JP6937007B2 - Digital watermarking device and method - Google Patents

Description

The present invention relates to a digital watermarking method for embedding copyright information or the like in digital image data, and more particularly to an invisible digital watermarking device and method having print resistance.

In the current digital information society, the duplication of information has made it possible for many people to share information, and the society has greatly developed. However, due to its convenience, cases such as copyright infringement have come to occur due to the illegal copying and distribution of personal works. With the recent improvement in the image quality of digital cameras and printers, it has become easier to obtain duplicate images that are exactly the same as the original images, and not only infringing copies but also forgery of banknotes and securities. This has resulted in the promotion of malicious criminal acts called acts.

Under such circumstances, digital watermarking technology that protects copyright by embedding other information, such as author information, in image information has been developed. This digital watermarking technology not only draws the attention of the user by embedding the name, contact information, handling items, etc. of the copyright holder in the copyrighted work, but also can track unauthorized use, etc. It is beginning to be widely used as a protection and security measure. Digital watermarking technology is required to have strong resistance to attacks in order to ensure safety and reliability, assuming an attack for deletion from a third party.

Digital watermarks as digital data do not deteriorate or disappear as long as they are duplicated as they are. However, when image editing, processing, or image compression is performed, the embedded information is likely to be lost because various operations and conversions are performed on the image data. Therefore, in order to withstand editing, it is necessary to increase the strength of the embedded data and embed it in a deep layer to improve the resistance.

Further, when the image is output to a printer, various image processing for printing is performed in addition to the operation of editing and processing the image. Further, at the time of printing, the embedded watermark information is more likely to be lost due to the engine characteristics such as the spatial frequency response characteristic (MTF) of the printer, the gradation reproduction characteristic, the scattering of printed dots, and the dot gain. Therefore, digital watermarking for printing requires more robust and robust technology to prepare for these situations.

So far, many methods have been proposed as digital watermarking techniques that are print-resistant to image data. One example is the patchwork method. Since information embedding by the patchwork method is an approach that uses statistics, it shows strong resistance to local attacks.
This method gives a slight deviation to a large number of pixel pairs in the real space (pixel space) and embeds them as watermark information. First, a pair of pixels is randomly selected, one brightness value is increased by δ, and the other brightness value is decreased by δ. By repeating this operation, the expected value of the distribution becomes biased. Watermark information can be extracted by statistically obtaining this bias.
However, since this method causes statistical bias, the amount of information that can be embedded is small. Furthermore, in order to satisfy the print resistance, it is necessary to increase the embedded data value δ, and since it is embedded in the real space (pixel space), the embedded area (patch) becomes visually noticeable and the brightness value. There is a problem that it is vulnerable to density unevenness, gradation fluctuation, etc. in printing and reading because it directly fluctuates.

Another method is to embed watermark information in the frequency space. This is a method of embedding the entire image in a specific intermediate frequency band by performing DCT (Discrete cosine transform), Fourier transform, and Wavelet transform. Since the embedded information is dispersed throughout the image in the real space, it is not visually noticeable. As the embedded frequency band, an intermediate frequency band that has little visual influence and can be read by a scanner is used. However, in order to increase the print resistance so that it can be read by a scanner, it is necessary to increase the embedding strength, and a pattern of embedding frequency is generated. Since this pattern is periodic, moire fringes with the image may occur, or the flat portion of the image may be easily noticeable, resulting in deterioration of image quality.

In addition, these watermarking technologies are often used as security technologies such as anti-counterfeiting and tracking of banknotes and securities, and guarantee of originals, and algorithms are usually not disclosed. Therefore, there is also 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). Tadashi Mizumoto, Koshio Matsui; "Study of print capture resistance of digital watermarks using DCT", Journal of the Institute of Electronics, Information and Communication Engineers A, Vol J85-A, No.4, pp.451-459 (2002)

Therefore, it is a problem to obtain a digital watermarking method that is print-resistant and has little moire and deterioration of image quality even if it is strongly embedded. Furthermore, the watermark information is not removed against various attacks from the outside, the watermark can be removed by using a key if necessary, and the security is maintained even if the algorithm is released for widespread use. Satisfying is also an issue.

In order to solve such a problem, the present invention provides a digital watermarking device and a method that are print-resistant with less occurrence of moire and image quality deterioration by embedding in a random cluster dot pattern with frequency band limitation. It is a thing.

