KR100753490B1 - Robust image watermarking using scale invariant feature transform - Google Patents

Abstract

The present invention relates to a digital watermarking method using size invariant feature transformation, and in particular, additional watermarking applied to circular patches extracted through size invariant feature transformation, which is robust to geometric transformation, in order to synchronize watermark positions. It is about a method.

The digital watermarking method using the size invariant feature transformation of the present invention is largely composed of an insertion step and a detection step.

In the watermark embedding step, a 2D rectangular watermark following a Gaussian distribution is generated by using a random number generator, a circular patch is generated by using SHIF, and the inverted polarization of the rectangular watermark according to the size of the circular patch is converted into a circular watermark. The watermark detection step includes generating a circular patch using a SHIF and calculating a noise by applying a Wiener filter to detect a watermark inserted in the circular patch and polarize the detected circular watermark. Converting into a rectangular watermark through the;

Digital watermarking, robust watermarking, watermark synchronization, size-invariant feature conversion

Description

Digital watermarking method using size-invariant feature transformation {ROBUST IMAGE WATERMARKING USING SCALE INVARIANT FEATURE TRANSFORM}

1 is a flowchart illustrating a digital watermarking insertion method using a size invariant feature transform (SIFT) of the present invention.

2 is a flowchart illustrating a digital watermarking detection method using a size invariant feature transform (SIFT) of the present invention.

3 is a diagram illustrating neighboring pixels for one pixel in a scale space and a scale space by a DoG function.

4 shows (a) original image, (b) average filtered image, (c) median filtered image, (d) additional uniform noise image, (e) 10 ° rotated image, and (f) 1.2 × scaling In the image obtained is a circular patch produced by the method of the invention.

5 is a diagram illustrating polar coordinates mapping between a rectangular watermark and a circular watermark of the present invention.

6 is a diagram illustrating a watermark embedding framework of the present invention.

7 illustrates a watermark detection framework of the present invention.

8 shows a histogram of the similarity value and gamma distribution thereof of the present invention.

9 is a table showing error probabilities and thresholds of the watermark detector of the present invention.

10 is a table showing the proportion of correctly redetected patches of the present invention and the number of failed images under various attacks.

11 is a table showing the proportion of correctly detected watermark patches, the number of failed images, and similarity under normal signal processing attacks of the present invention.

12 is a table showing the proportion of correctly detected watermark patches, the number of failed images, and the similarity under geometric deformation attack of the present invention.

FIG. 13 shows (a) original image, (b) watermarked image, and (c) residual image in lake, monkey, milk and leg images of the present invention.

The present invention relates to a digital watermarking method using a scale invariant feature transform (SIFT), and particularly to a size invariant feature transform (SIFT) that is robust to size transformation or position transformation in order to synchronize the position of a watermark. The present invention relates to an additional watermarking method applied to a circular patch extracted through the patch.

With the development of computer technology, various multimedia contents have been digitized. Digital multimedia has many advantages over analog media. In other words, digital multimedia can be easily accessed, copied, and distributed many times without degradation in quality. However, widespread use of digital multimedia has also caused problems with the protection of copyright.

Digital watermarking is an efficient solution for copyright protection, which inserts copyright information, i.e., a watermark, into the content itself. Ownership of the content can be verified by searching the inserted watermark. F. A. P. Petitcolas, R. J. Anderson, and M. G. Kuhn, "Attacks on copyright marking systems", Proc. Int. In Workshop on Information Hiding, 218-238 (1998)., Several attacks have been reported to be effective against watermarking methods. Among them, geometric deformation is known to be one of the most difficult attacks to prevent. Geometric deformations asynchronous the position of the watermark and thus cause inaccurate watermark detection. In this case, the watermark synchronization process must calculate the watermark position prior to watermark insertion and detection, which is very important for the robustness of the watermarking system.

Several studies related to watermark synchronization have been conducted for some time.

Among them are methods using periodic sequences, methods using templates, and methods using invariant transforms. There is a scheme for watermark synchronization using media content and the method of the present invention belongs to this scheme. Media content represents an invariant criterion for geometric deformations, and therefore, the problem of watermark synchronization can be solved by referring to the content. `` M. Kutter, S.K. Bhattacharjee, and T. Ebrahimi, "Toward second generation watermarking schemes", IEEE Int. Conf. on Image Processing 1, 320-323 (1999). The position of the watermark is in relation to image semantics, not image coordinates. In the following, a position for watermark embedding and detection is called a patch.

