1301 Sample 1702 on autocorrelation function value array Peak shape 1703 obtained from theoretical frequency distribution determined according to candidate point position Grid point position Theoretical value 1704 corresponding to the value 1704 Autocorrelation function value array 1705 Theoretical peak shape 1901 having a peak slightly shifted leftward from the base vector candidate point Autocorrelation function value array generation unit 1902 Fixed filter search unit 1903 90 degree pair Search unit 1904 Detailed filter search unit 1905 Input parameter 1906 Input signal 1907 Estimated primary conversion parameter 1908 Autocorrelation function value array 1909 Peak point candidate 1910 Base vector candidate 1911 Estimated base vector 1912 Dependent peak point remover 1913 Remove dependent peak point Estimated base vector 2200 digital watermark detection apparatus 2201 signal reduction unit 2202 autocorrelation function calculation unit 2203 autocorrelation function value normalization unit 2204 autocorrelation peak point evaluation unit 2205 autocorrelation peak point evaluation value normalization unit 2206 detection target determination unit 2207 Geometric correction unit 2208 Embedding information detection unit 2209 Input signal 2210 Watermark detection information

Claims (11)

A digital signal that is a two-dimensional image signal is embedded in advance as a digital watermark so that the predetermined pattern is not repeatedly perceived by human perception. As a modification, enlargement or reduction, rotation, and aspect ratio change are performed. When a digital signal subjected to any primary conversion is input , the digital watermark detection apparatus estimates a conversion parameter representing the primary conversion added to the digital signal,
A signal reduction means for reducing the image size of the digital signal, which is a two-dimensional image signal to which the input modification is applied, at a predetermined magnification;
An autocorrelation function calculating means for calculating an autocorrelation value of a digital signal which is a two-dimensional image signal reduced by the signal reduction means, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
In the autocorrelation function value array obtained by the autocorrelation function calculation means, an autocorrelation peak having a peak autocorrelation value is searched, and the position of the autocorrelation peak value stored in the autocorrelation function value array is searched. Near each position of the autocorrelation function value array within a predetermined range in the vicinity and the position where the theoretical autocorrelation peak value that appears in the autocorrelation function value array due to the periodicity of the pattern repetition is stored Autocorrelation peak point evaluation means for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within a predetermined range at,
In accordance with the evaluation value obtained by the autocorrelation peak point evaluation unit, a detection target determination unit that determines autocorrelation peak point candidates to be used as detection targets for embedded information;
The position where the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined by the detection target determining means is stored in the autocorrelation function value array, and the autocorrelation function when the alteration is not made Conversion parameter calculation means for calculating a conversion parameter for performing the primary conversion applied to the digital signal from a deviation from a position where the autocorrelation peak appearing in the value array is stored in the autocorrelation function value array; ,
A digital watermark detection apparatus comprising: A digital signal that is a two-dimensional image signal is embedded in advance as a digital watermark so that the predetermined pattern is not repeatedly perceived by human perception. As a modification, enlargement or reduction, rotation, and aspect ratio change are performed. When a digital signal subjected to any primary conversion is input, the digital watermark detection apparatus estimates a conversion parameter representing the primary conversion added to the digital signal,
A signal reduction means for reducing the image size of the digital signal, which is a two-dimensional image signal to which the input modification is applied, at a predetermined magnification;
An autocorrelation function calculating means for calculating an autocorrelation value of a digital signal which is a two-dimensional image signal reduced by the signal reduction means, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
In the autocorrelation function value array obtained by the autocorrelation function calculation means, an autocorrelation peak having a peak autocorrelation value is searched, and the position of the autocorrelation peak value stored in the autocorrelation function value array is searched. Near each position of the autocorrelation function value array within a predetermined range in the vicinity and the position where the theoretical autocorrelation peak value that appears in the autocorrelation function value array due to the periodicity of the pattern repetition is stored Autocorrelation peak point evaluation means for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within a predetermined range at,
In accordance with the evaluation value obtained by the autocorrelation peak point evaluation unit, a detection target determination unit that determines autocorrelation peak point candidates to be used as detection targets for embedded information;
The position where the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined by the detection target determining means is stored in the autocorrelation function value array, and the autocorrelation function when the alteration is not made Conversion parameter calculation means for calculating a conversion parameter for performing the primary conversion applied to the digital signal from a deviation from a position where the autocorrelation peak appearing in the value array is stored in the autocorrelation function value array; ,