Furthermore, because of the random pattern, there is no regularity, and it is highly resistant to external attacks. Also, if necessary, the watermark can be removed using the image-specific key.

The digital watermark of the present invention uses a dot pattern exhibiting cluster-type green noise characteristics. Therefore, when viewed in the spatial frequency space, there is no spectrum near the 0 frequency. On the other hand, since the image data is distributed in the vicinity of 0 frequency, there is little overlap between the two spectra. For this reason, the cluster dot pattern, which has good separability and exhibits green noise characteristics, is dispersed and has little anisotropy, and when observed at a clear vision distance, it is visually inconspicuous and has high quality. Can be maintained. Due to these characteristics, even if the watermark is tough, the strength of embedding (gain) may be small, and no deterioration in image quality is felt.

Therefore, the present invention embeds and extracts watermark information by the following method.
For watermark information embedding, the digital image data is divided into blocks of R pixels × R pixels, and the Fourier spectrum corresponds to the bit information (0 or 1) of the watermark information embedded in each block, and the Fourier spectrum is in the low frequency range and Watermark embedded image w (x, y) = i (x, ...) using multiple different dot patterns pi (x, y) (i = 0,1,2, ...) showing reduced green noise characteristics in the high frequency range Generate y) + gain * pi (x, y) (where gain is less than 1 in embedded strength),
Watermark information is extracted by correcting the received or read image data, dividing it into blocks, obtaining the distribution of the Fourier spectrum of each block, improving the contrast of the spectrum distribution, and then embedding information by pattern recognition of the distribution. Is extracted.

Furthermore, for the image w (x, y) with a watermark embedded, the dot pattern pi (x, y) at the time of embedding, the gain at the time of embedding, and the set of embedding information are obtained as a private key before embedding. The original image i (x, y) = w (x, y) --gain * pi (x, y) can be restored, and new information can be embedded as a digital watermark using the key.

INDUSTRIAL APPLICABILITY According to the present invention, it is possible to embed a tough watermark with print resistance while maintaining high image quality. Stable output is possible even in printing because the fluctuation of the main frequency of the cluster dots is small. Furthermore, the edited Affine coefficient can be detected from this dot pattern. Then, by obtaining the corrected image from this coefficient, highly accurate watermark extraction can be performed. Furthermore, it is possible to increase the print resistance by binarizing half toning.

Furthermore, it is possible to return to the image (original image) before embedding by obtaining the key for the image w (x, y) in which the watermark is embedded. The key is different for each delivered image and is powerless for other images. It is also possible to embed new secondary copyright information as a digital watermark using the key.

The figure of the apparatus for embedding and extracting a digital watermark of the present invention, Diagram showing the distribution of watermarked images on the Internet, A diagram showing the dot pattern for each gradation in the generated dither matrix and its Fourier spectrum. Diagram showing changes in the main wavelength of the green noise pattern for each gradation In the figure showing the pattern of digital watermark embedding, (a) has the spectral characteristics of a vertically elongated elliptical ring, (b) is a 90-degree rotation of it, and (c) is header and registration, and Affine coefficient detection. for. In the image embedded with different gains and a partially enlarged view, (a) is gain = 0.0625, (b) is 0.125, and (c) is 0.25. The figure which shows the image data of 512 pixels × 512 pixels at the time of embedding with gain = 0.125, the image data at the time of printing and reading, and the spectrum distribution thereof. The figure which shows the derivation of the Affine transformation coefficient 10 standard images used for evaluation Shows the evaluation results. The image embedded at the time of binarization and its spectrum, The figure when the embedded image is read by a digital camera and extracted, (a) is the image embedded at the time of binarization, (b) is the image taken by the camera with the monitor, and (c) is from (b). The clipped image, (d) is a spectrum image. A diagram showing two different patterns with the same spectral shape, The figure which shows the embedded image and the spectral distribution when using the dither matrix which shows the different spectral distribution (round type and square type). Diagram showing the mask pattern of the pattern recognition classifier, The figure which showed the image with and without correction image processing and its spectrum at the time of gain = 0.125. Diagram showing the response characteristics of the human visual system, The figure which showed the embedded image when the visual response characteristic was taken into consideration.

