JP2018201134A - Digital watermarking apparatus and method - Google Patents

Abstract

To provide a device and a method capable of extracting with high reliability a strong and robust digital watermark with print resistance without visually perceived degradation of image quality.SOLUTION: Highly reliable watermark information can be obtained by separating an image into blocks, embedding a cluster dot pattern indicating a plurality of different green noise characteristics in which the spatial frequency is limited to a band in each block according to watermark information, performing a masking process on a spectral distribution of each block for extraction, and selecting highly reliable watermark information.SELECTED DRAWING: Figure 2

Description

The present invention relates to a digital watermark method for embedding copyright information and the like in digital image data, and more particularly to an invisible digital watermark apparatus and method having printing resistance.

In today's digital information society, it has become possible for many people to share information by copying information, and societies have developed greatly. However, due to its convenience, illegal copying and distribution of personal works has caused incidents such as copyright infringement. In recent years, due to the high image quality of digital cameras and printers in recent years, it has become possible to easily obtain copies that are not inconsistent with the original image. In addition to illegal copies that infringe copyright, counterfeits such as banknotes and securities. This is a result of promoting a malicious criminal act.

Under such circumstances, digital watermark technology for protecting copyright by embedding other information such as author information in image information has been developed. This digital watermark technology embeds the name, contact information, and handling information of the copyright holder in the copyrighted work, not only alerts the user, but also enables tracking of unauthorized use. It is starting to be widely used as a protection and security measure. In order to ensure security and reliability, the digital watermark technology is required to have a strong resistance to an attack assuming an attack for deletion from a third party.

A digital watermark as digital data does not deteriorate or disappear as long as it is copied as it is. However, when image editing or processing by Affine transformation such as enlargement, reduction, rotation, etc., or when image compression such as JPEG is performed, various calculations and transformations are performed on the image data, so the embedded information Is easy to disappear. Therefore, in order to withstand the editing process, the strength of the embedded data must be increased and embedded in the deep layer to improve the durability.

Further, when the image is output to the printer, various image processing for printing is performed in addition to the image editing operation. For example, YMC conversion for subtractive color mixing, black plate generation processing for determining the black ink amount, halftoning processing for binarizing and reproducing gradation, and the like are performed. In addition, the embedded watermark information is more easily lost due to engine characteristics such as the printer's spatial frequency response (MTF), gradation reproduction, print dot scattering, and dot gain. Therefore, digital watermarking for printing requires a more robust and robust technique in preparation for these situations.

Until now, many methods have been proposed as digital watermark technology having printing resistance to image data. One example is the patchwork method. Information embedding by the patchwork method is an approach that uses statistics, and thus shows high resistance to local attacks.
This method gives a slight deviation to a large number of pixel pairs in real space (pixel space) and embeds them as watermark information. First, a pixel pair is selected at random, and one luminance value is increased by δ and the other luminance value is decreased by δ. By repeating this operation, the expected value of the distribution is biased. The watermark information can be extracted by statistically obtaining this bias. However, since this technique causes a statistical bias, the amount of information that can be embedded is small. Furthermore, in order to satisfy the printing durability, it is necessary to increase the embedded data value δ, and since it is embedded in the real space (pixel space), the embedded region (patch) becomes visually noticeable, and the luminance value In this case, there is a problem that it is vulnerable to density unevenness and gradation fluctuation in printing and reading.

Another method is to embed watermark information in a frequency space. In this method, the entire image is subjected to DCT (Discrete cosine transformation), Fourier transform, and Wavelet transform and embedded in a specific intermediate frequency band. Since the embedded information is distributed over the entire image in real space, it has the feature of being visually inconspicuous. The embedded frequency band uses an intermediate frequency band that has little visual impact and can be read by a scanner. However, in order to increase the printing durability so that it can be read by a scanner, it is necessary to increase the embedding strength, and a pattern corresponding to the embedding frequency is generated. Since this pattern is periodic, moiré fringes with the image are generated, or the flat portion of the image is easily noticeable, resulting in a reduction in image quality. In addition, in the method of passing embedded position information as a key, the algorithm is easily clarified by collusion attack or reverse analysis.

In addition, these watermark technologies are often used as security technologies such as forgery prevention and tracking of banknotes and securities, and guarantee of originals. Once the algorithm is released, various attacks to remove or avoid the watermark Can be considered. For this reason, the algorithm is usually not disclosed, and there is a problem that standardization and dissemination are prevented.

