CN113689319B

CN113689319B - Local multiple watermarking method for color image

Info

Publication number: CN113689319B
Application number: CN202110894246.8A
Authority: CN
Inventors: 赵鹏; 王保卫; 周瑞祥
Original assignee: Nanjing University of Information Science and Technology
Current assignee: Nanjing University of Information Science and Technology
Priority date: 2021-08-04
Filing date: 2021-08-04
Publication date: 2023-06-16
Anticipated expiration: 2041-08-04
Also published as: CN113689319A

Abstract

The invention discloses a local multiple watermarking method for a color image, which comprises the following steps: generating a rotation invariant region for the colored original carrier image; performing RGB channel separation on the generated rotation invariant region of the original carrier image and the watermark image, and then performing Anrold scrambling; dividing the rotation invariant region after the scrambling of the Anrolid into blocks of singular values, selecting the maximum singular value to form a feature matrix, and carrying out exclusive OR operation on the feature matrix and the watermark image after the scrambling of the Anrolid to obtain a zero watermark image with copyright information of each channel; carrying out RGB channel combination on the obtained zero watermark images of the three channels to generate a color zero watermark image; generating a key for the original carrier image, and embedding the watermark image into the original carrier image by using a DWT_SVD method based on the generated key; and extracting the embedded watermark from the original carrier image after embedding the watermark. The invention can better resist rotation attack, balances the problems of local area construction scale and algorithm performance, and improves the algorithm robustness.

Description

Local multiple watermarking method for color image

Technical Field

The invention relates to a local multiple watermarking method for a color image, belonging to the technical field of information security.

Background

With the rapid development of the internet, digital works are widely spread on networks, and people can easily acquire work information from the networks. But the piracy phenomenon is serious while the flow of the works is improved, so that the ownership of authors is seriously damaged. Digital watermarking (Digital Watermaking, DW) is one of the effective means for guaranteeing the security of works and copyrights of authors currently, has good invisibility, and robustness watermarking and multi-copyright authentication watermarking are targets of watermark field research.

The digital watermark is to embed the digital information of the watermark into the digital carrier such as image, audio/video file, etc., or to modify the characteristic information or information structure of the host, which still has high practicability, i.e. it is not easy to be detected and tampered, and can be extracted and identified by the original authority.

In 1993, the watermark appears for the first time in Tirkel publication, "Electronic watermark", which he called an electronic watermark, i.e. a digital watermark of today. After that, digital watermarking technology is rapidly developed at home and abroad.

In terms of implementation, digital watermarking techniques can be classified into spatial domain-based implementations and transform domain-based implementations. Early watermarking algorithms were airspace-based implementations, and Schyndel and Tirkel proposed the least significant bit algorithm in 1994. The algorithm decomposes the pixel values of the carrier image and the watermark image into 8 bits, and the lowest bit of the carrier image is replaced by the highest bit of the watermark to realize watermark embedding. Later researches show that the algorithm based on the space domain is poor in invisibility and robustness, so that people achieve watermark embedding by utilizing an image frequency domain to obtain high invisibility and strong robustness. The transform algorithms used for frequency domain watermarking mainly include discrete fourier transform DFT (Discrete Fourier Transform), discrete cosine transform DCT (Discrete Cosine Transform) and discrete wavelet transform DWT (Discrete Wavelet Transform). Since the DFT transform can reflect only global information of a signal and cannot process abrupt and unstable signals, the correlation algorithm is few. The DCT transform has the advantage that the energy of the image can be concentrated and the computation is relatively simple. In 1995, koch proposed a digital watermarking algorithm based on a block DCT. In 1996, cox et al proposed to embed a watermark in the low frequency region of the original image, thereby achieving a more robust DCT watermarking algorithm. The biggest feature of the DWT transform is that it can analyze both the time domain and the frequency domain. In 2011, yuan Xiugui et al propose a robust watermarking algorithm based on DCT_DWT_SVD, which complements DCT transformation, DWT transformation and SVD decomposition, thereby realizing a watermarking algorithm with higher security. In 2014, xiong Xiangguang et al propose an improved dwt_svd scheme, which solves the problem of too high false alarm rate by performing multi-stage wavelet decomposition on an image and then performing singular value decomposition.

Watermark algorithms can be divided into two classes from the watermark embedding area, one class being global watermarks and the other class being local watermarks. Local watermarking constructs local regions according to local features of the image to enable embedding of the watermark. Local feature detection of an image can be classified into a corner type and a region type. The corner point type detection operator mainly comprises Moravec algorithm proposed in 1977, SUSAN operator, harris operator improved based on Moravec algorithm proposed in 1988 and the like. The region type detection operator includes a Gaussian Laplace LoG (Laplace of Gaussian) detection operator, a Hessen determinant DoH (Determinant of Hessian) detection operator, a maximum stable extremum region MSER (Maximally Stable Extremal Region) detection operator and the like. However, none of the above detection operators solves the image scale problem. The feature points or feature regions detected using the above operators may be quite different at different scales of the same image. Thus, the scale invariant feature transform SIFT (Scale Invariant Feature Transform) algorithm was proposed by Lowe 1999 and fully elucidated in the published article of 2004. The algorithm searches for extreme points in the multi-scale space, and effectively solves the problems of scale and characteristics. In 2006 Bay et al improved the SIFT algorithm, proposed accelerating robust feature SURF (Speed Up Robust Features) algorithm, improved the robustness of feature point while greatly reduced the operation time.

