Low false alarm zero watermark method based on region-based XOR

Abstract

The invention discloses a low false alarm zero watermark method based on regional XOR, which expands the difference between similar features to solve the false alarm problem of image zero watermark through redundantly constructed features and watermarks, three-value quantization, and feature index regional XOR, thereby realizing copyright protection of digital information. Two feature extraction methods are used to extract the frequency domain features and spatial domain features of the image, and the frequency domain features are constructed into an index matrix. The redundantly constructed copyright watermark is divided into two regions according to the index matrix, and the redundantly constructed spatial domain features and the chaotic matrix are XORed respectively. The zero watermark is obtained after the regions are merged. In the copyright authentication process, the same feature extraction and regional XOR method is used to obtain the redundantly constructed copyright watermark image, which is quantized into a three-value matrix after de-redundancy operation, so as to realize copyright authentication of distinguishable similar images.

Description

Technical Field

[0001] The invention belongs to the technical fields of digital image processing, information security and copyright protection, and in particular relates to a low false alarm zero watermark method based on region-based XOR. Background Art

[0002] With the development of the information age, various forms of digital works can be easily obtained on the Internet, and copyright issues have also emerged. Digital watermarking technology, which embeds watermark information into carriers to achieve copyright protection, has received widespread attention in recent years. Embedding information in images will inevitably affect the image quality, and there is also a contradiction between robustness and invisibility. The proposal of zero watermark [Wenquan, Sun Fengfeng, Wang Shuxun. Concept and application of zero watermark [J]. Journal of Electronics. 2003, 31(02): 214-216.] solves this problem. It extracts a unique and robust feature sequence from the image and directly uses it as a zero watermark or establishes a connection with the copyright mark as a zero watermark, saves it to the certification center, and completes the registration of image copyright. The process of zero watermark generation can be divided into two parts: 1) image feature extraction, 2) establishing a connection between features and copyright (such as generating a zero watermark by XORing features with watermarks). Evaluation indicators of the zero watermark algorithm, such as robustness, operating efficiency and false alarm rate, are mainly affected by the image feature extraction part. After extracting features from the image, a connection is established with the copyright logo to generate a zero watermark. During the copyright authentication process, the copyright logo can be obtained from the zero watermark and image features to achieve copyright visualization. Usually, an encryption algorithm is added in the process of establishing a connection between features and copyright to enhance the security of the zero watermark.

[0003] In the current research on image zero watermarking, the methods of image feature extraction can be mainly divided into the following types. Based on frequency domain transformation, such as shift-invariant shearlet transform (SIST) [Shi S, Luo T, Huang J, et al. A Novel HDR Image Zero-Watermarking Based on Shift-InvariantShearletTransform[J]. Security and Communication Networks. 2021, 2021(8655): 1-12.], Gabor transform [Fan D, Li Y, Gao S, et al. A novel zero watermark optimization algorithm based on Gabor transform and discrete cosine transform[J]. Concurrency and Computation: Practice and Experience. 2020, n / a(n / a): e5689.], cellular automata transform (CAT) [Tzuo-Yau F, Her-Chang C, Bin-Chang C. Lossless medical image watermarking method based on significantdifference of cellular automata transform coefficient[J]. Signal Processing: Image Communication.2018,70], dual-tree complex wavelet transform (DTCWT) [Jing L, Jingbing L, Jixin M, et al. A Robust Multi-Watermarking Algorithm for Medical Images Based on DTCWT-DCT and Henon Map [J]. Applied Sciences.2019,9(4).], contourlet transform (contourlet transform, CT) [Xiaoqi W, Jingbing L, Rong T, et al.Contourlet-DCT based multiple robust watermarkings for medical images[J].Multimedia Tools and Applications.2019,78(7).], non-subsampled contourlet transform[Ayesha S,VMZero-Watermarking in Transform Domain and QuadtreeDecomposition for Under Water Images Captured by Robot[J].Procedia Computer Science.2018,133.], Curvelet transform[Wu Deyang, Zhao Jing, Wang Guoping, et al. An image zero watermarking technology based on improved singular value and sub-block mapping[J]. Acta Optica Sinica.2020,40(20):85-97.], etc. After obtaining the stable coefficient, it can be directly constructed as an image feature, or the coefficient can be further subjected to discrete cosine transform (DCT), singular value decomposition (SVD), improved singular value decomposition and other operations to extract its principal component construction features. Use polar harmonic transforms to calculate the invariant moment features of images, such as decimal-order polar harmonic transforms (DoPHTs) [Xia Z, Wang X, Han B, et al. Color image triple zero-watermarking using decimal-order polarharmonic transforms and chaotic system [J]. Signal Processing], quaternion polar harmonic transforms (quaternion polar harmonic transforms, QPHTs) [Zhiqiu X, Xingyuan W, Wenjie Z, et al. Color medical image lossless watermarking using chaotic system and accurate quaternion polar harmonic transforms [J]. Signal Processing. 2019, 157.], quaternion radial fractional Charlier moments (Quaternion Radial Fractional Charlier Moments, QRFrCMs) [MY, HK, MS, et al.Robust zero-watermarking scheme based on novelquaternion radial fractional Charlier moments[J].Multimedia Tools andApplications.2021(prepublish).], quaternion polar harmonic Fourier moments (QPHFMs)[Zhiqiu X,Xingyuan W,Xiaoxiao L,et al.Efficientcopyright protection for three CT images based on quaternion polar harmonicFourier moments[J].Signal Processing.2019,164.] etc. Extract features in the spatial domain, construct the stable relationship between the sub-block mean and the overall mean into image features [Xiong Xiangguang. Spatial domain strong robust zero watermarking scheme [J]. Acta Automatica Sinica. 2018, 44 (01): 160-175.], and build a neural network model to extract image robust features [Fierro-Radilla A, Nakano-Miyatake M, Cedillo-Hernandez M, et al. A Robust Image Zero-watermarking using Convolutional Neural Networks [C]. In: 2019 7th International Workshop on Biometrics and Forensics (IWBF). 2019.]. .