P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). Bas et al. Proposed a content-based synchronization method, which first extracts prominent feature points and then decomposes the image into a series of disjoint triangles through Delaunay tessellation. This series of triangles (patches) is used to insert and detect watermarks in the spatial domain. The disadvantage of this method is that the extracted feature points from the original image and the deformed image do not coincide. Thus, the series of triangles generated during watermark embedding and detection are different from each other, so that the patches do not match.

Thesis A. Nikolaidis and I. Pitas, "Region-based image watermarking", IEEE Trans. Image Process. 10 (11), 1726-1740 (2001). Nikolaidis and Pitas proposed an image-segmentation based synchronization method. By applying adaptive k-mean clustering techniques, they segment the image, select some of the largest regions, and approximate them as ellipsoids. The smallest rectangle surrounding the ellipsoid is used as the patch. The problem with this method is that image segmentation depends on the image content, so that image deformation seriously affects the segmentation result.

Thesis C-W. Tang and H-M. Hang, "A feature-based robust digital image watermarking scheme", IEEE Trans. Signal Process. 51 (4), 950-959 (2003). ”Tang and Hang introduced a synchronization method using feature-based feature extraction and image normalization. They extract feature points by Mexican hat wavelet scale interaction and perform normalization using the contents of a fixed radius disk centered on each feature point. These normalized disks are used as patches. However, normalization is sensitive to the image content used, so the robustness of these patches is reduced when the image is deformed.

In the content-based synchronization method, feature selection is important to achieve robust watermarking. However, conventional watermark synchronization has a problem in that it does not provide a watermarking method using the SIFT feature, which is robust to geometric deformation.

In order to solve the above problem, the present invention provides an additional water applied to a circular patch extracted through a size invariant feature transform (SIFT) that is robust to size conversion or position conversion in order to synchronize watermark positions. It is an object to provide a marking method.

The digital watermarking method using the size-invariant feature transformation of the present invention is largely composed of an inserting step and a detecting step. The first step of inserting a watermark includes a method of generating a two-dimensional rectangular watermark following a Gaussian distribution using a random number generator. Stage 1; Generating a circular patch using a scale invariant feature transform; A third step of converting a rectangular watermark into a circular watermark by inverse polar mapping according to the size of the circular patch; A fourth step of calculating a recognition mask for the circular patch; And a fifth step of additionally inserting a watermark in a spatial domain using the circular watermark and the recognition mask, wherein the watermark detection step includes a circular patch using a scale invariant feature transform. Generating a first step; A second step of detecting a watermark inserted in the circular patch by calculating a noise by applying a Wiener filter; A third step of converting the detected circular watermark into a rectangular watermark through polar mapping; And a fourth step of comparing similarity between the reference watermark used in the insertion process and the detected watermark through a circular convolution, and verifying the similarity. To present.

1 is a flowchart illustrating a process of embedding a digital watermark using the size invariant feature transformation of the present invention. First, a two-dimensional rectangular watermark along the Gaussian distribution is generated using a random number generator (S100). A circular patch existing on the image is generated using a size invariant feature transform (SIFT) (S200). For each circular patch, the rectangular watermark is converted into a circular watermark through inverse polar mapping to the size of the patch (S300). A recognition mask is calculated for each circular patch (S400). Using the circular watermark and the recognition mask, a watermark is inserted in an additional manner in a spatial area (S500).

2 is a flowchart illustrating a process of detecting a digital watermark using the size invariant feature transformation of the present invention. First, circular patches existing on an image are generated using a size invariant feature transform (SIFT) (E100). The inserted watermark is detected by applying a Wiener filter to the angular patch to calculate the noise embedded in the image (E200). The detected circular watermark is converted into a rectangular watermark through polar mapping (E300). Through the circular convolution, the ownership is determined by comparing the reference watermark used in the watermark embedding process with the detected watermark to verify similarity (E400).

Hereinafter, with reference to the accompanying drawings, the configuration and operation of the embodiment of the present invention will be described in detail.

1.Local Invariant Feature

Affine local invariant features in object recognition and image retrieval applications have recently been studied. These local invariant features are very distinct and are redetected with high probability for large image deformations. In content-based watermark synchronization, robust extraction of patches is associated with the robustness of the watermarking system, and consideration of local image characteristics helps in reliable extraction of patches. The present invention proposes a new synchronization method based on SIFT.

(1) SIFT

DG Lowe, "Distinctive image features from scale-invariant keypoints", Int. J. Comput. In Vision 60 (2), 91-110 (2004)., Lowe proposed SIFT, which is invariant to image rotation, scaling, shifting, partial illumination variation, and projective transform. In view of the local image characteristics, the SIFT extracts its properties such as the position of the feature (t ₁ , t ₂ ), the scale s and the direction θ.