The watermark algorithm can be divided into single-value watermarks and multiple watermarks according to the number of watermark embedding. Multiple watermarks were first applied by Cox in the video field, he breaks the video down into frames, embedding different watermarks onto different frames of the video. Later multiple watermarking technology developed on images, 2004 Ma Yide et al proposed a multiple watermarking algorithm combining DCT coefficients with Hadamard transform, which pulled the prologue of domestic multiple watermarking. In 2007, any et al proposed a dual watermarking algorithm that combines robustness and fragile watermarking, and the extraction of the watermark by this algorithm does not require the original image. In 2008, chen Jiandong et al propose a multiple watermark algorithm based on DWT transformation, which realizes multiple copyright protection by using DWT multiple frequency regions. In 2010, zhang Jinyan embeds the gray watermark into the low frequency domain of the image and the color watermark into the high frequency domain of the image based on wavelet packet and DCT transformation, thereby realizing double watermarking. In 2011, ma Rui and the cloud utilize Hadamard transformation, wavelet transformation and singular value decomposition technology to realize an algorithm for embedding multiple watermarks by combining a space domain and a frequency domain. In 2020, yu Lijun, a multiple digital watermarking algorithm is designed based on YIQ color space by utilizing Zigzag-Logistic double iterative scrambling, and the algorithm extracts a Y component image of the image YIQ color space to perform three-level wavelet decomposition, and embeds two robust watermarks at low frequency and high frequency.

Disclosure of Invention

Aiming at the contradiction between invisibility and robustness of the current global watermarking algorithm, the anti-rotation attack is poor, and the problems that the multi-copyright and the local area construction size and watermark capacity of the local watermarking algorithm are mutually contradictory can not be solved.

The technical scheme adopted by the invention specifically solves the technical problems as follows:

a method for localized multiple watermarking of color images, comprising the steps of:

step 1, generating a rotation invariant region for a color original carrier image;

step 2, performing RGB channel separation on the generated rotation invariant region of the original carrier image and the watermark image, and performing Anrold scrambling on the rotation invariant region of each channel and the watermark image; dividing the rotation invariant region after the scrambling of the Anrolid into blocks of singular values, selecting the maximum singular value to form a feature matrix, and carrying out exclusive OR operation on the feature matrix and the watermark image after the scrambling of the Anrolid to obtain a zero watermark image with copyright information of each channel; carrying out RGB channel combination on the obtained zero watermark images of the three channels to generate a color zero watermark image;

step 3, generating a secret key for the original carrier image, and embedding the watermark image into the original carrier image by using a DWT_SVD method based on the generated secret key; and extracting the embedded watermark from the original carrier image embedded with the watermark to obtain an original watermark image.

Further, as a preferred technical solution of the present invention, in the step 1, an inscribed square area of an inscribed circle of the original carrier image is selected as the rotation invariant area.

Further, as a preferred technical solution of the present invention, the step 2 further includes storing the scrambling parameters a and b and the scrambling times t as a key1 for color zero watermark image authentication.

Further, as a preferred technical solution of the present invention, the step 2 further includes authenticating the color zero watermark image: performing exclusive OR operation on the obtained feature matrix and the zero watermark image of each channel to obtain a scrambled watermark image of each channel; and carrying out inverse Arnold transformation on the scrambled watermark images of all channels by using a key1 to obtain three-channel watermark images, and carrying out RGB channel combination to generate a color watermark image.

Further, as a preferred embodiment of the present invention, the generating a key for the original carrier image in the step 3 includes the following steps:

RGB channel separation is carried out on the original carrier image, and three channel images are obtained;

and performing SIFT operation on the three channel images to obtain feature point information of each image, removing feature points outside the rotation-invariant region, forming a feature matrix A by pixel values of the rest feature points, and storing the feature matrix A as a key 2.