FIG. 1 is a block diagram of a digital watermarking system for embedding and extracting information of 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 image-processed by the image processing program in the program memory 6 using the CPU 11, ROM 4, RAM 5, and the like, and displayed on the monitor 8. A scanner 1 and a printer 2 are connected to the computer 3, the processed image is output from the printer, and the image can be read from the scanner 1. Since such image processing is often a high-load process, a processing board such as a GPU may be included to increase the speed.

Here, in order to embed the copyright information in the image, for example, the name of the author and character information such as URL are input from the keyboard 9. The number of characters that can be inserted depends on the image size and block size. When the block size is 64x64 with 512 pixels x 512 pixels, a maximum of 8 bytes is possible, and when the block size is 32x32, a maximum of 32 bytes is possible. It is appropriately selected according to the application and purpose. Since this information is embedded in the image and is difficult for the human eye to see as invisible digital watermark information, it is difficult for a third party to notice.

FIG. 2 shows a processing procedure for Internet distribution. The image in which the watermark information is embedded is distributed by the Internet 12 from the computer of the copyright holder 16 (13). 17 who wants to purchase after seeing it, and inform the distributor of the desire to purchase (14). After the prescribed procedure, the copyright holder distributes the private key for embedding and removing the watermark (15). Based on the contract with the copyright holder, the purchaser removes the digital watermark embedded in the sent image and edits the image.

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

To monitor illegal copying and use, you can use watermark extraction software to read watermark information from images and confirm that it is your own copyrighted work. This watermark extraction software is not confidential, and can be freely downloaded from the copyright holder's homepage, etc., and widely published, and anyone can use it. This allows a third party to know the owner, copyright information, contact information, etc. of the image, which also warns of unauthorized use.

There is a one-to-one correspondence between the private keys and each image to be embedded. For this reason, copyright holders use different keys for each image to prevent misuse. Therefore, this key cannot be used to remove watermarks in other images. 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.

Hereinafter, description will be given with reference to Examples.
The present invention is a method of adding a cluster-type dot profile showing green noise characteristics as a digital watermark to image data. Therefore, first, a dot profile creation algorithm showing green noise characteristics will be described, but since it is described in detail in the applied Japanese Patent Application Laid-Open No. 2016-163197, an outline will be described here.
Now, let the dither matrix size to be obtained be R × R (R = 2 ^ m), and set the blackening rate representing gradation to g (0 ≤ g ≤ 1: g = 1 is all black, g = 0 is all white). do. First, the intermediate (g = 1/2) dot profile is set to the blackening rate g, and the dot profile at the point (i, j) is p (i, j, g), and the following is performed.
(1) First, the random number generator generates R ^ 2/2 random dots (white noise in the initial state), and sets them as p (i, j, 1/2). Two-dimensional Fourier transform of the dot profile is performed to obtain P (u, v, 1/2).
(2) Apply the filter D (u, v, 1/2) to P (u, v, 1/2) to obtain a new P'(u, v, 1/2). Here, D (u, v, 1/2) is a filter whose radial frequency fr has a value in the region of fmin to fmax.
(3) Perform an inverse Fourier transform on P'(u, v, 1/2) to obtain a multi-valued point profile p'(i, j, 1/2).
(4) Find the error function e (i, j, 1/2) = p'(i, j, 1/2) -p (i, j, 1/2) in descending order of error at each pixel position. Invert white and black.
(5) Repeat the above operation until the error is within the permissible amount, and finally obtain a dot profile with g = 1/2.

Then, a point profile with g ± Δg is created by the following procedure.
(1) Fourier transform of p (i, j, g ± Δg) is performed to obtain P (u, v, g ± Δg).
(2) Apply the filter D (u, v, g ± Δg) to P (u, v, g ± Δg) to obtain a new P'(u, v, g ± Δg). Here, D (u, v, g ± Δg) is a filter in which fmin and fmax change depending on g.
(3) Perform an inverse Fourier transform on P'(u, v, g ± Δg) to obtain a multi-valued dot profile p'(i, j, g ± Δg).
(4) Find the error function e (i, j, g ± Δg) = p'(i, j, g ± Δg) -p (i, j, g ± Δg) in descending order of error at each pixel position. Only N = R ^ 2/2 ^ n are inverted in white and black.
Repeat steps (1) to (4) above, and end this operation with g = 0,1.