[0004] In the above-mentioned zero watermark research, most of them focus on improving the stability of image features. When the number of images to be protected is large or the similarity between images is high, false alarm problems will occur, and unprotected images will be mistakenly judged as protected images. Existing studies mostly solve the false alarm problem of similar images from the perspective of feature extraction, extracting robust and highly unique features from images to generate zero watermarks. The literature [Wang W, Li Y, Liu SA Polar Complex Exponential Transform-Based Zero-Watermarking for Multiple Medical Images with High Discrimination [J]. Security and Communication Networks. 2021, 2021 (2): 1-13.] studied the similarity distinction problem of zero watermarks for multiple medical images, constructed features based on the stability of the same-order amplitude and the same-repetition amplitude relationship, and specified a reference value for all amplitudes to reduce the similarity of similar image features. In the false alarm experiment of 10 cancer medical images, the average bit error rate and the minimum bit error rate were 0.3631 and 0.2629, respectively. Reference [Xiaobing K, Fan Z, Yajun C, et al. Combining polar harmonic transforms and 2D compound chaotic map for distinguishable and robust color image zero-watermarking algorithm [J]. Journal of Visual Communication and Image Representation. 2020, 70 (C).] simultaneously calculates the three polar harmonic transforms (PHTs) of the image, selects the precise moment to enhance the robustness, constructs a binary feature sequence based on the relationship between adjacent moment amplitudes to generate a zero watermark, and performs a false alarm test on the baboon image in the USC-SIPI dataset and other similar baboon images. The bit error rate between the zero watermarks generated by similar images is about 0.5.The literature [Beiji Z, Jingyu D, Xiyao L, et al. Distinguishable zero-watermarking scheme with similarity-based retrieval for digital rights management of fundus image [J]. Multimedia Tools and Applications. 2018, 77 (21).] divides the circular area of ​​the fundus medical image into multiple fan-shaped sub-blocks, calculates the difference between adjacent sub-blocks, quantizes them into four-value feature vectors as image features, and establishes a feature library. During the copyright authentication process, a similarity search is first performed in the feature library to distinguish between protected images and similar images. The literature [Liu X, Lou J, Wang Y, et al. Discriminative and robust zero-watermarking scheme based on completed local binary pattern for authentication and copyright identification of medical images [C]. In: Imaging Informatics for Healthcare, Research, & Applications. 2018.] first uses the complete local binary pattern (CLBP) to construct three CLBP symbol images and three CLBP amplitude images, which are divided into multiple concentric ring areas. The mean, variance, skewness, and kurtosis of the area are calculated and used as image features after binarization. In the similar image false alarm test, the average and minimum bit error rates are 0.4636 and 0.1111, respectively. In the literature [LiuX, Sun Y, Wang J, et al. A novel zero-watermarking scheme with enhanced distinguishability and robustness for volumetric medical imaging [J]. SignalProcessing: Image Communication. 2021, 92 (prepublish). The image is divided into multiple concentric ring regions, and the mean, variance, skewness, and kurtosis information residual features of each region are calculated. After binarization, the four features in the region are constructed into the features of one region. In the similar image false alarm test, the average and minimum bit error rates are about 0.4 and 0.2, respectively.The literature [Wang W, Li Y, Liu SA Polar Complex Exponential Transform-Based Zero-Watermarking for Multiple Medical Images with High Discrimination[J]. Security and Communication Networks. 2021, 2021(2): 1-13.] and the literature [Beiji Z, Jingyu D, Xiyao L, et al. Distinguishable zero-watermarking scheme with similarity-based retrieval for digital rights Management of Fundus Image[J]. Multimedia Tools and Applications. 2018, 77(21).] respectively solve the false alarm problem of zero watermarks of similar images from the perspectives of invariant moment features and circular block feature construction. Although the zero watermarks generated by similar images can reach a distinguishable level, they are both based on a fixed feature extraction algorithm and for a fixed type of image, with poor scalability, and there is still room for improvement in the distinguishability of zero watermarks of similar images.