Further, as a preferred technical solution of the present invention, the step 3 of embedding the watermark image into the original carrier image by using dwt_svd method based on the generated key includes the following steps:

extracting feature point coordinates of a key2, selecting a peripheral 2 x 2 region and forming a 64 x 64 local feature matrix B according to pixel values of feature points in the region; performing discrete wavelet transformation on the local feature matrix B and the watermark image respectively to obtain respective frequency bands, and selecting other frequency bands except a low frequency band as a key3 to store;

respectively carrying out singular value decomposition on low frequency bands in respective frequency bands, selecting a plurality of matrixes obtained by respective singular value decomposition as a key4 for storage, and then additively embedding singular values obtained by watermark images into singular values of watermark embedding areas; and then, carrying out inverse singular value decomposition and inverse wavelet transformation on singular values of the watermark embedding region to obtain a characteristic matrix C containing watermark information, and finally replacing the pixel value of the characteristic matrix C containing watermark information with the pixel value of the original position on the original carrier image to realize watermark embedding.

Further, as a preferred technical solution of the present invention, the extracting of the embedded watermark from the original carrier image after embedding the watermark in the step 3 includes the following steps:

extracting the position of the feature point according to the key 2;

extracting pixel values of the feature points and the feature points in the surrounding 2 x 2 area to form a 64 x 64 local feature matrix B';

discrete wavelet transformation is respectively carried out on the local feature matrix B' and the local feature matrix B of the original carrier image to obtain respective low frequency bands; after singular value decomposition is carried out on the low frequency bands obtained respectively, a singular value matrix of the watermark image is obtained by utilizing a reverse additive rule; performing inverse singular value decomposition on a singular value matrix of the watermark image in combination with a key4 to obtain a low frequency band of the watermark image after wavelet transformation;

and performing inverse wavelet transformation on the obtained low frequency band combined with the key3 after the wavelet transformation of the watermark image to obtain a raw watermark image.

By adopting the technical scheme, the invention can produce the following technical effects:

the method of the invention is mainly divided into three parts: the generation of a rotation invariant region, the realization of a zero watermarking algorithm based on Arnold scrambling and blocking singular values, and the realization of a DWT_SVD watermarking method based on a SIFT feature point matrix. The invention mainly solves the problems of the prior watermarking algorithm, cuts out a rotation invariant region for coping with the rotation attack which is difficult to process in the geometric attack, and regenerates a SIFT feature point embedded region with rotation invariance in the region, thereby being capable of better resisting the rotation attack; searching feature points by using a SIFT algorithm capable of obtaining more feature points, preparing for constructing a large-capacity local feature area, aiming at the problem of constructing a scale of the local area, applying the idea of constructing a feature matrix in a zero watermark to an embedded watermark, and better balancing the problem of constructing the scale of the local area and the performance of the algorithm; multiple watermarks are embedded by using a DWT_SVD algorithm, so that watermark images with the same size as an embedded area are embedded, and the robustness of the algorithm is improved; the solution scheme of multiple authentication is considered, the multiple watermark mode is realized by combining the embedded watermark and the zero watermark, the contradiction between invisibility and robustness is balanced, and the algorithm performance is improved.

Therefore, the method utilizes the rotation invariant region and SIFT local characteristics, mixes the DWT_SVD embedded watermark and the blocking SVD zero watermark to carry out multiple watermark embedding, realizes the balance between the invisibility and the robustness of the digital watermark, and realizes the balance between the size and the watermark capacity of the local region construction, the rotation attack resistance and the multiple authentication; meanwhile, the effective watermark can be extracted through the reverse process, so that the digital watermark embedding and extraction of the color image of the patent are realized, the anti-rotation property and the multi-copyright authentication are realized, and the copyright of the digital image can be better protected.

Drawings

Fig. 1 is a Lena image after a rotation attack in an embodiment of the present invention.

Fig. 2 is a Lena marker diagram in an embodiment of the present invention.

Fig. 3 is a diagram of Lena markers after a rotation attack in an embodiment of the present invention.

FIG. 4 is an image of a Lena original image and a narrow Arnold scrambled 67 times in an embodiment of the invention.

Fig. 5 is a schematic diagram of generating a copyright zero watermark in an embodiment of the present invention.

Fig. 6 is a schematic diagram of authentication of a copyright zero watermark in an embodiment of the present invention.

Fig. 7 is a schematic diagram of the establishment of a gaussian differential pyramid in an embodiment of the present invention.

Fig. 8 is a schematic diagram of extreme point detection in an embodiment of the present invention.

Fig. 9 is a schematic diagram of two-dimensional discrete wavelet decomposition and reconstruction in an embodiment of the present invention.

Fig. 10 is a schematic diagram of key generation in an embodiment of the present invention.

Fig. 11 is a schematic diagram of local watermark embedding in an embodiment of the present invention.

Fig. 12 is a schematic diagram of local watermark extraction in an embodiment of the invention.

Fig. 13 is an original carrier image of an experiment in an embodiment of the present invention.

Fig. 14 is a watermark image of an experiment in an embodiment of the present invention.

Fig. 15 is a watermark embedded carrier image in an embodiment of the invention.

Fig. 16 is an image of a carrier after various attacks in an embodiment of the present invention.

Fig. 17 is a watermark image extracted in an embodiment of the present invention.