Here, the settings of the radial frequencies fmax and fmin will be described.
The fmax and fmin of the filter D (u, v, g) are made variable by g so that they approach the blue noise characteristics in the bright and dark areas. The frequency due to the average dot spacing in g is
When 0 ≤ g ≤ 0.5 fo = √g ・ fn
When 0.5 <g ≤ 1, fo = √ (1-g) ・ fn
Given by, fmax and fmin are based on fo
a ・ fn≡fmax --fo
b ・ fn≡fmin --fo
Is expressed using the parameters (a, b) (fn is the Nyquist frequency). By changing (a, b), dither matrices with different cluster sizes can be obtained.

Fig. 3 shows the dot profile of (a, b) = (1/8, -1/8) and its spectral characteristics, and Fig. 4 shows the variation of the principal wavelength with respect to the gradation. It can be seen that the main wavelength is constant over a wide range of gradations. This shows that stable image output is possible even with an electrophotographic printer that is vulnerable to environmental changes. As described above, the dot profile p (i, j) with green noise characteristics is obtained, but since the target spectral distribution is symmetrical with respect to the x-axis and y-axis, the imaginary part of p (i, j) is It becomes 0.

Next, a watermark embedding algorithm for multi-valued image data will be described.
Now, let's assume that different spectral patterns are embedded in block units in the frequency space, and the different spectral patterns are Pi (u, v) (i = 0, ..., N). Pi shows the above-mentioned green noise characteristics. Ideally, fmin should be chosen to have a frequency higher than or equal to the maximum frequency of the embedded image, but it is not necessarily limited to this.
Assuming that 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)
Will be. Here, gain << 1 must be satisfied in order to be invisible. In order for the watermark information Pi (u, v) to be extracted accurately, it is desirable that there is little overlap between I (u, v) and Pi (u, v).

Watermark information is embedded in real space. When the previous equation is converted to real space (pixel space) (represented in lowercase letters),
w (x, y) = i (x, y) + gain · pi (x, y) (1)
It becomes. For the watermark information, prepare the dither pattern pi (x, y, g) showing the different green noise characteristics explained earlier. For the bit information (0,1) to be embedded
When embedded bit = 0 → p0 (x, y, 1/2)
When embedded bit = 1 → p1 (x, y, 1/2)
Embed as. Here, p0 and p1 mean a pattern with a gradation of 1/2. When the image data is 1 pixel and 8 bits, it means a 128-level dot pattern. Also, p0 and p1 are binary values of (1,0), but they are set to (1/2, -1 / 2) to preserve the average brightness. Therefore, gain = 1 means to add a value of ± 128.

However, a pattern other than g = 1/2 may be used. In this case, it is necessary to make the values distributed to the plus side and the minus side uneven in order to match the average brightness. It is better to set g = 1/2. It has the best symmetry and is simple.

Next, watermark embedding in binarization will be described. Watermark information on image data is not completely guaranteed because it is easily lost by multiple editing processes and part of it is lost by processing such as cropping and trimming. Therefore, it is effective to insert a digital watermark in the binarization process at the time of printing. It can also be used as a watermark for tracking illegal copies such as forgery by entering the printer equipment number.
Embedding of watermark information at the time of binarization is done in R × R block units, as in the case of multi-valued images.
When the embedded bit = 0, the dither matrix → p0 (x, y, g)
When the embedded bit = 1, dither matrix → p1 (x, y, g)
Binarization is performed by the dither method using.

Only the pattern of g = 1/2 was used for embedding the multi-valued image, but since the embedding at the time of binarization uses the entire g, the quality of the output image is determined by the workmanship of the dither mask. A cluster-type halftone screen that exhibits green noise characteristics is often used as a cluster-type FM screen, and has excellent uniformity with dispersed dots. Also, the boundaries of the dither matrix are usually noticeable, but this problem does not occur because of the random dots.