[0005] After extracting the image features, the image features are linked to the watermark image through logical operations, visual cryptography, neural network training and other methods to obtain a zero watermark, which is then saved to the authentication center. As an information hiding algorithm, although the zero watermark does not embed the watermark information into the carrier, the zero watermark generated by it is a form of expression for establishing a connection between the carrier and the copyright logo. If the generated zero watermark has low security, when the zero watermark generation algorithm is made public, attackers can analyze the corresponding relationship between the protected electronic carrier and the copyright owner. In special application environments such as military and medical, this corresponding relationship, as a piece of information, has an inestimable value, so the zero watermark needs to have a certain degree of security. Usually copyright watermarks and image features are binary matrices, so logical operations (such as XOR) have the advantages of low complexity and reversibility, and have been widely used in zero watermark algorithms, such as [Shi S, Luo T, Huang J, et al. A Novel HDR Image Zero-Watermarking Based on Shift-Invariant Shearlet Transform [J]. Security and Communication Networks. 2021, 2021 (8655): 1-12, Fan D, Li Y, Gao S, et al. A novel zero watermark optimization algorithm based on Gabor transform and discrete cosine transform [J]. Concurrency and Computation: Practice and Experience. 2020, n / a (n / a): e5689.5-13, Xiong Xiangguang. Spatial robust zero watermark scheme [J]. Acta Automatica Sinica. 2018, 44 (01): 160-175, Fierro-Radilla A, Nakano-Miyatake M, Cedillo-HernandezM, et al. A Robust Image Zero-watermarking using Convolutional Neural Networks [C]. In: 2019 7th International Workshop on Biometrics and Forensics (IWBF). 2019.], and in order to improve security, most encryption algorithms are added in this step; visual cryptography, as a secret-sharing encryption scheme, is also used to construct zero watermarks, such as in the literature [Beiji Z, Jingyu D, Xiyao L, et al.Distinguishable zero-watermarking scheme with similarity-based retrieval for digital rights Management of Fundus Image[J].Multimedia Tools andApplications.2018,77(21), Liu X, Lou J, Wang Y, et al. Discriminative and robust zero-watermarking scheme based on completed local binary pattern for authentication and copyright identification of medical images[C]. In: ImagingInformatics for Healthcare,Research,&Applications.2018,Liu X,Sun Y,Wang J,etal.A novel zero-watermarking scheme with enhanced distinguishability and robustness for volumetric medical imaging[J].Signal Processing: Image Communication. 2021, 92 (prepublish), Qu Changbo, Yang Xiaotao, Yuan Duoning. Wavelet domain visual cryptographic zero watermark algorithm [J]. Journal of Image and Graphics. 2014, 19 (03): 365-372.] This method itself has high security and generally does not require the addition of additional encryption algorithms. Therefore, it can improve the operating efficiency of the zero watermark algorithm to a certain extent, and the image superposition can display the secret information, which has the characteristics of simple authentication; neural network-based methods, such as the literature [Wang X, Ma D, Hu K, et al. Mapping based Residual Convolution Neural Network for Non-embedding and Blind Image Watermarking [J]. Journal of Information Security and Applications. 2021, 59.], after extracting features from the image, the features are used as the input of the residual convolutional neural network (RCNN), the copyright watermark image is used as the output, the network model is trained, and the trained model parameters are saved as zero watermarks to the authentication center. Since the parameters of the neural network are used as zero watermarks, it is difficult for attackers to analyze the copyright watermark or carrier features from the parameters, and this method has high security. Among the above three methods, the logical operation needs to be supplemented with an encryption algorithm, which may reduce the efficiency of the algorithm while improving security; the visual password will expand the copyright watermark pixels and increase the amount of zero watermark data; the neural network method needs to train the network parameters when generating the zero watermark of each carrier image, so it has the limitation of the number of parameters and the complexity of the network model. . Summary of the invention