The digital signal subjected to the modification is subjected to geometric correction by performing an inverse transformation on the conversion parameter, and the geometrically corrected digital signal is translated by a predetermined movement amount within a predetermined region, Attempts to detect a digital watermark from a digital signal that has been translated, and a translation amount estimation means that sets the amount of movement when detected to be a translation amount added to the digital signal;
A digital watermark detection apparatus comprising: If the predetermined magnification is plural, each digital signal is a two-dimensional image signal applied with the modified reduced at each magnification, prior Symbol autocorrelation value obtained by the autocorrelation function calculating means mean and variance or as a standard deviation is the same, further has a self-correlation function value normalization means for normalizing the values of the autocorrelation, the digital watermark detection apparatus according to claim 1 or 2 wherein the. If the predetermined magnification is plural, each digital signal is a two-dimensional image signal applied with the modified reduced at each magnification, all evaluation values obtained by the autocorrelation peak point evaluation unit mean and standard deviation further comprises an autocorrelation peak point evaluation value normalizing means for normalizing the evaluation values such that the same becomes, the electronic watermark detecting apparatus according to any one of claims 1 to 3. When the predetermined magnification is plural, the detection target determining means
Priority peak point candidate number storage means for storing the number L of peak point candidates to be prioritized;
Embedded information detection upper limit number storage means for storing the number M of peak points to be output as detection targets;
For each digital signal that is a two-dimensional image signal that has been reduced and reduced at each magnification, L evaluation values are selected in descending order of evaluation values obtained by the autocorrelation peak point evaluation means. The evaluation value obtained by the autocorrelation peak point evaluation means for each digital signal which is the two-dimensional image signal with the modification reduced at each magnification is added to the selected L evaluation values. Predetermined value p greater than any value ₀ Autocorrelation peak point evaluation means for adding
Evaluation values added by the autocorrelation peak point evaluation value increasing means obtained for each digital signal which is a two-dimensional image signal with the modification reduced at each magnification, and evaluation values not added, all Among the evaluation values of the embedded information detection upper limit number M stored in the embedded information detection upper limit number storage means in order from the largest evaluation value, and the peaks corresponding to the selected M evaluation values are selected. Embedded information detection target peak point candidate selection means for outputting point candidates as detection targets;
To determine peak point candidates for embedded information detection
The digital watermark detection apparatus according to claim 1. A digital signal that is a two-dimensional image signal is embedded in advance as a digital watermark so that the predetermined pattern is not repeatedly perceived by human perception. As a modification, enlargement or reduction, rotation, and aspect ratio change are performed. A digital watermark detection method in a digital watermark detection apparatus for estimating a conversion parameter representing the primary conversion added to the digital signal when a digital signal subjected to any primary conversion is input,
A signal reduction step of reducing the image size of the digital signal, which is a two-dimensional image signal to which the input modification is applied, by a predetermined magnification;
An autocorrelation function calculating step of calculating an autocorrelation value of a digital signal, which is a two-dimensional image signal reduced in the signal reduction step, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
In the autocorrelation function value array obtained in the autocorrelation function calculation step, an autocorrelation peak having a peak autocorrelation value is searched, and the position of the autocorrelation peak value stored in the autocorrelation function value array is searched. Near each position of the autocorrelation function value array within a predetermined range in the vicinity and the position where the theoretical autocorrelation peak value that appears in the autocorrelation function value array due to the periodicity of the pattern repetition is stored Autocorrelation peak point evaluation step for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within a predetermined range at
In accordance with the evaluation value obtained in the autocorrelation peak point evaluation step, a detection target determination step for determining autocorrelation peak point candidates to be used as detection targets for embedded information;
The position where the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined in the detection target determination step is stored in the autocorrelation function value array, and the autocorrelation function when the alteration is not made A conversion parameter calculation step for calculating a conversion parameter for performing the primary conversion applied to the digital signal from a deviation from a position where the autocorrelation peak appearing in the value array is stored in the autocorrelation function value array; ,