Embedding at the time of binarization is performed after the watermark information of the multi-valued image data is once removed by the key. However, if the embedding is not strong, it may be binarized as it is without removing it. Since it is output as green noise type dispersed cluster dots, it can absorb fluctuations in the recording engine and output with high image quality.
It is desirable to incorporate this binarization method into a printer. This is because sometimes there is interference with the binarization method inside the printer (for example, the error diffusion method). For color images, for example, it is desirable to embed in yellow ink / toner. At that time, since the spectral characteristics of the YMC ink overlap, it is desirable that the scanner reads through a blue narrow band interference filter in order to improve the color separation performance. However, since it is a special reading system, it is usually output in a single color.
Digital watermarking at the time of binarization can be applied to applications such as capturing images with a smartphone or digital camera and linking them to the Internet, such as a two-dimensional bar code.

To extract the watermark information, the size and inclination of the read image are corrected, and then the spectrum image is obtained in block units.
Fourier transform Eq. (1)
F {w (x, y)} = F {i (x, y) + gain · pi (x, y, 1/2)}
= F {i (x, y)} + gain · F {pi (x, y, 1/2)}
Will be. However, F {} represents the Fourier transform. Normally, F {i (x, y)} is localized near 0 frequency, and F {pi (x, y, 1/2)} is a ring-shaped spectrum surrounding it as shown in Fig. 5. Therefore, the overlap between the two is small, and it is easy to separate the two even if the gain is made small.

To correct the image, first find the coefficient of the applied Affine transformation. Now, if the embedded image w (x, y) is magnified a times, the Fourier transform is that 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. In addition, the image rotated by θ is also rotated by θ in the Fourier spectrum. Therefore, the Affine coefficient applied by measuring the distribution of the spectrum of the embedded image can be obtained. The size of the detection window may be small. For example, when attempting extraction with a window size of 40x40 without considering registration, detection is possible even if FFT (Fast Fourier Transform) is performed with a margin of 64x64.

The watermark information can be removed to obtain the original image by knowing the embedding pattern. From equation (1)
i (x, y) = w (x, y) --gain · pi (x, y)
By receiving the embedded pattern pi (x, y) and gain as "keys", the original image can be restored. In order to completely restore the original image, it is necessary to limit the dynamic range in advance to avoid overflow and underflow of the embedded image.
As will be described later, since there are many patterns having the same spectral characteristics, it is not possible to remove the embedded watermark data of another image with the same key.

To embed the watermark information, the color image data is converted into Y, Cb, Cr, and the watermark information is embedded in the brightness Y. It is most ideal to embed in Blue (yellow when printing). However, it is necessary to perform color separation with high accuracy. If you want to embed a lot of information, you can embed 32 bytes by setting 32 pixels x 32 pixels as the block size. It is also possible to repeatedly embed the same information to increase the reliability at the time of extraction. At this time, it is necessary to put the header information described later in the first bit of the sentence.

Figures 5 (a) and 5 (b) show the two types of dither masks p0 (x, y) and p1 (x, y) used for embedding. p0 (x, y) was (a, b) = (-1/16, -5/16), and the ellipticity was 1.3. p1 (x, y) is a 90 ° rotation of it. The boundary between the two is inconspicuous due to the random dots. Figure 6 shows the image embedded with gain = 0.0625, 0.125 and 0.25 and a partial enlargement of the image. If you increase the gain, the cluster dots will stand out.

If one of the two types of dither mask is rotated by 90 ° from the other dither mask, only one type of "key" needs to be sent. However, in order to improve security, it is preferable to obtain them from different random numbers. In this case, the watermark can be removed only when both keys are available.

FIG. 7 shows image data in which 8-character alphanumeric characters are embedded in an image of 512 pixels × 512 pixels with gain = 0.125, a printed image, and its spectral characteristics. One bit of information is embedded in units of 64 pixels x 64 pixels, and a total of 8 bytes of character information ("Kawamura") is embedded and read and extracted by a scanner. The printer and scanner may be commercially available household items. The print was output at 188 ppi (69 mm x 69 mm for an image of 512 pixels x 512 pixels), and the reading was performed at 600 dpi. Watermark extraction was performed by performing a Fourier transform in units of 64 pixels x 64 pixels, and then watermark information was extracted by pattern recognition of the spectral distribution.

For Affine transformation detection, a dot profile having a rectangular spectral distribution shown in FIG. 5 (c) can be introduced to detect the variable magnification and the angle of rotation. FIG. 8 is enlarged to 125%, rotated by 8 °, printed, and read by a scanner. The Affine coefficient can be obtained by measuring the rectangular pattern after the Fourier transform. Such an Affine detection pattern can be assigned to the most significant bit (MSB) if the embedded characters are limited to ASCII code. It can also be used as a registration pattern and string delimiter.