[0006] Based on the above deficiencies in the prior art, the technical problem solved by the present invention is to provide a low false alarm zero watermark method based on regional XOR, which can solve the false alarm problem of similar images while ensuring a certain robustness, and can be applied to the zero watermark algorithm of various feature extraction methods.

[0007] The low false alarm zero watermark method based on region-based XOR of the present invention includes copyright watermark embedding and copyright authentication:

[0008] Zero watermark construction includes the following steps:

[0009] 1.1 Extracting image feature sequence: The brightness component of the color image of size 512×512×3 is divided into non-overlapping blocks. To increase the stability of the feature, the size of each sub-block is set to 16×16. The feature is constructed based on the size relationship between the sub-block mean and the overall mean of the brightness component, and the spatial domain feature f of size 32×32 is obtained. SP , then extract the frequency domain features of the image, reduce the dimension of the color image features, extract the brightness component into non-overlapping blocks, each sub-block size is 2×2, calculate the sub-block mean to construct a matrix of size 256×256, use discrete cosine transform to extract the frequency domain features of the image, select a coefficient matrix of size 4×8 in the low-frequency part, and convert it into a 32-bit binary sequence f according to the positive and negative coefficients. FR As frequency domain features;

[0010] 1.2 Redundant construction of carrier features: The feature f of size 32×32 SP Redundancy is constructed into a feature F of size 128×128 SP , according to the 32-bit signature sequence f FR Redundant Structure F FR ;

[0011] 1.3 Redundant construction of copyright watermark: Divide the 32×32 copyright logo w into non-overlapping blocks, with a sub-block size of 32 / p×32 / q, i∈{1,2,3,...,128}, generate a random number matrix L of size (4×p)×(4×q), quantize the values ​​into p*q types, with the same number of each type of value, save them as the sub-block arrangement key K1, and arrange the sub-blocks redundantly;

[0012] 1.4 Regional XOR of feature index: The frequency domain feature F after redundant construction FR As the index matrix, the redundant copyright logo W is divided into two regions, which are respectively connected with the redundant constructed spatial domain features F SP XOR with the randomly generated binary matrix key K2 to obtain a zero watermark;

[0013] Copyright certification includes the following steps:

[0014] 2.1 Extract spatial features F' of size 128×128 from the image to be verified for copyright SP and frequency domain feature F' FR , select the pixel points in the zero watermark Z according to the frequency domain characteristics, and compare them with the spatial domain features F' SP XOR with key K2 to obtain a redundant copyright watermark.

[0015] Optionally, it also includes watermark de-redundancy and ternary quantization;

[0016] Watermark redundancy removal: Redundant copyright watermarks are not overlapped in blocks, the sub-block size is 32 / p×32 / q, and the sub-block is marked as b' m,n , where m∈{1,2,3,...,4×p}, n∈{1,2,3,...,4×q}, and the key K1 is de-redundant according to the arrangement.

[0017] Furthermore, when the quantization thresholds are s1=1 and s2=15, the redundant copyright watermark W' is obtained, and after de-redundant ternary quantization, the final ternary matrix w is obtained. final .