SIFT extracts features through a series of stepwise filtering processes that identify stable points in the scale-space. That is, (1) the candidate of the feature is selected by searching for the local maximum or minimum point in the scale space generated from the difference-of-Gaussian (DoG) function, and (2) the feature of each feature using its stability dimension. And localize (3) assign the direction of the feature based on the directional information of the adjacent pixels.

를 곱하여 스케일을 증가시킨다.To extract candidate positions of features, the scale space D (x, y, σ) is calculated using the DoG function. As shown in Fig. 3, the scale-space image is calculated by continuously smoothing the original image with the variable scale (σ1, σ2 and σ3) Gaussian functions and subtracting the two successive smoothed images. The parameter σ is the variance (called scale) of the Gaussian function. The scale of the scale-space image is determined by the adjacent scale (σ1, σ2 or σ3) of the Gaussian smoothed image. In these scale-space images, the eight nearest neighbors on the same scale and the nine neighbors on the lower and upper scales are examined to retrieve all local maximums and minimums. These extremes determine the position (t ₁ , t ₂ ) and scale s features of the SIFT feature, which are invariant to changes in the scale and orientation of the image. In the present invention, to generate a scale-space image, the scale of the Gaussian function is applied from 1.0 to 32.0 and a constant factor

Multiply by to increase the scale.

After the candidate positions are found, the precise values are calculated after detailed modeling of the positions (t ₁ , t ₂ ) and scale s of each feature with 3-D quadratic functions. In addition, candidate positions having low contrast or poorly localized along the edges are removed after measuring the stability of each feature using a 2 × 2 Hessian matrix H as shown in Equation 1 below.

[Equation 1]

,

, 및

는 스케일-공간 이미지의 도함수이다.Here, the value e is the ratio between the largest and smallest eigenvalues and is used to control the stability and uses 10 as the value of e.

,

, And

Is the derivative of the scale-space image.

To achieve invariance to image rotation, assign a consistent direction to each feature. In a Gaussian smoothed image with the scale of the extracted SIFT feature, the directional information of all sample points in the circular window of the feature location is calculated and used to form a directional histogram. The maximum point in this histogram corresponds to the dominant direction of the feature.

(2) variants for watermarking

Here, we describe how to create a circular patch for watermark embedding and detection using SIFT. Local features from SIFT cannot be applied directly to watermarking because the number and distribution of features depends on the image content and texture. In addition, SIFT was originally designed for image-matching applications, thus extracting many features with dense distribution over the entire image. Thus, we adjust the number, distribution, and scale of features and remove features that are vulnerable to watermark attacks.

SIFT extracts features with attributes such as position (t ₁ , t ₂ ), scale s and direction θ. In fact, the orientation properties of the SIFT extracted from the original image and the deformed image do not exactly match. Therefore, a circular patch is made using only the position (t ₁ , t ₂ ) of the extracted SIFT feature and scale s as shown in Equation 2.

[Equation 2]

Where k is a magnification factor that controls the radius of the circular patch. The method of determining this coefficient is described in the fourth paragraph. These patches are immutable to signal scaling attacks as well as image scaling and movement. By applying a prefilter such as a Gaussian filter before feature extraction, the interference of noise can be reduced and the robustness of the extracted circular patch can be improved.

The scale of the feature derived from the SIFT is related to the scaling factor of the Gaussian function in scale space. Features with a small scale are less likely to be redetected because they are easily lost when the image content is modified. Features with large scales are also less likely to be redetected in the deformed image because their positions easily move to other positions. In addition, the use of large scale features means that the patches overlap each other, resulting in degradation of the perceptual quality of the watermarked image. Thus, in the present invention, features having a scale above a or below b are removed and a and b are set to 2.0 and 10.0, respectively.

SIFT features extracted through SIFT have different values in scale s depending on the image content. Watermark embedding and detection must interpolate to convert the rectangular watermark to match the shape of the circular patch or vice versa. In order to minimize deformation of the watermark through interpolation, the size of the patch (ie, the radius) must be similar to or larger than the size of the watermark. The scale s of the extracted SIFT feature varies from 2.0 to 10.0. Thus, the scale of the feature is divided into two ranges to apply different magnification coefficients k1 and k2, which are determined empirically on the assumption that the size of the watermarked image does not change excessively.

The distribution of local features is related to the performance of the watermarking system. In other words, the distance between adjacent features must be carefully determined. If the distance is small, the patch has a large area overlapping with the adjacent patch, and if the distance is large, the number of patches for effective insertion of the watermark is insufficient. To control the distribution of the extracted features, see P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). We applied a circular neighborhood constraint similar to that used by Bas et al., Whereby the feature with the largest intensity is used to create a circular patch. The value from the DoG function is used to measure the strength of each feature. The distance between adjacent features depends on the dimensions of the image and is quantized by the r value as shown in equation (3).