W. Blender, D. Gruhl, N. Morimoto; “Techniques for data hiding”, Proceedings of SPIE, Vol. 2020, pp. 2420-2440 (1995). Minoru Mizumoto, Koko Matsui; “Examination of digital watermark printing tolerance using DCT”, IEICE Journal A, Vol J85-A, No.4, pp.451-459 (2002)

For this reason, it is a problem to obtain a secure digital watermarking method which has printing resistance and has little moire and image quality deterioration even when embedded strongly. In other words, it is necessary to have resistance to editing processing, image compression, printer engine characteristics, etc., and resistance to various attacks that attempt to remove watermark information from the outside. The problem is to improve the image quality while increasing the tolerance. In general, there is a trade-off between increasing tolerance and improving image quality, but image quality can be satisfied by using human visual characteristics and making it visually inconspicuous.

Furthermore, it is a problem that a watermark can be removed using a key as necessary. Another challenge is to create a system that does not spread damage to a wide range of lost keys. In addition, it is a problem to satisfy the fact that safety is maintained even if an algorithm is disclosed in order to spread.

Further, in printing, there are many examples in which an image is edited and cut out and used, which makes it difficult to extract a watermark. In order to perform watermark extraction with high accuracy, it is possible to improve the extraction accuracy by obtaining at least information on how the original image has been edited. That is, it is also a problem to be able to specify the size and direction of the original image from the embedded image data.

In order to solve the above-described problems, the present invention realizes a digital watermark with high image quality but printing durability by superimposing and embedding a random cluster dot pattern with band limitation in a frequency space on an original image. An apparatus and method are provided.

Random cluster dots used for embedding show spectral characteristics with reduced intensity at high and low frequencies. Since it is random, there is no moiré with the original image. The printer engine is characterized by high printing response characteristics and is difficult to recognize from human visual characteristics. For this reason, even if the reproducibility of watermark information is high in printing, the watermark information does not matter when the printed image is observed at a clear distance.

In addition, it has a high resistance against external attacks and has a key that can remove a watermark if necessary. The key includes a dot pattern of watermark information to be embedded, and such a pattern is composed of a large number of random dots, so that it is difficult to perform reverse analysis. Further, by changing the pattern configuration for each image data, a single image is assumed. Even if the watermark information is decoded, it is almost impossible to remove the watermark of other images with this information.

In order to solve this problem, the present invention embeds an original image using a dot pattern exhibiting a cluster type green noise characteristic. When this green noise pattern is viewed in the spatial frequency space, there is no spectrum near 0 frequency. On the other hand, since the image data is distributed in the vicinity of 0 frequency, there is little overlap between the spectra. For this reason, when extracting the embedded information by separating each spectrum, it is possible to extract well. In addition, the cluster dot pattern exhibiting green noise characteristics is dispersed and has low granularity, and when viewed at a clear distance, it is visually inconspicuous, so that the high quality of the original image can be maintained. For this reason, even if the embedding strength (gain) is increased, it is difficult to visually perceive image quality deterioration, and printing resistance is enhanced.

The watermark information embedding method divides digital image data into blocks of R pixels x R pixels, and the Fourier spectrum has a low frequency range corresponding to the bit information (0 or 1) of the watermark information embedded in each block. Also, watermark embedding is performed using a plurality of different dot patterns exhibiting green noise characteristics reduced in a high frequency range. In the method of extracting watermark information, received or read image data is corrected and then divided into blocks, the Fourier spectrum distribution of each block is obtained, and information embedded by mask processing is obtained.

Further, when extracting watermark information, a parameter called “reliability” of extracted data is introduced and quantified, and watermark information with high reliability can be extracted by selecting watermark information with high reliability.

According to the present invention, robust watermark embedding with high image quality and printing durability can be performed, and highly reliable watermark extraction can be performed from a printed image. In printing, stable output is possible because the main frequency of cluster dots is highly reproducible in the printer engine. In addition, cluster dots can detect the Affine transform coefficient (enlargement / reduction ratio and rotation angle) applied in editing from this dot pattern, and can perform highly accurate watermark extraction by obtaining a corrected image. .

Furthermore, it is possible to return to the image (original image) before embedding by obtaining a key for the image in which the watermark is embedded. By changing the key for each distribution image, it is powerless to other images even if it is decrypted or stolen. It is also possible to embed new secondary copyright information as a digital watermark using the key.