[0018] From the above, the present invention, based on the existing feature extraction algorithm, expands the difference between similar features through redundant construction of features and watermarks, three-value quantization, and feature index regional XOR to solve the false alarm problem of image zero watermark, thereby being able to achieve copyright protection of digital information. Two feature extraction methods are used to extract the frequency domain features and spatial domain features of the image, and the frequency domain features are constructed into an index matrix. According to the index matrix, the redundantly constructed copyright watermark is divided into two regions, which are XORed with the redundantly constructed spatial domain features and the chaotic matrix respectively. After the regions are merged, a zero watermark is obtained. In the copyright authentication process, the same feature extraction and regional XOR method is used to obtain a redundantly constructed copyright watermark image, which is quantized into a three-value matrix after de-redundancy operation to achieve copyright authentication of distinguishable similar images. The beneficial effects of the low false alarm zero watermark method based on regional XOR of the present invention are as follows:

[0019] (1) Based on the existing feature extraction algorithm, a feature-sensitive zero watermark is constructed, which provides a new solution to the problem of false alarm of zero watermark in images.

[0020] (2) A feature index region-based XOR zero watermark algorithm is proposed. In the process of establishing a connection between features and copyright, the carrier features are integrated to achieve multi-feature fusion to reduce the false alarm rate of image zero watermark.

[0021] (3) Redundantly construct image features and copyright watermarks, and in the copyright authentication process, the extracted copyright watermarks are de-redundant and quantized into a three-value matrix. When unprotected feature-similar images are used for copyright authentication, the copyright information cannot be extracted.

[0022] (4) Based on two existing feature extraction algorithms, a low false alarm zero watermark algorithm is proposed, which provides a new solution to the image zero watermark false alarm problem, and the feature extraction algorithm used can be replaced arbitrarily.

[0023] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the following is a detailed description in combination with the preferred embodiments and the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solution of the embodiment of the present invention, the drawings of the embodiment are briefly introduced below.

[0025] Figure 1 is the value distribution histogram;

[0026] Figure 2 This is a diagram of the zero watermark construction process;

[0027] Figure 3This is a diagram of the copyright certification process. DETAILED DESCRIPTION

[0028] The following is a detailed description of the specific implementation of the present invention in conjunction with the accompanying drawings, which is a part of this specification and illustrates the principles of the present invention through examples. Other aspects, features and advantages of the present invention will become clear from the detailed description. In the drawings referred to, the same or similar parts in different figures are represented by the same reference numerals.

[0029] like Figures 1 to 3 As shown in the figure, in the zero watermark generation process, the image feature matrix and copyright watermark are first redundantly constructed, and then a 32-bit robust binary sequence is extracted from the image and constructed into a regional index matrix. The redundant watermark image is divided into regions according to the index matrix, and the redundant features and the chaotic matrix are XORed respectively. The zero watermark is obtained after the regions are merged. In the copyright authentication process, the feature matrix and the 32-bit robust binary sequence are extracted from the image, and the redundantly constructed watermark image is extracted using the same method as the zero watermark generation process. After the redundancy operation, it is quantized into a matrix with only three values ​​as the extraction result. The experimental results show that the algorithm can solve the false alarm problem of similar images under the premise of ensuring a certain robustness, and can be applied to the zero watermark algorithm of various feature extraction methods.

[0030] The technical solution adopted by the present invention is:

[0031] 1. Analyze the causes of false alarm problems of zero watermark

[0032] Usually in the zero watermark algorithm based on image feature extraction, the robust features extracted from the image uniquely represent an image, so the false alarm problem is mainly caused by the lack of uniqueness of the features. According to whether the carrier images are similar, it can be divided into two categories: 1) False alarm problem between similar images: When the similarity of two images is high, the features extracted from the two images have a high similarity, which makes the zero watermarks generated by the two images have a high similarity, which will cause false alarm problems. 2) False alarm problem between dissimilar images: In the zero watermark algorithm, binary matrices are usually extracted from grayscale images with 8 bit depths and three-channel color images as features to construct zero watermarks. At the same time, the size of the feature matrix is ​​smaller than the original image, so the diversity of image features must be much smaller than the diversity of the original image. As the number of images increases, the feature similarity between all images is greater than the original image similarity. There may be images with obvious visual differences but high feature similarity, which will also cause false alarm problems.