In order to improve the extraction accuracy of the watermark information, it is essential to correct the scanned image. The correction is to correct the size and slope from the Affine coefficient, and to emphasize the pattern by edge enhancement. After that, Fourier transform is performed for each block. If nothing is done, the average brightness will differ for each block, making it difficult to recognize the next pattern. Therefore, as a pre-processing for pattern recognition, we tried to improve and standardize the contrast by histogram equalization.

Subsequently, the watermark information is extracted by pattern recognition from the spectrum distribution after the Fourier transform. FIG. 15 shows a mask pattern of the classifier used in pattern recognition. The classifier 1 is a vertically long and horizontally long elliptical ring-shaped mask, and this mask is superimposed on the spectrum with a watermark, and it is determined whether the bit is 0 or 1 from the difference of the integrated luminance values of the overlapping portions. The classifier 2 is a classifier that detects only the periphery of the ellipse, and the classifier 3 excludes this part because the spectrum of the image spreads as noise along the x-axis and the y-axis. The most reliable output from each classifier may be selected, or a linear combination may be used to make a comprehensive judgment. In addition, patterns other than these patterns are also conceivable, and it is possible to increase the recognition rate by machine learning.

Next, an evaluation was performed to confirm the accuracy of this method. Figure 9 shows the image used for the evaluation. 16-bit data was embedded in 10 types of standard images of 256 pixels x 256 pixels, printed at 188 ppi (34 mm x 34 mm), and then read and extracted with a scanner. Figure 10 shows the correct answer rate for extracting watermark information from printed images. When gain = 0.25 and 0.125, 100% correct answers were obtained, but when gain = 0.0625, errors of 1 to 3 bits out of 16 bits occurred. Therefore, if it is used with gain = 0.125 or higher, it seems that it is almost possible to extract the watermark information from the printed image.

FIG. 11 shows an accuracy evaluation performed by scanning the embedded material during the binarization process of printing with a scanner. Since the binary image has high contrast, it can be extracted more stably than the multi-valued image.
Further, FIG. 12 shows the result of taking a picture of the display screen with a digital camera and extracting the watermark. An image of 256 pixels x 256 pixels is binarized, 2 bytes of information are embedded, a digital camera of about 20 million pixels is used to take a picture as shown in Fig. (B), and a watermark is extracted by cutting out as shown in (c). However, it was possible to extract with high reliability. This is because the contrast is high regardless of gain because it is binarized. It is not easily affected by lighting unevenness and image distortion, and extraction is possible if the dots can be reproduced. Since the binarized halftone image is equivalent to the cluster type FM screen, the image quality is not a problem.

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

In addition, a collusion attack that identifies a watermark pattern from an original image without a watermark or a plurality of embedded images can be avoided by changing the embedding pattern for each image.

FIG. 13 shows different cluster dot patterns with the same spectral characteristics. Different cluster dots are formed depending on the initial value (random number SEED value) at the time of dot profile creation, and innumerable profiles can be generated. 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 profile of the dot pattern for each image. The copyright holder provides the purchaser with a dot profile obtained from different random numbers as a "private key". Since this key is different for each purchaser, it is not possible to remove the watermark from the images of other purchasers.

For a normal digital watermark, the watermark embedding algorithm can be roughly estimated by comparing it with the original image. For example, what is embedded in the frequency space can be roughly guessed by trying all orthogonal transformations after taking the difference from the original image. It takes time, but the algorithm is elucidated. Therefore, in principle, the watermark algorithm is not disclosed.
On the other hand, even if the algorithm is disclosed, the key to remove the watermark is basically one for one image, so even if the key information can be decrypted, the key is used for other images. The watermark cannot be removed. Since the number of possible keys is a combination of R × R / 2 in R × R (half of that when considering symmetry) in the block size R × R.
Number of keys = (1/2) x (R ^ 2) C (R ^ 2/2)
Here, C indicates a combination. Therefore, the probability that the same key will appear is extremely low.

It can also be used without removing the watermark as it is. When observing an embedded image at a normal size and a clear vision distance, the pattern is hardly recognized due to 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 print resistance can be maintained even if the gain is reduced. In the experiment, gain = 0.125 was found to be sufficiently resistant. This value corresponds to embedding ± 16 data when the pixel data is 8 bits.