Fig. 18 is a graph of the PSNR values of an attacked carrier image in an embodiment of the present invention.

Fig. 19 is a diagram showing NC value extraction by watermark after attack in the embodiment of the present invention.

Detailed Description

Embodiments of the present invention will be described below with reference to the drawings.

The invention relates to a local multiple watermarking method for a color image, which mainly comprises the following steps:

step 1, generating a rotation invariant region for a colorful original carrier image, wherein the rotation invariant region is specifically as follows:

for a common watermarking algorithm, two main methods for resisting rotation attack are: firstly, constructing a local area by utilizing characteristic points with a rotation invariance algorithm like SIFT and SURF, wherein the watermark embedded in the local area can effectively resist rotation attack; and secondly, correcting the pixel position by carrying out rotation correction on the rotated image.

The image after the rotation attack is shown in (a) - (c) in fig. 1, the Lena image has corner pixel loss during the rotation of 15 degrees, 30 degrees and 45 degrees, and the loss area is the area of four black areas and the loss area. Fig. 1 (e) - (g) are rotation corrected carrier images, but the missing pixels cannot be compensated.

The rotation attack is not only severe in that the pixel values of the image are lost, but also the in-situ pixel values are changed, resulting in missing images, and thus difficult to handle. As can be seen from fig. 1, when the rotation degree is not an integer multiple of 90 degrees, some pixels are more or less lost from the image, and the rotation degree of the image and the pixel loss area are found to have a periodicity rule through the subsequent increase of the rotation angle, that is, the rotation degree x loses the pixel value area and the rotation angle x mod 45 are the same.

For unnecessary pixel loss, the invention needs to construct a region with relative rotation unchanged for watermark embedding. From a mathematical point of view, the inscribed circle area of the image remains unchanged all the time when the image is rotated. Thus, in order to construct square embedded regions, the present invention marks the region between the Lena image inscribed circle region and the inscribed square of the inscribed circle as white, as shown in the labeled diagram of fig. 2. It is readily apparent from the rotational dynamic evolution of the marked areas of fig. 3 that the white marked areas do not appear black. Therefore, the invention verifies that the pixel value of the inscribed circle region of the image is not lost when the image is subjected to rotation attack and after rotation correction. Therefore, the rotation-invariant region determined by the present invention is the inscribed circle region of the image, and the inscribed square of the inscribed circle thereof is selected as the watermark embedding region due to the square watermark employed by the present invention.

Therefore, the rotation invariant region to be cut is an inscribed square region of the inscribed circle of the carrier image, and the side length of the inscribed square region is calculated according to Pythagorean theorem

In order to better construct an embedded region for operation, the embedded region is reduced to a square region with the side length of M/2, the region is cut off and is independently stored as an image, and a zero watermark with copyright identification is generated on the basis of the image.

And 2, generating a color zero watermark image for the original carrier image. In this embodiment, for an original carrier image with a size of m×m, a watermark image with a size of n×n, a process for generating a color zero watermark image is shown in fig. 5, and the steps are as follows:

(1) The channels are separated. And R, G, B channel separation is carried out on the rotation invariant region and the watermark image of the original carrier image generated in the step 1, so that the rotation invariant region and the watermark image of the original carrier image under each channel are obtained.

(2) And performing Anroli scrambling. And carrying out Anroli scrambling on the original carrier image rotation unchanged area and the watermark image under each channel to obtain the rotation unchanged area and the watermark image after the Anroli scrambling of each channel, wherein scrambling parameters a and b and scrambling times t are stored as a key1 and used for a subsequent color zero watermark image authentication process.

(3) Blocking singular values. Singular values are calculated in blocks of the rotation invariant region scrambled by each channel Anrolid, the size of the blocks is M/2n, and the maximum singular value is selected to form a feature matrix.

(4) The exclusive or generates a gray zero watermark image. And performing exclusive OR operation on the feature matrix obtained by the Anroli scrambling and the scrambled watermark image to obtain gray zero watermark images with copyright information of each channel.

(5) Zero watermark channel merging. And R, G, B, combining the obtained zero watermark images of the three channels to generate a color zero watermark image.

(6) Uploading to a rights issuer. And uploading the generated color zero watermark image to a corresponding copyright center, and providing credentials for subsequent copyright authentication. The purpose of watermark existence is to protect copyright, and watermark information can be obtained after zero watermark is authenticated, so that copyright protection is realized.

The Arnold scrambling transformation used in the present invention was proposed by Russian math, friedel Mi Er Arnold (Vladimir Igorevich Arnold), an algorithm also known as cat face transformation.

(x, y) as coordinates on the image, the generalized Arnold scrambling is transformed into the one shown in formula (1):

wherein (x ', y') is the position of the point (x, y) after Arnold transformation, a and b are any positive integer values, and can be saved as the key of Arnold scrambling algorithm, mod is the remainder operation, and N is the image length and width, because the Arnold scrambling algorithm is realized under the condition that the image length and width are equal. In particular, when a=b=1, it is called narrow Arnold scrambling.