 [Equation 3]

Here, the width and height of the image are represented by w and h, respectively. The r value is a constant that controls the distance between adjacent features and is set to 16 and 32 respectively in the insertion and detection process.

4 shows a circular patch from the synchronization method proposed in the present invention in the signal processing filter, additive uniform noise, rotation and scaling of the image. For convenience of identification, only one patch is shown and it can be seen that the patch is robustly generated even when the image is deformed.

2. Watermarking method

In the present invention, a circular patch is extracted using a size invariant feature transform (SIFT) as described above. A 2-D watermark is generated and converted into a circular shape to fit each patch. A circular watermark is inserted and detected in addition to the patch. First, a watermark generation procedure will be described, followed by watermark embedding and detection.

(1) watermark creation

A random number generator is used to generate a 2-D rectangular watermark that follows a Gaussian distribution. To be inserted into a circular patch, this watermark must be converted so that its shape is circular. Consider a rectangular watermark as a polar-mapped watermark and inversely polar-map it to assign the insertion location of the circular patch. As such, the rotational attack is mapped as a movement in a rectangular watermark, which can still be detected using a correlation detector. Note that the size of the circular patches is different and therefore a separate circular watermark must be created for each patch.

The x and y sizes of the rectangular watermark are denoted by M and N, respectively. The radius r of the circle is equal to the radius (size) of the circular patch. As shown in FIG. 5, the circle, ie, circular patch, is divided into homocentric regions. The x and y axes of the rectangular watermark are inverted polar coordinates in the radial and angular directions of the circle. The relation between the coordinates of the rectangular watermark and the circular watermark is expressed as in Equation 4.

 [Equation 4]

으로 설정하였다. 효과적인 변환을 위해, r_o은

보다 커야 하고 r_M와 r_o간의 차가 N보다 커야 한다. 이들 제약이 만족되지 않는 경우, 직사각형 워터마크는 샘플링되어야만 하고, 그로 인해 효율적으로 변환하기가 어렵다.Where x and y are rectangular watermark coordinates and r _i and θ are the coordinates of the circular watermark. r _M is the radius of the circle r and r _o is the radius of the circle determined relatively from r _M. In the present invention, r _o

Set to. For effective conversion, r _o is

It must be greater than and the difference between r _M and r _o must be greater than N. If these constraints are not satisfied, the rectangular watermark must be sampled, thereby making it difficult to convert efficiently.

To improve the robustness and invisibility of the embedded watermark, the rectangular watermark to be mapped is converted only to the upper half of the circle, i.e., the y axis of the rectangular watermark is scaled by the angle π of the semicircle, not the angle of the complete circle 2π. The lower half of the circle is set symmetrically using the upper half of the circle (see FIG. 5). If noise of similar intensity gathers in adjacent areas, the noise can be recognized. In the present invention, pixels in a rectangular watermark are mapped to several adjacent pixels in the circular watermark during polar mapping. In other words, the same noise (watermark constituting some kind of noise) is inserted in the concentric region of the circular watermark. Therefore, when the size of the concentric region is large, the inserted watermark appears as an embossing effect. By symmetrical mapping, the size of the concentric region can be made small, and thus a watermark can be inserted invisibly. In addition, it is possible to improve the likelihood that the watermark will withstand attacks such as cropping.

(2) Insert watermark

The first step of watermark embedding is to analyze the image content to extract the patch. The watermark is then repeatedly inserted in all the patches. The watermark embedding process is shown in FIG.

(Step a) Extract the circular patch using SIFT. A single image can contain multiple patches. Insert watermarks into all patches to improve robustness.

(Step b.1) A circular watermark is generated according to the radius (size) of each patch. While extracting the circular patch, we tried to configure the patch so that the radius r of the patch is similar to or larger than the x and y size of the rectangular watermark, so that the pixels in the rectangular watermark (w) may be a few pixels in the circular watermark (wc). Mapped to pixels This may compensate for errors in the position and size of the circular patch in the watermark detection process.

를 적용한다.(Step b.2) The embedding of the watermark should not affect the quality of the image. This constraint is related to the insertion strength of the watermark. Perceptual mask of the following equation (5)

Apply.

[Equation 5]

Where α is the lower limit of visibility in flat and flat areas and β is the upper limit in textured areas with edges. The noise visibility function NVF (i, j) is calculated by equation (6).