1 is a diagram of an electronic watermark embedding and extracting apparatus according to the present invention. FIG. The figure which showed the delivery on the internet of the image with a watermark. This figure shows the mask pattern used for embedding and its spectrum. (A) is the one with the spectral characteristics of a vertically long elliptical ring, (b) is rotated 90 degrees, and (c) is for Affine coefficient detection. , (D) is for header detection. Figure (a) is an embedded image, (b) is a partially enlarged view, and (c) is an extracted spectral image. An image embedded with different gains and a partially enlarged view. (a) gain = 0.0625, (b) 0.125, (c) 0.25 The figure showing the processing flow of watermark embedding. Diagram showing gain automatic adjustment processing flow when embedding watermark The figure showing the flow which acquires the correction image at the time of watermark extraction. Diagram showing derivation of Affine transform coefficient The figure showing the processing flow of watermark extraction. The figure of the mask pattern of watermark extraction. The figure which shows the flow of the mask process of watermark extraction. The figure which shows a reliability process flow in the mask process of watermark extraction. The figure which shows the mask process by 4 divisions by watermark extraction. The figure which showed the image of the presence or absence of edge processing, and its spectrum. The figure which showed the different dot pattern by the same spectrum.

This digital watermark is embedded in an image whose copyright is to be protected, and tracks and protects the copyright in image data (electronic data) and printed matter. Copyright holders embed and publish watermarks in images that they want to claim their copyrights over the Internet, and because of the watermarks, it is possible to track unauthorized use later. On the other hand, the purchaser can perform watermark cancellation and rewriting by using a key, and can also embed secondary author information.

Illegal copying and usage can be monitored by watermark extraction software. This watermark extraction software has no secrecy, and can be freely used by anyone who can make it freely downloadable from the homepage of the copyright holder, for example. This makes it possible for a third party to know the owner of the image, copyright information, contact information, and so on, which also serves as a warning of unauthorized use.

In order to implement such digital watermarking, the present invention uses a plurality of different dot patterns for embedding in the original image. All of these dot patterns exhibit green noise characteristics, and an embedding pattern is selected according to 0 and 1 of bit information to be embedded and header information (the beginning of a character string). Moreover, the pattern which analyzes the coefficient of Affine transformation is included everywhere.

Such a dot pattern is a green noise pattern obtained by repetitive calculation from a random dot pattern generated from a pseudo-random number. Using these patterns, for example, a character string such as a copyright holder's name, e-mail address, or URL is converted into bit information and embedded in image data. It is also used as a “key” for removing embedded information and returning to the original image. Such a key is composed of dot pattern information, embedding strength (gain) information, and embedded character information. Such key information becomes a different dot pattern by changing the initial value of the random number of the generated dot pattern. For this reason, the key information is different for each dot pattern.

Hereinafter, it demonstrates in detail along an Example.
FIG. 1 is a block diagram of a digital watermark system for embedding and extracting information according to the present invention. The copyright holder's computer 3 stores copyrighted image data in a data memory 7 such as a hard disk. The image data is subjected to image processing by the image processing program in the program memory 6 using the CPU 11, ROM 4, RAM 5, etc. and displayed on the monitor 8. A scanner 1 and a printer 2 are connected to the computer 3, and processed images are output from the printer, and images can be read from the scanner 1. Since such image processing is heavily loaded, there may be a processing board for speeding up the GPU or the like.

FIG. 2 shows an Internet distribution processing procedure. The image in which the watermark information is embedded is distributed over the Internet 12 from the computer of the copyright holder 16 (13). The purchase applicant 17 who sees it notifies the distributor of the purchase request (14). After a predetermined procedure, the copyright holder distributes a secret key for embedding and removing the watermark (15). Based on the contract with the copyright owner, the purchaser removes the digital watermark embedded in the sent image and edits the image.

Further, the purchaser can embed the edited image data as the secondary author using his / her secret information and information such as his / her copyright information and URL. Images embedded with such watermark information are distributed to the market and used in various ways.

To monitor illegal duplication and usage, watermark information can be extracted from an image by watermark extraction software, and it can be confirmed that it is a copyrighted work. This watermark extraction software has no secrecy, and can be freely used by anyone who can make it freely downloadable from the homepage of the copyright holder, for example. This makes it possible for a third party to know the owner of the image, copyright information, contact information, and so on, which also serves as a warning of unauthorized use.

The secret key corresponds to each embedded image on a one-to-one basis. For this reason, the copyright holder uses a different key for each image so as not to be abused. For this reason, the watermark of other images cannot be removed using this key. It is also resistant to collusion attacks as described below. At this time, the watermark reading software can be used in common even if the keys are different, as described above.