The Arnold scrambling transform has a periodicity that is determined by the matrix length, N. Along with the increase of the scrambling times of the images, the images are changed from clear to fuzzy and from fuzzy to clear, and finally, the images are restored to the original images when the scrambling times reach the period value. Fig. 4 (a) shows the Lena original image and (b) shows the image obtained by scrambling 67 times in the narrow range of Arnold in the embodiment of the present invention.

Correspondingly, decryption of the Arnold algorithm is the encrypted inverse operation, and the inverse matrix is transformed, as shown in the formula (2):

the original Lena diagram of FIG. 4 (a) was obtained by performing 67 inverse narrow Arnold transformations on (b) of FIG. 4.

The singular value decomposition SVD (Singular Value Decomposition) used in the present invention means that any matrix can be decomposed into the form of orthogonal matrix multiplication with a unique form. For any matrix a of size m x n, it can be expressed as the product of m x m orthogonal matrix U, diagonal element non-negative m x n diagonal matrix S and n x n orthogonal matrix V transpose. The singular value decomposition of matrix a is represented as formula (3):

A＝USV ^T (3)

wherein the expression of S is shown as formula 4. The matrix has non-negative values only on the diagonal, the other terms are all 0, and the relationship of the values is sigma ₁ ≥σ ₂ …≥σ _m ≥0。

Due to sigma _i (i=1, 2 … m) is a numerical value that can be uniquely identified in the decomposition expression, and is therefore called a singular value. Singular values have various advantages such as stability, scale invariance, rotational invariance, etc. Therefore, the singular value decomposition technology is applied to digital watermarking, so that the robustness and the robustness of a watermarking algorithm can be improved.

Therefore, the method realizes the generation of the color zero watermark image by utilizing Arnold scrambling transformation and singular value decomposition through the steps, so that the watermark algorithm has stronger invisibility, robustness and robustness.

In addition, as shown in fig. 6, the step further includes an authentication process for the color zero watermark image, wherein the authentication process is an inverse process of the generation process, and the color zero watermark image can be obtained by performing exclusive or on the color zero watermark image instead of the watermark image and the maximum singular value matrix, and finally performing inverse Arnold scrambling, namely: performing exclusive OR operation on the obtained feature matrix and the zero watermark image of each channel to obtain a scrambled watermark image of each channel; the scrambled watermark images of all channels are subjected to inverse Arnold transformation by using a key1 to obtain three-channel watermark images, and RGB channel combination is performed to generate color watermark images, wherein the method comprises the following steps:

(1) The channels are separated. Using a rotation invariant region generated by an original carrier image, separating the obtained carrier rotation invariant region from a color zero watermark image by R, G, B channels to obtain an original carrier image rotation invariant region and a zero watermark image of each channel;

(2) And performing Anroli scrambling. Scrambling the rotation invariant region of the original carrier image under each channel by using the key1 to obtain the rotation invariant region of the original carrier image under each scrambled channel;

(3) Blocking singular values. Dividing the rotation invariant region of the original carrier image under each channel after disorder into blocks to obtain singular values, and selecting the maximum singular value to form a feature matrix;

(4) The exclusive or generates a scrambling watermark. Performing exclusive OR operation on the feature matrix formed by the maximum singular value selected by each channel and the zero watermark image under each channel separated by the channel to obtain a scrambled watermark image;

(5) Carrying out inverse Arnold transformation on the obtained scrambled watermark image by using a key1 to obtain a three-channel watermark image;

(6) Watermark channels merge. And R, G, B, combining the obtained watermark images of the three channels to generate a color watermark image.

Step 3, generating a secret key for the original carrier image, and embedding the watermark image into the original carrier image by using a DWT_SVD method based on the generated secret key; extracting the embedded watermark from the original carrier image embedded with the watermark to obtain an original watermark image, wherein the method comprises the following steps of:

first, a key is generated for an original carrier image, as shown in fig. 10, which is generated by:

(1) The color original carrier image was subjected to R, G, B channel separation to obtain three channel images. Because the three channels all adopt the same method, the three channels are collectively called a channel image for illustration;

(2) Performing SIFT operation on the three channel images to obtain characteristic point information of each image;

(3) Taking the rotation invariant region of each channel image, namely an inscribed circle as a boundary, removing the characteristic points outside the rotation invariant region, and selecting 1024 characteristic points from the rest characteristic points. If the number of the characteristic points is more than 1024, the first 1024 characteristic points are taken; and if the number of the feature points is less than 1024, zero padding is carried out on the feature points.

(4) And constructing a feature matrix A with the size of 32 x 32, and storing pixel values corresponding to the 1024 selected feature points in the feature matrix A, wherein the feature matrix A is stored as a key 2.