[Equation 6]

및

는 각각 5개의 픽셀 내의 이웃 픽셀의 국소 분산 및 최대값을 나타낸다. D는 크기를 나타내는 상수값이다.here,

And

Represents the local variance and the maximum of neighboring pixels in each of the five pixels. D is a constant value representing magnitude.

(Step b.3) The circular watermark is additionally inserted into the spatial domain. The insertion of the watermark is expressed as a spatial addition between the pixels of the image and the pixels of the circular watermark as shown in equation (7).

[Equation 7]

및

는 각각 이미지 및 원형 워터마크의 픽셀을 표시한다.

는 워터마크의 삽입 강도를 제어하는 인식 마스크를 나타낸다.here,

And

Displays pixels of an image and a circular watermark, respectively.

Denotes a recognition mask that controls the embedding strength of the watermark.

(3) watermark detection

Similar to watermark embedding, the first step of watermark detection is to analyze the image content to extract the patch. The watermark is then detected from the patch. If the watermark is correctly detected from at least one patch, then ownership can be proved successfully. The watermark detection process is shown in FIG.

(Step a) To extract the circular patch, use the SIFT. There are several patches in the image, and I try to detect the watermark from all the patches.

(Step b.1) The additional watermark method in the spatial domain embeds the watermark as noise in the image content. Therefore, Wiener filter is applied to extract this noise. There are various kinds of filters such as Wiener filter, Inverse filter, Gaussian filter, but here Wiener filter is applied. That is, the difference between the watermarked image and its Wiener-filtered image was calculated and considered as a retrieved watermark. As in the watermark embedding process, correction is compensated by the recognition mask, but this compensation does not significantly affect the performance of watermark detection.

(Step b.2) In order to measure the similarity between the reference watermark generated during watermark insertion and the retrieved watermark, the retrieved circular watermark must be converted to a rectangular watermark by applying the polar mapping. Taking into account the fact that the watermark is symmetrically inserted, the average value from the two areas is taken. By this mapping, the rotation of the circular patch is represented as a movement, thus achieving rotational invariance for the watermarking scheme of the present invention.

(Step b.3) Apply circular convolution to the reference watermark and the retrieved watermark. The similarity between the two, called the response of the watermark detector, is expressed by the maximum value of the cyclic convolution as shown in Equation (8).

[Equation 8]

Where w is the reference watermark and w * is the retrieved watermark. Similarity values range from -1.0 to 1.0. If the X and Y sizes of the watermark are M and N, respectively, m and n satisfy 0 ≦ m <M and 0 ≦ n <N as coordinates on the watermark. The rotation angle (π / r) of the patch can be checked by finding r having the maximum value. If the similarity exceeds a predetermined threshold, it can be considered that the reference watermark has been inserted. The method of determining the threshold is described in the section below.

(Step c) As described above, there are several circular patches in the image. Thus, if a watermark is detected from at least one patch, ownership is verified, otherwise it is not. The fact that watermarks are inserted into several circular patches rather than just one increases the likelihood that the proposed scheme will detect watermarks even after image transformation.

The watermarking method of the present invention is robust to geometric deformation attack as well as signal processing attack. Scaling and movement invariance are achieved by extracting circular patches from SIFT. Rotational invariance is successfully obtained by using an attribute whose rotational transformation is represented as a translational transformation for polar mapped circular patches.

(4) error probability

Since ownership is verified by determining whether the similarity exceeds a predetermined threshold, the probability that the watermark detector will generate an error depends on which threshold is selected. Errors are classified into two types: false positive error and false negative error. False positive determination error refers to the probability that a watermark will be detected correctly when the image is not watermarked. False speech determination error refers to the probability that the method fails to retrieve the inserted watermark from the watermarked image. Indeed, it is difficult to analyze false negative judgment errors due to the wide range of possible attacks. Therefore, it is common to select a threshold based on a false positive decision error.

In order to estimate the probability of false positive determination error of the watermark detector, an attempt was made to retrieve 100 random watermarks from 100 randomly collected images including natural scenery and portraits. A total of 168,500 circular patches were processed to detect the watermark, because each image contains several patches. The size of the rectangular watermark was 32 x 32 pixels. A histogram of the similarity value (normal correlation value) and its gamma distribution are shown in FIG. 8.

『I. In most cases, J. Cox, ML Miller, and JA Bloom, "Chap. 5 in Digital Watermarking, Morgan Kaufmann Publishers", San Francisco, CA (2002). Suppose that the distribution of follows a Gaussian or approximated Gaussian distribution model. However, the distribution of the response of the watermark detector further follows the gamma distribution model because it takes the maximum value from the cyclic convolution. The gamma distribution is defined as in Equation 9.