First, the dot profile creation algorithm showing the green noise characteristics used in the present invention will be described.
Assume that the dither matrix size to be calculated is R × R (R = 2 ^ m; where ^ represents a power), and the blackening rate representing gradation is g (0 ≦ g ≦ 1: g = 1 is all black, = 0 is all white). A dot profile with an intermediate density (g = 1/2) is determined as follows, with the blackening rate g and the dot profile at point (i, j) as p (i, j).
(1) First, R ^ 2/2 random dots (white noise in the initial state) are generated by a pseudo-random number generator and set to p (i, j). At this time, the dot profile in the initial state can be changed by changing the SEED value of the pseudo random number generator. A two-dimensional Fourier transform of the dot profile is performed to obtain P (u, v).
(2) Multiply P (u, v) by filter D (u, v) to obtain a new P ′ (u, v). Here, D (u, v) is a filter having a value in the region where the radial frequency fr is fmin to fmax.
(3) An inverse Fourier transform is performed on P ′ (u, v) to obtain a multivalued point profile p ′ (i, j).
(4) An error function e (i, j) = p ′ (i, j) −p (i, j) is obtained, and white and black are inverted in descending order of error at each pixel position.
(5) The above operation is repeated until the error is within the allowable amount, and finally a dot profile of g = 1/2 is obtained.

Here, the setting of the radial frequencies fmax and fmin will be described. The frequency due to the average dot spacing at g = 1/2 at fmax and fmin of filter D (u, v) is
fo = √g · fn = √ (1/2) · fn
Where fmax and fmin are based on fo
a≡ (fmax-fo) / fn
b≡ (fmin-fo) / fn
The parameters (a, b) are defined as follows. Here, fn represents the Nyquist frequency. By changing (a, b), dot patterns with different cluster sizes can be obtained.

As an example, a dot profile is obtained when R = 64, (a, b) = (-1/16, -5 / 16), and ellipticity 1.3. Such a dot profile p0 (i, j) is shown in FIG. An ellipticity of 1.3 means that the filter D (u, v) is enlarged 1.3 times in the y-axis direction. Since the spectrum distribution is targeted for the x-axis and y-axis, the imaginary part of p0 (i, j) is 0. The dot pattern p1 (i, j) in FIG. 9B is obtained by rotating p0 (i, j) by 90 °, and the spectrum is similarly rotated by 90 °. (C) and (d) of FIG. 5 will be described later with an affinity coefficient detection pattern and a header pattern, respectively.

Next, a watermark embedding algorithm for multi-value image data will be described.
In embedding watermark information, color image data is converted into Y, Cb, and Cr, and watermark information is embedded in luminance Y. Ideally embedded in Blue (yellow when printing). However, accurate color separation is required. The following description is based on embedding in a block of 64 pixels × 64 pixels. However, when a large amount of information is to be embedded, the block size is 32 pixels × 32 pixels.

Now, suppose that different spectral patterns are embedded in the frequency space in units of blocks and Pi (u, v) (i = 0, 1, 2,...). The spectrum of Pi shows the aforementioned green noise characteristics. Ideally, fmin is selected to be larger than the maximum frequency of the embedded image, but is not necessarily limited thereto.
If the frequency spectrum of the image block is I (u, v), the embedded spectrum W (u, v) is
W (u, v) = I (u, v) + gain ・ Pi (u, v) ---------- (1)
It becomes. Here, in order to become invisible, gain << 1 must be satisfied. In order to extract the watermark information Pi (u, v) with high accuracy, it is desirable to reduce the overlap of I (u, v) and Pi (u, v).

The actual watermark information embedding work is performed in real space. Converting the previous expression into real space (pixel space) (represented in lower case)
w (x, y) = i (x, y) + gain ・ pi (x, y) ---------- (2)
It becomes. For the watermark information, a dot pattern pi (x, y) (i = 1,2) indicating such different green noise characteristics is prepared, and for the embedded bit information (0,1),
When embedded bit = 0 → p0 (x, y)
When embedded bit = 1 → p1 (x, y)
Embed as Here, although p0 and p1 are binary values of (1,0), they are assumed to be (1/2, -1/2) in order to preserve the average luminance.

Moreover, since p0 and p1 are random dots, the boundary between them is not noticeable. In addition, cluster type halftone screens exhibiting green noise characteristics are often used as cluster type FM screens during printing, and are uniform in terms of human vision due to their excellent uniformity with dispersive dots. I don't feel it.

(C) and (d) in FIG. 3 are obtained by obtaining a green noise pattern using a rectangular filter. (C) is a generation pattern from a rectangular filter of (a, b) = (0, -1 / 8). That is, it is a dot pattern generated from a rectangular band filter of fmin = −1 / 8 · fn, fmax = 0 in the x and y directions. (D) is a dot pattern generated from a rectangular band filter of (a, b) = (1/8, -1 / 8). The pattern (c) is used for searching for an Affine transform coefficient, which will be described later, and the pattern (d) is used at the beginning of a character string, and is used as a repeated delimiter pattern.