In the invention, the robustness of the local feature region is about that of the local feature points, and the SIFT algorithm with good scale invariance and attack resistance is the key for constructing the strong robustness region.

The Scale-invariant feature transform (Scale-invariant feature transform) algorithm adopted in the invention is supplemented and perfected by David Lowe published 1999, 2004. The feature points extracted by the SIFT algorithm have the advantages of scale invariance, rotation invariance, strong noise resistance, strong geometric attack resistance and the like, so that the robustness of the algorithm can be greatly enhanced when the SIFT algorithm is developed to the field of local watermarking. The implementation process of the SIFT algorithm can be divided into four steps: and establishing a Gaussian differential pyramid, positioning the key points, determining the direction of the key points and describing the key points.

The first step of the algorithm is to carry out smoothing processing on the image, and then convolve the image I (x, y) by utilizing a two-dimensional Gaussian function G (x, y, sigma) (formula 5) to construct a scale space L (x, y, sigma). The specific operation is as shown in formula (6):

L(x,y,σ)＝G(x,y,σ)·I(x,y)(6)

the constructed scale space is a gaussian pyramid constructed which contains the number of groups O, as shown in formula (7), the number of layers S as shown in formula (8), and the multiple k as shown in formula (9). Where M, N is the image length and width, and n is the number of images from which features are to be extracted.

O＝log ₂ (min(M,N)-3)(7)

S＝n+3(8)

The origin of the gaussian pyramid is that each set of images is decreasing in size and therefore only appears to be a pyramid in shape, with the completed gaussian pyramid being shown to the left in fig. 7. In the gaussian pyramid, the first set of first layer images is a smoothed carrier image with a scale σ=1.6 and the n-th layer has a scale k ^n-1 Sigma. For the same group, each layer differs in the scale sigma, the upper layer is k times the scale of the lower layer, while for the upper and lower groups the last third layer image of the upper group is downsampled as the image of the first layer of the next group, so that the scale of the first layer of the next group is k ⁿ Sigma. The gaussian pyramid is subtracted from the upper and lower layers of the group to obtain a gaussian differential pyramid, as shown on the right of fig. 7.

The second step of the algorithm is to detect extreme points of the upper, middle and lower layers of the same group of Gaussian differential pyramids, as shown in FIG. 8. The middle cross point needs to compare 8 pixels of the same layer and 18 pixels of the upper layer and the lower layer, and only when the value is maximum or minimum, the point is taken as an extreme point and stored for subsequent calculation. Since the detected extreme points are discrete and cannot be associated with adjacent coordinate points, the extreme points in the continuous space are detected by Talor expansion for the scale invariance of the reinforcing feature points. Taylor expansion formula at point (x 0, y0, σ) for the scale-space function D (x, y, σ) is equation (10):

wherein X is the offset of the point from the exact extreme point. Deriving the above equation, and making the result be 0, and then the accurate extreme point position offset is equation (11):

the corresponding extreme point equation is equation (12):

if all three components of x' are less than 0.5, it is assumed that the accurate extreme point has been shifted by a small amount from the current point, and it is regarded as an accurate extreme point. If any one of the components is greater than 0.5, iterating the point, the iteration times are not more than five times, and if the point is still not converged, discarding the point.

In addition, among the obtained extreme points, points of low contrast are also to be discarded. If abs (D (x) ^′ ) If T/n), the point contrast is too low and becomes unstable when disturbed by noise, thus being dropped.

Then, the watermark embedding of the DWT_SVD is realized according to the key2, as shown in FIG. 11, and the specific steps are as follows:

(1) Extracting feature point coordinates according to a key2, selecting an area containing 2 x 2 around the feature points, forming a local feature matrix with the size of 64 x 64 by using pixel values of the feature points in the selected area, and marking the local feature matrix as a local feature matrix B;

(2) Performing discrete wavelet transformation on the local feature matrix B and the watermark image respectively to obtain respective frequency bands, obtaining eight frequency bands LL, LH, HL, HH, ll, lh, hl and hh in total, selecting other frequency bands except the low frequency band in the respective frequency bands as a key3 to store, namely selecting the lh, hl and hh frequency bands outside the low frequency band as the key3 to store; the frequency bands lh, hl, hh are chosen as keys because only their low frequency band ll is embedded for greater robustness in watermark embedding, and these three components do not operate on them. The three frequency bands are closely related to the whole image, namely, the three components generated by performing discrete wavelet transform on the image embedded with the watermark are greatly different from the original image components, so that the three components are stored as keys.

(3) Singular value decomposition is respectively carried out on each low frequency band LL and LL, a matrix u and v obtained by singular value decomposition of the low frequency band LL are stored as a key4, and a singular value matrix S of the LL is embedded into a singular value S of the LL by utilizing additive embedding, so that a singular value S' of a watermark embedded area containing watermark singular value information is obtained. Performing inverse singular value decomposition on U, S ', V to obtain a frequency band LL';

(4) Performing inverse wavelet transformation on four frequency bands of LL', LH, HL and HH to obtain a characteristic matrix C containing watermark information;

(5) And replacing the pixel value of the original position on the original carrier image with the pixel value of the characteristic matrix C containing watermark information, so as to realize the embedding of the watermark.