[0033] 2. Expanding the differences in carrier characteristics

[0034] In the zero watermark algorithm, when the similarity between image features is high, false alarm problems will occur. By expanding the differences between different image features, false alarm problems can be effectively avoided, such as redundant construction of features and watermarks, three-value quantization, feature index region XOR and other operations, to expand the differences between similar features, thereby solving the zero watermark false alarm problem.

[0035] 2.1 Redundant Construction Carrier Features

[0036] Extract the binary feature matrix f of size m1×m2 from the color image FEA , when a false alarm problem occurs, it is usually manifested as a high similarity between the features of the two images. There is another image feature matrix f SIM , and f FEA The bit error rate B f and the number of error bits E f Calculate as formula (1) (2).

[0037]

[0038] f FEA and f SIM Redundantly constructed into F of size m1×m2×R FEA and F SIM , bit error rate B F and the number of error bits E F Calculate as follows.

[0039]

[0040] Because F FEA and F SIM It is f FEA and f SIM The redundant construction is obtained, so

[0041] F FEA (i,j,1)=F FEA (i,j,2)=...=F FEA (i,j,R)=f FEA (i,j) (5)

[0042] F SIM (i,j,1)=F SIM (i,j,2)=...=F SIM (i,j,R)=f SIM (i,j) (6)

[0043] Where i∈{1,2,3,...,m1}, j∈{1,2,3,...,m2}, we can get the bit error rate relationship and the error bit number relationship before and after the feature redundancy construction, as shown in equations (7) and (8).

[0044]

[0045] Although feature redundancy construction cannot affect the bit error rate between features, it can increase the number of error bits between different features, satisfying E F =R×E f , R is the redundancy.

[0046] 2.2 Redundant watermark construction

[0047] The binary watermark M with a size of m3×m4 w Redundancy is constructed into a redundant binary watermark M of m3×m4×R' ww , in the redundant watermark M ww Modify the binary data at random positions in , to simulate the process of extracting copyright information from a zero watermark generated by an image using an irrelevant image, the number of modifications is β×m3×m4×R', where β∈[0,1] represents the redundant watermark M ww The degree of modification after copyright certification, matrix M ww The probability of each bit being modified can be calculated by formula (9).

[0048]

[0049] in Represents the element M in the matrix ww The probability that (i,j,k) is modified, i∈{1,2,3,...,m1}, j∈{1,2,3,...,m2}, k∈{1,2,3,...,R'} is the matrix M ww In the subscript index, R' is the redundancy of the redundantly constructed watermark. ww After the value in is modified, we get M' ww The redundant watermark of copyright authentication is removed by the method of formula (10), and the matrix M' is obtained w , where the value range is {0,1,2,3,...,R'}.

[0050]

[0051] In the matrix M' w The change in the value of each element Expectations Calculated by formula (11).

[0052]

[0053] Represents the matrix M' wAn element in, in formula (11), i∈{1,2,3,m.1.}., j∈{1,2,3,...,m2} is a matrix M' ww and M' w Because β∈[0,1], as the redundancy R' increases, M' w The expected change in the element It also increases accordingly, that is, the redundancy matrix M' w With M w The difference between ×R' becomes larger.

[0054] 2.3 De-redundant watermark ternary quantization

[0055] In order to expand the difference between the features of different images, the three-value quantization processing matrix M' is used in the zero watermark copyright authentication process. w , as shown in formula (12).

[0056]

[0057] Where R' is the redundancy of the redundant watermark, i∈{1,2,3,...,m3}, j∈{1,2,3,...,m4}, α∈[0,1] is the three-value quantization parameter, (α×R') and (1-α)×R' are the quantization thresholds. Matrix M' w In the example, let the change of each value be

[0058]

[0059] That is, from formula (12) (13) we can get

[0060]

[0061] From formula (14), we can see that when the matrix M' w The change in the value When it is greater than or equal to the three-value quantization threshold (α×R'), the matrix The corresponding position in is marked as 0.5. When it is less than the three-value quantization threshold (α×R'), The value at the corresponding position in will be marked as the original binary watermark M w The value of the modified matrix M ww The values ​​at random positions in the matrix M' w The amount of change in each value in Not exactly the same, the test is carried out with parameter β = 0.25, the binary watermark image is redundantly constructed, the random position value is modified and de-redundant, and the change amount of all elements in the matrix is ​​calculated Its numerical distribution is as follows Figure 1 shown.