FIG. 18 is a simulation of an image in which a watermark is embedded with gain = 0.25 and viewed through a human visual system response characteristic VTF (Visual Transfer Function). FIG. 17 shows the VTF curve. When viewed at a clear vision distance (L = 300 mm), it can be seen that the high-frequency component is attenuated by VTF and the cluster dot pattern becomes inconspicuous even though it is embedded with a large gain. Therefore, it can be said that gain = 0.125 is hardly noticeable.

In addition, since information is extracted from the spatial distribution of dots, it is not easily affected by uneven density and uneven lighting. Extraction errors occur 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 moving 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 restrictions on the quality of the printing paper become strict. If there is a margin in the number of characters to be embedded, one solution is to repeatedly embed the watermark information and select a bit with high extraction reliability to improve the accuracy. Although the amount of information to be embedded is reduced, the embedded block size should be 32 pixels x 32 pixels.

Figure 14 shows another example of the dither pattern, in which the cluster dot pattern showing the spectral characteristics of the round and square shapes is used as the dither threshold. The block boundaries of cluster dots are inconspicuous due to the random dot arrangement. In addition, the Affine coefficient can be easily measured from the square shape.

The image size does not have to be a power of 2. Divide by block size and fill the less than a certain value. Further, since the Fourier transform is performed in block units instead of the entire image, a large-capacity memory is not required and high-speed processing is possible.

In this method, edge enhancement processing that enhances cluster dots is effective for watermark extraction. Figure 16 shows that after reading a watermark image with gain = 0.125 with a scanner, the image is edge-enhanced to improve the detection accuracy. In the case of no processing, the spectrum of the original image jumped out of the elliptical ring, but due to the enhancement processing, it gathered in the low frequency region near 0 frequency and became confined inside the fmin of the spectrum distribution of green noise, and was detected. The accuracy is greatly improved. Since such edge enhancement processing is performed only for recognition, strong edge enhancement is highly effective.

The digital watermarking method of the present invention has been described above. However, since this method embeds a dot pattern showing green noise characteristics in real space, there is little visual discomfort and it is possible to embed it strongly. Since it is not easily affected by distortion and can be restored to the original image by key information, it is not only for copyright protection but also for various applications by linking printed images and digital data. It is valid.

In addition, the watermark algorithm and extraction software may be disclosed. It cannot be unlocked without the image-specific 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 is distribution of image data, 14 is contact of purchase request, 15 is sending of private key,
16 represents the computer of the copyright holder, and 17 represents the computer of the purchaser.

Claims (2)

In the digital watermarking method that embeds copyright information in digital image data i (x, y)
Watermark embedding divides digital image data into blocks of R pixels x R pixels, and the Fourier spectrum corresponds to the bit information (0 or 1) of the watermark information embedded in each block, and the Fourier spectrum is in the low frequency range and high frequency. Watermark embedded image w (x, y) = i (x, y) using multiple different dot patterns pi (x, y) (i = 0,1,2, ...) showing reduced green noise characteristics in the region ) + Gain · pi (x, y) is generated (however, gain is the embedding strength, gain << 1),
Watermark extraction divides receives or reads image data into blocks, obtains the distribution of the Fourier spectrum of each of the blocks, the spectrum distribution after improved contrast, extracts the embedded information by pattern recognition distribution,
Watermark removal is performed by obtaining a set of dot pattern pi (x, y) at the time of embedding, gain at the time of embedding, and embedding information as a secret key for the image w (x, y) with the embedded watermark. Image i (x, y) = w (x, y) -gain · pi (x, y) before embedding can be generated.
The watermark is embedded using a dot pattern generated by an initial value (random number seed value) that is different for each image.
The watermark can be removed by using a secret key having a different dot pattern for each image, and secondary copyright information or the like can be embedded by using this key.
The watermark extraction can use common software for different images even if the private key is different.
A digital watermarking method characterized by that.
As the dot pattern pi (x, y) showing the green noise characteristic, the coefficient of the Affine transformation applied to the image using the pattern showing the rectangular spectral characteristic is obtained, and it is also used as the head (header) information of the embedded information. The digital watermarking method according to claim 1 .

Images