For two-dimensional discrete wavelet transform of an image, the most common method is to separate two-dimensional scale functions and respectively perform one-dimensional discrete wavelet transform on the two-dimensional scale functions, namely:

φ(x,y)＝φ(x)φ(y)(13)

as shown in fig. 9, in the two-dimensional discrete wavelet transform process of the image according to the embodiment of the present invention, first, one-dimensional discrete wavelet transform is performed on each line of the image to obtain a low-frequency component L and a high-frequency component H of the original image in the horizontal direction, and then one-dimensional discrete wavelet transform is performed on each column of the transformed data to obtain a low-frequency component LL of the original image in the horizontal and vertical directions, a low-frequency component LH of the original image in the horizontal and vertical directions, a high-frequency component HL of the original image in the horizontal and vertical directions, and a high-frequency component HH of the original image in the horizontal and vertical directions. In addition, the watermark embedding on different components is also characterized, for example, the watermark embedding on the low-frequency component LL can enhance the robustness of the watermark, but the invisibility and imperceptibility of the human eye vision system are poor because the human eye vision system is more sensitive to the low-frequency component, and on the contrary, the invisibility and the robustness of embedding the watermark on the high-frequency component are strong. The invisibility of the present invention has been excellent so that the watermark is embedded in the low frequency component to enhance the robustness of the algorithm. The wavelet decomposition of component LL may then continue to be performed further to obtain a more robust low frequency component.

Finally, extracting the embedded watermark from the original carrier image after embedding the watermark, as shown in fig. 12, specifically comprising the following steps:

(1) Extracting the position of the feature point according to the key 2;

(2) Extracting pixel values of the feature points and the feature points in the surrounding 2 x 2 area to form a 64 x 64 local feature matrix, and marking the local feature matrix as a local feature matrix B';

(3) Performing discrete wavelet transformation on the local feature matrix B 'and the local feature matrix B of the original carrier image to obtain respective low frequency bands LL', LL;

(4) Performing singular value decomposition on each low frequency band LL', wherein LL is subjected to singular value decomposition, and a singular value matrix s of the watermark image is obtained by using a reverse additive rule;

(5) Performing inverse singular value decomposition on a singular value matrix s of the extracted watermark image in combination with a key4 to obtain a low frequency band ll of the watermark image after wavelet transformation;

(6) And carrying out inverse wavelet transformation on the obtained low frequency band ll after wavelet transformation of the watermark image in combination with the key3 to obtain the original watermark image.

In order to verify that the method can utilize the rotation invariant region, the embedded watermark is combined with the zero watermark, and the zero watermark construction feature matrix is applied to the embedded watermark, so that the balance between invisibility and robustness of a watermark algorithm is met, and an experimental example is specifically listed for illustration.

The color original carrier image and watermark image of the experiment of the invention are shown in fig. 13 and 14, the effect after embedding the watermark is shown in fig. 15, and the effects of various attacks and watermark extraction are shown in fig. 16 and 17.

Table 1 comparison of the present invention with the existing algorithm

Color space

Embedded region

Watermark type

Watermark capacity

Number of watermarks

Extraction mode

The method of the invention

RGB

Local area

Color image

64*64

Multiple ones

Non-blind extraction

DCT

RGB

Global situation

Binary image

64*64

One or more of

Blind extraction

DWT_SVD

RGB

Global situation

Color image

64*64

One or more of

Non-blind extraction

Multistage DWT

RGB

Global situation

Color image

64*64

Multiple ones

Non-blind extraction

DWT optimization

RGB

Global situation

Color image

64*64

One or more of

Non-blind extraction

Local SURF

YCbCr

Local area

Text information

1*32

Multiple ones

Non-blind extraction

TABLE 2 vector image PSNR values after challenge

PSNR(db)	Lena	Baboon	Pepper	Plane	Sailboat	Tiffany
							The method of the invention	40.2392	40.3158	40.0533	40.2843	40.4836	41.7297
DCT	36.4156	36.5684	36.3192	36.5623	36.6752	36.8267
							DWT_SVD	35.1193	35.1223	35.1236	35.1345	35.1367	35.1349
Multistage DWT	38.4165	38.4259	38.3492	38.6156	38.3214	38.5826
							DWT optimization	35.7981	35.8025	35.8193	35.8342	35.8328	35.8672
Local SURF	45.3287	40.2346	42.3516	41.1456	42.3851	42.3564