 [Equation 9]

Where x is a continuous random variable. The parameters α and β satisfy the α> 0 and β> 0 and are calculated using the mean and variance of the random variable x as shown in equation (10).

[Equation 10]

Based on this result, the mean and variance of the detector response x were 0.1012 and 1.2191e-004, respectively, and the values of the parameters α and β were 83.9865 and 0.0012, respectively. 9 shows the probability that the watermark detector will generate a false positive determination error and the selected threshold of the watermark detector when using this gamma distribution model.

4. Simulation Results

Here, the performance of the watermarking method of the present invention will be described. In order to measure (1) the performance of watermark synchronization based on SIFT and (2) the robustness of the watermarking scheme of the present invention, two experiments were performed. 100 512 × 512 images randomly collected from the Internet, including images of Lena, monkeys, lakes and bridges, were used.

(1) SIFT performance

A synchronization method based on SIFT is described in the paper P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). The performance of the synchronization method of the present invention is shown by comparison with the method of. During watermark detection, redetection of patches extracted during watermark insertion is important for robustness. To measure the redetection rate, a circular patch was first extracted from both the original image and the attacked image and then the position and radius of the patch from the original image were compared with that of the patch from the attacked image. Prior to the comparison, the position and radius of the circular patch from the geometrically attacked image were converted to the position and radius of the patch in the original image. If the difference between the position and the radius of the circular patch from the original image and the patch from the attacked image is 2 pixels or less, the patch was considered to be correctly redetected.

Various attacks: median filters (2 × 2, 3 × 3, and 4 × 4), JPEG compression (Quality factor 40, 50, 60, 70, 80, and 90), Gaussian filtering (3 × 3) ), Additional uniform noise, centered cropping (5%, 10%, 15%, 20%, and 25%), rotation + cropping (0.25 °, 0.5 °, 1 °, 5 °, 10 °, and 30 °), and scaling (0.75 ×, 0.9 ×, 1.1 ×, 1.3 ×, and 1.5 ×) were applied.

The average number of circular patches extracted from 100 original images is described in P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). 22.60 in the method and 16.85 in the method of the present invention. Comparison results for the attack are shown in FIG. 10. Detection ratio refers to the ratio of the number of patches extracted from the original image to the number of correctly redetected patches from the attacked image. The detection rate is high when watermark synchronization is performed more strongly. Detection failure refers to the number of images in which none of the 100 images have been redetected.

As shown in Fig. 10, the detection rate drops in proportion to the strength of the attack in both synchronization methods, but the synchronization method of the present invention is described in P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). Better performance than the In particular, as the intensity of the attack increases, the paper P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). The feature points of the method are removed or newly created. As a result, the detection rate of the patches generated in consideration of the relationship between the feature points is drastically reduced and the number of failed images is increased. However, the method of the present invention is performed reliably because we only consider local image characteristics.

These results support the claim that SIFT is a reliable technique for extracting features (patches) and is useful for robust watermarking against signal processing attacks and geometric deformation attacks.

(2) Watermarking performance

Here, the performance of the watermarking method of the present invention was tested.

The size of the rectangular watermark was 32 × 32 pixels, and the weighting factors α and β of the noise visibility function were set to 5.0 and 1.0, respectively. By setting the threshold to 0.182, a 10-9 error probability (reliability) was obtained for the proposed scheme, and a circular patch in the range of -2 to +2 to compensate for the error of patch size caused by the synchronization method of the present invention. An attempt was made to detect the inserted watermark by searching the radius. There are several overlapping areas between the circular patches, but the inserted watermarks in these areas are not visible to the naked eye because the watermarks are inserted in consideration of the perceptual masking.

The overall PSNR value between the original image and the watermarked image was over 40 dB. Here we inserted a watermark in the circular patch, and the image was only partially corrected. As a result, the scheme of the present invention achieved a high PSNR value. In highly textured images, such as known monkey images, the PSNR value is relatively low because the watermark is strongly inserted due to the fact that the noise cannot be perceived.

Most of the Stirmark 3.1 attacks here, namely median filter, JPEG compression, Gaussian filtering, additional uniform noise, linear geometric transformation, random bending, row-column rejection, shearing, centered cropping ), Rotation + cropping, scaling, and rotation + scaling + cropping were applied. These attacks attempt to remove or attenuate the embedded watermark and to asynchronous the position of the embedded watermark.

Simulation results under attack are shown in FIGS. 11 and 12. The detection rate refers to the ratio of the number of inserted watermarked patches to the number of watermarked patches accurately detected in watermark detection. Of 100 images, the detection failure represents the number of images in which the inserted watermark was not detected from any patch, thus failing to prove ownership. Similarity is the average of similarity values from watermarked patches that are correctly detected.