The amount of data (maximum number of bits) N that can be embedded in the original image is calculated for an image with an image size of W pixels x H pixels.
N = int (W / R) × int (H / R) ---------- (3)
Given in. Here, int () represents rounding down to the nearest whole number. The number of characters that can be embedded is int (N / 8) for ASCII characters. Usually, there are many target images for copyright protection that are HD size (1980 pixels x 1024 pixels) or more. In HD images, N = 480 bits = 60 bytes, which is sufficient for embedding copyright information. However, if a small size image requires a sufficient amount of information, if R = 32 is embedded, for example, if 512 pixels × 512 pixels, an information amount of N = 256 bits = 32 bytes can be embedded. is there.

FIG. 4 shows an image of R = 64 embedded in an image of 1024 pixels × 576 pixels. Therefore, this image is
N = int (1024/64) x int (576/64) = 16 x 9 = 144 bits = 18 bytes, and character information of up to 18 bytes (in the case of ASCII) can be embedded. The figure (a) shows an embedded image, and the figure (b) shows a partial enlarged view. The embedding gain is 0.0625. In other words, the dot pattern of ± 8 is synthesized with the original image (as 8 bits / pixel). In FIG. 5C, the correct answer is obtained by the extraction result described later.

For embedded images, printing gain can be increased by increasing the gain. However, increasing gain makes the dot pattern visible. FIG. 5 shows an image when embedding with changing gain. When gain = 0.0625 in the figure (a), the PSNR (Peak Signal-to-Noise Ratio) is 30.1, and the embedding is hardly visible visually, but when gain = 0.25 in the figure (c), the watermark The extraction is almost 100%, but PSNR = 18.06, and the dot structure becomes noticeable. As a result of the experiment, it was found that if gain is 0.125 or more, it has printing resistance. In this case, PSNR = 24.08, but it is not noticeable due to the low-pass effect of visual characteristics.

FIG. 6 shows a watermark embedding process flow. First, a watermark information bit string wm (i, j) to be embedded is prepared (20). Here, i represents the i-th character to be embedded, and j represents the j-th bit (MSB is the 0th bit). Since ASCII characters are 8 bits / character, j = 0,1, ... 7. I is 1 ≦ i ≦ int (N / 8). Here, N is the maximum number of embedded bits shown in Equation 3. Subsequently, the image data is divided into R × R blocks (21), and a dot pattern is synthesized in units of blocks. Whether or not it is the head of the character string (22). Embed (pHeadder). Further, it is determined whether or not the MSB (Most Significant Bit) is a character (23). In the case of the MSB, an Affine detection pattern (pAffine) described later is embedded. Otherwise, the pattern of p0 is embedded when the embedded bit is 1, and the pattern of p1 is embedded when the embedded bit is 0 (24). The process ends when all blocks have been embedded.

The embedded watermark information may not be extracted because the embedded gain is small. Such a case occurs when the spatial frequency of the corresponding block of the original image is high. FIG. 7 shows a flow for performing automatic gain adjustment. First, watermark information wm (i, j) is embedded in all blocks as gain = g (30). After all blocks are embedded, the watermark is read (32). Here, the extracted watermark information wm ′ (i, j) is compared with the watermark information wm (i, j) before embedding to determine whether or not they match. At the same time, it is determined whether the reliability wmrel ′ (i, j) of the extracted bit, which will be described later, is greater than a certain threshold (33). If the watermark information does not match, or if the watermark information matches but the reliability is equal to or less than the threshold, the watermark is embedded again by adding a constant value Δg to gain (34). Since the reliability can be quantified, embedding with the optimum watermark strength can be automatically performed as described above. Note that if the reliability is equal to or less than the threshold even if the watermark information matches, the print image may not be read.

In the extraction of watermark information, the spectrum image is obtained in units of blocks after correcting the size and inclination of the read image. Fourier transform of equation (2)
F {w (x, y)} = F {i (x, y)} + gain ・ F {pi (x, y)} ---------- (4)
It becomes. However, F {} represents a Fourier transform. Usually, F {i (x, y)} is the spectrum of the original image and is localized near the 0 frequency, and F {pi (x, y)} is a ring-shaped spectrum surrounding it. For this reason, there is little overlap of both and it is easy to isolate | separate both.

FIG. 8 shows the processing flow of the corrected image. First, the printed image is read with a scanner or camera (35). Subsequently, the target image is cut out (36), an R × R window is scanned, and the corresponding portion is subjected to FFT to search for a rectangular pattern (37). Since the printed image is reduced or enlarged, in some cases, the window size may be R / 2 ^ n (where n is an integer). If a rectangular pattern is found, an Affine coefficient can be obtained (38). Thereafter, inverse Affine transformation is performed to restore the original image size and direction (39). Subsequently, edge enhancement is applied to the image (40) to obtain a corrected image.