Table 3 watermark extraction NC values after attack

In contrast to the current advanced algorithm, its basic information pair is shown in table 1. In terms of invisibility, it can be seen from Table 2, FIG. 18 that the PSNR values of the method of the present invention are above 40db, and the overall curve is relatively smooth, with no significant fluctuations. Compared with other algorithms, although the local SURF algorithm is higher than the method of the invention, the whole curve is more tortuous and the stability is not enough; the other four algorithms are stable in overall curve but lower than the method of the invention. Therefore, the method of the invention not only maintains a higher PSNR value, but also ensures the stability of invisibility after watermark embedding. For the local SURF algorithm, even a PSNR value of 45 is reached, because the capacity of embedded watermarks is very small, which is only 1/128 of the capacity of the watermark of the method of the invention. Therefore, the method of the invention can be said to have stronger invisibility than the existing algorithm.

In terms of robustness, as can be seen from table 3, fig. 19, NC values of the watermark extracted by the method of the present invention under the attacks of noise, sharpening, compression, translation, shearing, etc. are all above 0.99, and the highest NC value in the comparison algorithm is still about 0.02 lower than the method of the present invention; under the rotation attack, the NC value extracted by the method is kept at about 0.985, and the height drop is the largest under all attacks from the line diagram. From the integral line graph, the NC value of the method is higher than that of a comparison algorithm under all attacks, and the curve is smoother except for shrinking the attacks. It can be said that the algorithm herein is more robust than the existing algorithm.

In summary, the method solves the problems of the prior watermarking algorithm, cuts out a rotation-invariant region for coping with the rotation attack which is difficult to process in the geometric attack, and regenerates a SIFT feature point embedded region with rotation invariance in the region, thereby being capable of better resisting the rotation attack; searching feature points by using a SIFT algorithm capable of obtaining more feature points, preparing for constructing a large-capacity local feature area, aiming at the problem of constructing a scale of the local area, applying the idea of constructing a feature matrix in a zero watermark to an embedded watermark, and better balancing the problem of constructing the scale of the local area and the performance of the algorithm; watermark embedding is carried out by using a DWT_SVD algorithm, so that watermark images with the same size as the embedded area are embedded, and the robustness of the algorithm is improved; the solution scheme of multiple authentication is considered, the multiple watermark mode is realized by combining the embedded watermark and the zero watermark, the contradiction between invisibility and robustness is balanced, and the algorithm performance is improved.

The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for localized multiple watermarking of color images, comprising the steps of:

step 2, respectively carrying out RGB channel separation on the rotation invariant region of the generated original carrier image and the watermark image, and carrying out Anrold scrambling on the rotation invariant region of each channel and the watermark image; dividing the rotation invariant region after the scrambling of the Anrolid into blocks of singular values, selecting the maximum singular value to form a feature matrix, and carrying out exclusive OR operation on the feature matrix and the watermark image after the scrambling of the Anrolid to obtain a zero watermark image with copyright information of each channel; carrying out RGB channel combination on the obtained zero watermark images of the three channels to generate a color zero watermark image;

step 3, generating a secret key for the original carrier image, and embedding the watermark image into the original carrier image by using a DWT_SVD method based on the generated secret key, wherein the method comprises the following steps:

performing SIFT operation on the three channel images to obtain feature point information of each image, removing feature points outside the rotation-invariant region, forming a feature matrix A by pixel values of the rest feature points, and storing the feature matrix A as a key 2;

respectively carrying out singular value decomposition on low frequency bands in respective frequency bands, selecting a plurality of matrixes obtained by respective singular value decomposition as a key4 for storage, and then additively embedding singular values obtained by watermark images into singular values of watermark embedding areas;

then, carrying out inverse singular value decomposition and inverse wavelet transformation on singular values of a watermark embedding region to obtain a characteristic matrix C containing watermark information, and finally replacing pixel values of the characteristic matrix C containing watermark information with pixel values of an original position on an original carrier image to realize watermark embedding;

and extracting the embedded watermark from the original carrier image embedded with the watermark to obtain an original watermark image.

2. The method according to claim 1, wherein the step 1 selects an inscribed square region of the inscribed circle of the original carrier image as the rotation invariant region.

3. The method according to claim 1, wherein the step 2 further comprises storing the scrambling parameters a and b and the scrambling times t as a key1 for color zero watermark image authentication.

4. The method of partial multiple watermarking of color images according to claim 2, wherein step 2 further includes authenticating the color zero watermark image: performing exclusive OR operation on the obtained feature matrix and the zero watermark image of each channel to obtain a scrambled watermark image of each channel; and carrying out inverse Arnold transformation on the scrambled watermark images of all channels by using a key1 to obtain three-channel watermark images, and carrying out RGB channel combination to generate a color watermark image.

5. The method for locally multiplexing color image watermarking according to claim 1, wherein the extracting of the embedded watermark from the original carrier image after embedding the watermark in step 3 includes the steps of:

extracting the position of the feature point according to the key 2;