In most attacks, the watermarking scheme of the present invention was able to detect embedded watermarks from a significant number of circular patches and the similarity between the inserted watermarks and the detected watermarks was high enough to prove ownership. When converting a rectangular watermark to a circular watermark, the pixels of the rectangular watermark are mapped to concentric regions of the circle, ie multiple pixels, which compensate for the position error for the circular patch. As a result, the watermarking scheme of the present invention could function well even in linear geometric transformations, random blending, and shearing attacks. Since only local properties of the features are considered here, the scheme of the present invention could also function in cropping attacks. In rotation + cropping and rotation + scaling + cropping attacks, the intensity of cropping and scaling increased in proportion to the angle of rotation, thus reducing the detection rate.

Although watermark synchronization is well performed, when the image is strongly attacked by methods such as JPEG compression 40, scaling 0.70 ×, and center cropping 50%, an additional watermarking method in the spatial domain is limited in several images. It didn't work. In these attacks, the inserted watermark is almost eliminated, attenuated or partially cropped by deformation.

Nevertheless, these performances are described in the paper P. Bas, J-M. Chassery, and B. Macq, "Geometrically invariant watermarking using feature points", IEEE Trans. Image Process. 11 (9), 1014-1028 (2002). And the paper C-W. Tang and H-M. Hang, "A feature-based robust digital image watermarking scheme", IEEE Trans. Signal Process. 51 (4), 950-959 (2003).

As mentioned above, other synchronization methods exhibit lower performance than the synchronization method of the present invention. Thesis C-W. Tang and H-M. Hang, "A feature-based robust digital image watermarking scheme", IEEE Trans. Signal Process. 51 (4), 950-959 (2003).], Their schemes were robust against several signal processing attacks, such as JPEG compression, Gaussian filters, and additional uniform noise.

However, when the image is deformed by geometric deformation or the pixels of the image are removed by median filter, row-column removal, or cropping, the detection rate drops considerably and the probability of detection failure is high.

The above results support the claim that the watermarking scheme of the present invention is robust against various image attacks. An advantage of the present invention is that the watermark is inserted into the image several times, and as a result the watermark remains under attack, thereby correctly proving ownership.

5. Review and conclusion

The difference between the matching items in Figs. 10, 11 and 12 indicates the number of circular patches when the circular patches are well synchronized but the additional watermarking method does not detect the inserted watermark. As mentioned above, the watermark synchronization of the present invention based on SIFT can accurately extract patches even after an image attack and even on images with complex textures. However, this is because the additional watermarking method in the spatial domain was not able to successfully detect the inserted watermark.

FIG. 13 shows several original images, a watermarked image, and a residual image between the original image and the watermarked image. We have modified the histogram of the residual image for convenience. As shown in (a) and (b) of FIG. 13, a watermark is inserted in the image so that it is not visible to the naked eye. As shown in Fig. 13C, the residual image shows the position and radius of the circular patch, how the rectangular watermark is formed concentrically, and how the additional watermarking method embeds the watermark in the spatial domain. The scheme of the present invention achieves a high PSNR value because the image is partially modified.

If the image is well-textured, the patch is distributed throughout the image, thereby allowing the watermark to be inserted over the entire image. However, if the texture of the image is simple, for example in the sleeping area in milk or in the sky area of the lake, the watermark is concentrated in the area near the object.

In the present invention, a robust watermarking method using local invariant features is proposed. That is, in order to prevent geometric deformation, the present invention extracts a circular patch using SIFT which is invariant to movement and scaling deformation. These circular patches were additionally watermarked in the spatial domain. Rotational invariance was achieved using the movement properties of polar mapped circular patches.

Through various experiments, the simulation results show that the method of the present invention is robust against geometric deformation attack as well as signal processing attack. Therefore, consideration of local features is important for the design of a robust watermarking scheme, and it is believed that the method of the present invention will be a solution using these features.

It will be apparent to those skilled in the art that various changes and modifications can be made without departing from the technical spirit of the present invention through the above description. Therefore, the technical scope of the present invention should not be limited only to the contents described in the embodiments, but should be defined by the claims.

As described above, the watermarking method according to the present invention is capable of extracting a clear circular patch even after an image attack by using size invariant feature transform (SIFT) to perform efficient watermark synchronization, and based on polar mapping. By using an additional watermarking method, the embedded watermark is invariably detected for rotational distortion, and thus it is flexible to geometric distortion attack and signal processing attack.