An electronic watermark detection method comprising: A digital signal that is a two-dimensional image signal is embedded in advance as a digital watermark so that the predetermined pattern is not repeatedly perceived by human perception. As a modification, enlargement or reduction, rotation, and aspect ratio change are performed. A digital watermark detection method in a digital watermark detection apparatus for estimating a conversion parameter representing the primary conversion added to the digital signal when a digital signal subjected to any primary conversion is input,
A signal reduction step of reducing the image size of the digital signal, which is a two-dimensional image signal to which the input modification is applied, by a predetermined magnification;
An autocorrelation function calculating step of calculating an autocorrelation value of a digital signal, which is a two-dimensional image signal reduced in the signal reduction step, and storing it in a two-dimensional autocorrelation function value array corresponding to the image signal;
In the autocorrelation function value array obtained in the autocorrelation function calculation step, an autocorrelation peak having a peak autocorrelation value is searched, and the position of the autocorrelation peak value stored in the autocorrelation function value array is searched. Near each position of the autocorrelation function value array within a predetermined range in the vicinity and the position where the theoretical autocorrelation peak value that appears in the autocorrelation function value array due to the periodicity of the pattern repetition is stored Autocorrelation peak point evaluation step for calculating an evaluation value based on a correlation value with each value of the autocorrelation function value array within a predetermined range at
In accordance with the evaluation value obtained in the autocorrelation peak point evaluation step, a detection target determination step for determining autocorrelation peak point candidates to be used as detection targets for embedded information;
The position where the autocorrelation peak value of the autocorrelation peak point candidate of the detection target determined in the detection target determination step is stored in the autocorrelation function value array, and the autocorrelation function when the alteration is not made A conversion parameter calculation step for calculating a conversion parameter for performing the primary conversion applied to the digital signal from a deviation from a position where the autocorrelation peak appearing in the value array is stored in the autocorrelation function value array; ,
The digital signal subjected to the modification is subjected to geometric correction by performing an inverse transformation on the conversion parameter, and the geometrically corrected digital signal is translated by a predetermined movement amount within a predetermined region, A translation amount estimation step in which digital watermark detection is attempted from the translated digital signal and the amount of movement when detected is added to the digital signal as a translation amount;
An electronic watermark detection method comprising: If the predetermined magnification is plural, each digital signal is a two-dimensional image signal applied with the modified reduced at each magnification, prior Symbol autocorrelation value obtained by the autocorrelation function calculating step The digital watermark detection method according to claim 6 , further comprising an autocorrelation function value normalizing step of normalizing the autocorrelation value so that the mean and variance or standard deviation of the same are equal . When the predetermined magnification is plural, all evaluation values obtained by the autocorrelation peak point evaluation unit for each digital signal which is the modified two-dimensional image signal reduced at each magnification The digital watermark detection method according to claim 6, further comprising an autocorrelation peak point evaluation value normalizing step of normalizing the evaluation value so that an average and a standard deviation are the same . In the detection target determining step, when the predetermined magnification is plural,
For each digital signal that is a two-dimensional image signal that has been modified and reduced at each magnification, L evaluation points are selected in descending order of evaluation values obtained in the autocorrelation peak point evaluation step. Any one of the evaluation values obtained in the autocorrelation peak point evaluation step for each digital signal which is a two-dimensional image signal in which the modification is applied to the selected L evaluation values and reduced by the respective magnifications. A predetermined value p greater than the value of ₀ Autocorrelation peak point evaluation value increasing step for adding
Evaluation values added and not added by the autocorrelation peak point evaluation value increasing step obtained for each digital signal which is a two-dimensional image signal with the modification reduced at each magnification, In order from the largest evaluation value, the embedding information detection upper limit count M stored evaluation values stored in the embedding information detection upper limit count storage means are selected, and the peaks corresponding to the selected M evaluation values are selected. Embedded information detection target peak point candidate selection step for outputting point candidates as detection targets;
The digital watermark detection method according to claim 6, wherein: A digital signal that is a two-dimensional image signal is embedded in advance as a digital watermark so that the predetermined pattern is not repeatedly perceived by human perception. As a modification, enlargement or reduction, rotation, and aspect ratio change are performed. When a digital signal subjected to any primary conversion is input, a digital watermark detection program for estimating a conversion parameter representing the primary conversion added to the digital signal,
  A digital watermark detection program for causing a computer to function as each unit of the digital watermark detection apparatus according to any one of claims 1 to 5.

Images