The Affine coefficient is obtained as follows. Now, assuming that the embedded image w (x, y) is enlarged a times, the Fourier transform is P (fx, fy), where the Fourier spectrum before scaling is P (fx, fy)
F {w (x / a, y / a)} = F {i (x / a, y / a)} + gain ・ F {p (x / a, y / a)}
= F {i (x / a, y / a)} + gain ・ | a | ・ P (au, av)
And is reduced to 1 / a in the spectral space. Further, an image obtained by rotating the image by θ is also rotated by θ in the Fourier spectrum. Therefore, the Affine coefficient applied by measuring the spectrum distribution of the embedded image can be obtained. The detection window size may be small. For example, when extraction is attempted with a window size of 40 × 40 without considering registration, detection is possible even if a margin is provided and 64 × 64 is applied to perform FFT (Fast Fourier Transform).

FIG. 9 specifically shows the detection of the Affine conversion coefficient. The print image is obtained by enlarging the original image by 1.25 times and rotating it clockwise by 8 degrees. Scanning the extraction window, find a rectangular pattern in the spectrum and adjust the position and margins so that it is most clear. It is possible to obtain an Affine coefficient by comparing the size of the obtained rectangular pattern with the size when embedded in the original image. The rotation can be obtained from the rotation angle of the obtained rectangular pattern.

FIG. 10 shows a processing flow of watermark extraction. First, the corrected image is divided into R × R blocks (41), and FFT is performed on the first block (42). Subsequently, histogram equalization is applied to the obtained spectral image (43). This is for improving the contrast of the spectral pattern and at the same time standardizing by quantifying the reliability described later. Subsequently, a mask process described later is performed (44) to obtain watermark information wm '(i, j) and reliability wmrel' (i, j). It is determined whether or not all the blocks are finished (46). If not finished, the process goes to the next block and the same processing is performed again.

FIG. 11 shows a mask pattern as a discriminator that extracts watermark information by pattern recognition from the spectrum distribution after Fourier transform. A vertically long and horizontally long elliptical ring-shaped mask, which corresponds to the spectral characteristics of the dot pattern for embedding. In the figure, the black region has a value of 1, and the white portion has a value of 0. This mask is overlaid on the spectrum with the watermark, and it is determined whether the bit is 0 or 1 from the difference in the integrated luminance value of the overlapping portion. That is, in FIG. 12, outputs Q0 and Q1 having the following integrated luminance values are obtained for the spectrum pattern W (i, j) of the block (40).
Q0 = M0 ◎ W = 1 / ZΣM0 (i, j) ・ W (i, j)
Q1 = M1 ◎ W = 1 / ZΣM1 (i, j) ・ W (i, j) ---------- (5)
Here, ◎ represents a mask operation, and Z represents the number of pixels having a mask value of 1. From this output,
When Q0> Q1, extraction bit = wm '(i, j) = 0
When Q1> Q0, extraction bit = wm '(i, j) = 1
(41). At the same time, the reliability is obtained as follows (42).
wmrel '(i, j) = | Q0-Q1 |
The higher the reliability, the higher the reliability of the extracted bits.

FIG. 13 illustrates an improvement in the accuracy and reliability of watermark information. The maximum number of embedded bits N is given from the equation (3) depending on the size of the image. If the number of bits to be embedded is n and n <N, the bit string can be embedded in duplicate. In the extraction of the watermark information in FIG. 10, continuous watermark information wm ′ and reliability wmrel ′ up to N are obtained regardless of the number of bits (without considering the repetition of the character string). From the continuous character string, a highly reliable one is selected to obtain watermark information up to n.

First, when i is the i-th character number (0 <i ≦ int (N / 8)) and j is the bit number (0 ≦ j ≦ 7), the remainder k is modulo n. , K = i mod n (50), and the (i, j) th reliability wmrel ′ is
When wmrel '(i, j)> wmrel (k, j), the extraction reliability is high.
wm (k, j) = wm '(i, j)
wmrel (k, j) = wmrel '(i, j) ---------- (6)
(52), the watermark information is replaced. This is performed for all bits and all blocks, that is, until the number of embedded bits N is completed.

Through the above operation, the watermark information wm2 (i, j) and wmrel2 (i, j) with the highest reliability of the embedded bit string is selected. Thereby, the extracted value for each bit and its reliability are quantified. From this value, the reliability wmrel2 (i) of the extracted character unit is calculated from the average of the two-level reliability of each bit.
Here, Σ is the sum of j = 0-7. The reliability of the characters extracted in this way can be quantified.