Claims (18)

In digital watermarking method using Scale Invariant Feature Transform (SIFT), Generating a two-dimensional rectangular watermark along a Gaussian distribution using a random number generator; A second step of generating a circular patch using a scale invariant feature transform (SIFT); A third step of converting a rectangular watermark into a circular watermark by inverse polar mapping according to the size of the circular patch; A fourth step of calculating a recognition mask for the circular patch; And A fifth step of additionally inserting a watermark in a spatial area using the circular watermark and the recognition mask; Digital watermarking method using a size-invariant feature transformation comprising a. The method according to claim 1, The circular watermark is generated for each patch, digital watermarking method using a size invariant feature transformation (SIFT). The method according to claim 1, The rectangular watermark is mapped to the upper half of the circle, and the lower half of the circle is set symmetrically using the upper half of the circle. The method according to claim 1, And converting the rectangular watermark into a circular watermark by inverse polar mapping, using Equation 11. The digital watermarking method using a size invariant feature transform (SIFT).

Where x and y are the rectangular watermark coordinates, M is the x size of the rectangular watermark, N is the y size of the rectangular watermark, r _i and θ are the coordinates of the circular watermark, r _M is the radius of the circle r, r _o is the radius of the circle determined relatively from r _M ) The method according to claim 4, R _o in Equation 11

A digital watermarking method using a size invariant feature transform (SIFT), characterized in that the difference between r _M and r _o is greater than N. The method according to claim 1, The circular patch extraction is a digital watermarking method using a size invariant feature transform (SIFT) characterized in that the extraction using the equation (12).

(Where k is a magnification factor that controls the radius of the circular patch, t ₁ , t ₂ are the center coordinates of the circular patch, and s is the extracted scale of the circular patch.) The method according to claim 6, The scale s is a digital watermarking method using a size invariant feature transformation (SIFT), characterized in that varying from 2.0 to 10.0. The method according to claim 1, And inserting the circular watermark in a spatial manner in an additional manner using Equation 13. The digital watermarking method using a size invariant feature transform (SIFT).

(here,

Is an image with a watermark,

And

Are the pixel values of the image and the circular watermark,

Indicates a recognition mask that controls the embedding strength of the watermark.) The method according to claim 1, The recognition mask is a digital watermarking method using a size invariant feature transformation (SIFT), characterized in that using the equation (14).

Where α is the lower limit of visibility in flat and flat regions, β is the upper limit in textured regions with edges, and NVF (i, j) is a noise visibility function.) The method according to claim 9, The noise visibility function NVF (i, j) is a digital watermarking method using a size invariant feature transform (SIFT) characterized in that using the equation (15).

(here,

And

Are the local variances and maximum values of neighboring pixels in each of the five pixels, (i, j) are pixel coordinates on the image, and D is a constant value representing magnitude.) Watermarking detection in digital watermarking method using Scale Invariant Feature Transform (SIFT) Generating a circular patch using a scale invariant feature transform; A second step of detecting a watermark inserted in the circular patch by calculating a noise by applying a filter; A third step of converting the detected circular watermark into a rectangular watermark through polar mapping; And A fourth step of verifying similarity by comparing the detected watermark with the reference watermark used in the embedding process through a circular convolution; Digital watermarking method using a size-invariant feature transformation comprising a. The method according to claim 11, The circular watermark is generated for each patch, digital watermarking method using a size invariant feature transformation (SIFT). The method according to claim 11, The filter is a digital watermarking method using a size invariant feature transformation (SIFT), characterized in that any one of the Wiener filter, Inverse filter, Gaussian filter. The method according to claim 11, And polarizing the circular watermark to transform the rectangular watermark into a rectangular watermark using Equation (16).

Where x and y are the rectangular watermark coordinates, M is the x size of the rectangular watermark, N is the y size of the rectangular watermark, r _i and θ are the coordinates of the circular watermark, r _M is the radius of the circle r, r _o is the radius of the circle determined relatively from r _M ) The method according to claim 14, In equation (16) r _o is

A digital watermarking method using a size invariant feature transform (SIFT), characterized in that the difference between r _M and r _o is greater than N. The method according to claim 11, The similarity between the reference watermark and the detected watermark is represented by Equation 17. The digital watermarking method using a size invariant feature transform (SIFT).

(Where w is the reference watermark, w * is the retrieved watermark, and m and n are each pixel coordinate on the rectangular watermark.) The method according to claim 16, The similarity (similarity) is from -1.0 to 1.0 Digital watermarking method using a size invariant feature transformation (SIFT). The method according to claim 11, The watermark detection is digital watermarking using size invariant feature transform (SIFT) characterized in that it is detected by using a correlation coefficient detector.

Images