However, when high-frequency components exist in the original image data, it may be extracted with low reliability. FIG. 14 shows an example. Figure (a) shows an RxR block image with a uniform sky at the top and a treetop at the bottom. As a result of extraction from this image, as shown in Figure (b) Therefore, the reliability is low and extraction is difficult. This is because the frequency of the image exists in the frequency band of the green noise spectrum. In such a case, when the four sides are divided into R / 2 rectangles and a Fourier spectrum is obtained for each rectangle, an image with no spatially high frequency appears as shown in Fig. 3 (c), and a clearly oblong elliptical spectrum appears. It is confirmed that

Extraction with high reliability is possible by subdividing and extracting an embedded block with low reliability in this way. All four divided blocks exhibit the same spectral characteristics, so if even one high reliability is extracted, the entire block is upgraded to that reliability. In this way, highly reliable watermark extraction can be performed on a printed image.

The edge enhancement process in the correction process makes the spectrum distribution clear as shown in FIG. 15 and improves the recognition accuracy. It is desirable to apply this edge enhancement process strongly with a 3 × 3 Laplacian filter. Usually, a printer has a strong low-pass action, so the printed image has a lower high frequency range. For this reason, the watermark extraction reliability is improved by emphasizing the high frequency range.

The watermark is extracted as described above, and FIG. 4C described above shows the spectrum distribution at the time of extraction. Embedding is an ASCII character with 8 characters “kawamura” embedded. Each character is expanded into bits, and the pattern of p0 is embedded for “0” and p1 is embedded for “1” as described in FIG. Since the first bit (MSB) of the ASCII character is all 0, it is embedded with the pattern for detecting the Affine conversion coefficient, and the first bit of the first character of the character string is embedded with the header pattern for character string delimiter. After the corrected image is created, first, a character string is extracted starting from the header pattern. Since an Affine detection pattern appears every 8 bits, characters are extracted using this pattern as a synchronization signal. When the header pattern is extracted again, it is known after the end of the character string, the second extraction is performed, and the one with high reliability is selected. In Fig. 3 (c), the character string of 8 characters is embedded twice and twice, and the accuracy of the extraction result can be improved by selecting a highly reliable one from the duplicated embedded information. , The reliability of each extracted character can be quantified.

The watermark information can be removed and returned to the original image by knowing the embedding pattern. That is, from equation (2)
i (x, y) = w (x, y)-gain ・ pi (x, y)
The embedding pattern pi (x, y), gain, and the embedded character string can be received as a “key” to return to the original image. In order to restore the original image completely, it is necessary to limit the dynamic range in advance in order to avoid overflow and underflow of the embedded image.
The key size is 4096 bits (512 bytes) when R = 64, and 1024 bits (128 bytes) when R = 32.

Next, attack resistance of this method will be described. Examples of attacks on digital watermarks include signal enhancement (sharpness adjustment), noise addition, filtering (linear and nonlinear), lossy compression (JPEG, MPEG), and transformation (rotation and scaling). These attacks are attacks on luminance information, and embedding with a set of dispersed dots as in this method can be maintained unless there is an attack in the resolution direction (for example, a strong low-pass filter), and is generally robust. I can say that. Furthermore, tolerance can be increased by increasing gain.

In addition, a collusion attack in which a watermark pattern is specified from an original image in which a watermark is not embedded or a plurality of embedded images can be avoided by changing the embedded pattern for each image. Since there are a large number of patterns having the same spectral characteristics, it is not possible to remove watermark data embedded with other images with the same key.

FIG. 16 shows a different cluster dot pattern p0 generated by changing the initial value (SEED value of the random number) at the time of creating the dot pattern. The spectral distribution is the same and the extraction software is common, but the dot profiles are different. When R = 64, the number of different patterns can be (4096) C (2048) only (C represents a combination), and the probability of the same pattern is very low, so the safety is high. Since the spectrum distribution is the same, there is no need to change the watermark extraction software. Therefore, collusion attacks can be avoided by changing the dot pattern profile for each image. The copyright holder provides the purchaser with a dot profile obtained from different random numbers as a “secret key”. Since this key is different for each purchaser or image, the watermark cannot be removed from the images of other purchasers.
A normal digital watermark is easy to attack if the algorithm is known. Therefore, in principle, the watermark algorithm is not disclosed. On the other hand, even if this algorithm is made public, the key for removing the watermark is, in principle, one for one image. Even if the key information can be decrypted, the key is used for other images. The watermark cannot be removed.

Furthermore, the safety is further enhanced by creating a new dot pattern p1 rotated by 90 ° from different random number patterns. In this case, two key patterns are required, and the watermark can be removed only when both keys are available.

Also, the watermark can be used without being removed. When the embedded image is observed at a normal distance at a clear distance, the pattern is hardly recognized from the MTF characteristics of the visual system. The digital watermarking method of the present invention is an analogy of spread spectrum in real space, and printing resistance can be maintained even if gain is reduced.

In addition, since information is extracted from the spatial distribution of dots, it is less susceptible to density variations and illumination variations. An extraction error occurs when the original image itself has a spatial frequency close to the frequency of the cluster dots. In this case, the extraction accuracy is improved by bringing the main frequency of the cluster dots to the high frequency side, but high response characteristics in the high frequency range of the printer are required, and the restriction on the paper quality of the printing paper becomes severe. If there is a margin in the number of bits to be embedded, one solution is to embed watermark information repeatedly and select bits with high extraction reliability to improve accuracy. The amount of information to be embedded is reduced, but the embedded block size should be 32 x 32 pixels.

The embedding of the watermark can be performed at a high speed in the same manner as the binarization by the dither method by embedding in units of blocks. Extraction is somewhat complicated and computationally expensive, but because it is independent in units of blocks, parallel processing reduces the computation time in inverse proportion to the number of parallel operations, so it is effective for large images. .

As described above, the digital watermarking method of the present invention has been described. Since this method embeds a dot pattern showing green noise characteristics in real space, there is little visual discomfort and it can be embedded strongly. Since it is not easily affected by distortion and can be restored to the original image by key information, it can be used not only for copyright protection but also for various applications by linking the printed image and electronic data. It is valid.
The watermark algorithm and extraction software may be disclosed. A general user can confirm copyright information, conditions, and the like of an image on the network using extraction software. The embedded watermark information cannot be released unless there is a unique key.

1 is scanner, 2 is printer, 3 is computer system, 4 is ROM, 5 is RAM, 6 is program memory, 7 is data memory, 8 is monitor, 9 is keyboard, 10 is communication function, 11 is CPU, 12 is Internet, 13 distribution of image data, 14 contact for purchase request, 15 send private key,
Reference numeral 16 denotes a copyright owner's computer, and 17 denotes a purchaser's computer.

Claims (5)

In the digital watermarking method in which n-bit information including copyright information is embedded in digital image data i (x, y),
To embed watermark information, digital image data is divided into N blocks of R pixels x R pixels (N ≥ n), and bit information (0 or 1) of the watermark information embedded in each block corresponds to , Using multiple different dot patterns pi (x, y) (i = 0,1,2,...) Showing green noise characteristics with reduced Fourier spectrum in the low and high frequency regions, the watermark embedded image w (x , y) = i (x, y) + gain-pi (x, y) is duplicated (where gain is the embedding strength, gain << 1),
In the extraction of watermark information, the received or read image data is divided into blocks, and the embedded bit information and the reliability of the bit information are obtained by masking the Fourier spectrum distribution of each block and the extraction mask. An electronic watermarking system and method characterized in that watermark information is obtained by selecting the most reliable bit information. The dot pattern pi (x, y) indicating the green noise characteristic has a point profile at an intermediate value (g = 1/2) of the blackening rate g (0,1), a lower limit value fmin of the spatial frequency, and an upper limit. 2. The electron according to claim 1, which is obtained by filtering with a band-pass filter D (u, v) having a finite value between 0 and fmax (where u and v are two-dimensional spatial frequencies). Watermark system. The digital watermark extraction mask includes masks M0 and M1 corresponding to the spectral characteristics P0 and P1 of the embedded dot pattern, and the mask processing is performed by integrating luminance values Q0 and Q1 of the mask and the extracted spectrum distribution of the block. ,
When Q0> Q1, extraction bit = 0
Extraction bit = 1 when Q1> Q0
And the reliability is reliability = | Q0-Q1 |
3. The digital watermark system and method according to claim 2, wherein watermark information is extracted by selecting a highly reliable one from watermark information embedded in duplicate. In embedding the watermark information, the watermark information is first embedded in the image with a predetermined gain value, and then the embedded watermark information is extracted. If all bits are not extracted correctly, or the reliability of each bit is below a certain threshold value. In this case, the gain is increased and embedding is performed again, and the above operation is automatically performed until all bits are correctly extracted and the reliability of each bit is equal to or higher than a certain threshold. 4. The digital watermark system and method according to 3. 4. The digital watermark system and method according to claim 3, wherein in extracting watermark information, the embedded character and the extraction reliability of the character are simultaneously displayed, and the authenticity of the extraction result is visualized.