[0062] Change amount of elements in the matrix It basically satisfies the normal distribution, so according to formula (11), the change of the element is expected to be It can be seen that there is a certain relationship between the parameter β that represents the degree of difference between the copyright watermark before and after authentication and the three-value quantization parameter α. For example, when α=β=0.25, the three-value matrix About half of the values ​​are 0.5, and the three-value matrix can be affected by adjusting α and β. The number of 0.5 values ​​in the three-value matrix The values ​​0 and 1 in the watermark image M mainly come from w , the value 0.5 is represented by a gray dot. When a small number of values ​​in the matrix are 0.5, due to the visual characteristics of the human eye, copyright information can still be identified from it. When a large number of values ​​in the matrix are 0.5, the three-value image There are a lot of gray pixels in the image, and the copyright information cannot be identified.

[0063] Since the copyright authentication result is a three-value matrix, it is difficult to use the bit error rate and NC value to represent the difference between it and the original copyright image. After the three-value matrix is ​​displayed as an image, there are only three colors in the image: black, white and gray. Compared with the binary copyright image, gray pixels are added. The larger the number of gray pixels, the greater the difference from the binary copyright image. The difference between the three-value copyright image and the binary copyright image can be represented by calculating the percentage γ of the elements with a value of 0.5 in the three-value matrix, as shown in formula (15).

[0064]

[0065] in, is the ternary matrix obtained after copyright authentication. The count(·) function calculates the number of pixels that meet the conditions. i∈{1,2,...,32} and j∈{1,2,...,32} are the element indices in the matrix. The percentage parameter γ takes a value between 0 and 1. When the gray pixels in the ternary image are between 12.5% ​​and 25%, the information of the copyright image can be seen. When the gray pixels in the ternary image are 75%, no valid information can be seen in the image. When the gray pixels in the ternary image are 50%, the copyright information can be vaguely seen, but the damage is serious.

[0066] When the parameter β is determined, the parameter α can be adjusted to remove the redundant matrix M' w The different numbers of values ​​in are quantized to 0.5; when the parameter α is determined, different three-value images can be obtained according to the difference degree β between the two image features to distinguish the copyright authentication results of the attacked image and the false alarm image, avoiding the zero watermark false alarm problem caused by the similarity of image features.

[0067] 2.4 Regional XOR of Feature Index

[0068] In order to improve the security of the zero watermark algorithm, reduce the false alarm rate of the zero watermark algorithm, and visualize the copyright authentication information, a feature indexed regional XOR algorithm is proposed to extract binary features from the image and reconstruct it into a matrix F' of size M×N. FEA , and at the same time the binary watermark M w Redundancy is constructed into a watermark M of size M×N ww , according to the matrix F' FEA Select watermark M for the value in the middle ww The pixel in is XORed with two binary matrices to obtain a zero watermark Z, as shown in the formula.

[0069]

[0070] In the copyright authentication process, a matrix F of size M×N is extracted from the image" FEA , according to the matrix F' FEA The pixel in the zero watermark Z is XORed with two binary matrices to obtain the copyright authentication image M' ww , as shown in formula (17).

[0071]

[0072] In formula (17), i∈{1,2,3,...,M}, j∈{1,2,3,...,N} are the element position indices in the matrix, M1 and M2 are both randomly generated binary matrices of size M×N. When using images with features similar to those that cause false alarms for copyright authentication, extract the features similar to those in the matrix F' from the false alarm images. FEA The feature matrix F" with higher similarity FEA , matrix F' FEA It can be viewed as the matrix F' FEA Modify some elements to get the difference matrix D of the feature matrix F and difference rate D F , and the difference matrix D of random matrices M As shown in the following formula.

[0073] D F =xor(F' FEA ,F” FEA ) (18)

[0074]

[0075] D M =xor(M1,M2) (20)

[0076] D F and DM is a binary matrix, representing the difference between the feature difference and the random matrix. F and D M A point in the equation is 0, indicating that there is no difference at that point. F and D M A certain point in is 1, indicating that there is a difference at this point. In the copyright authentication process of formula (16), when M1(i,j)=M2(i,j), It can be deduced from formula (16) that when there is no difference in the features at a certain point, that is, D F When (i, j) = 0, the M' obtained by using formula (16) is ww (i,j)=M ww (i,j); when there is a difference in the features at a certain point, that is, D F When (i, j) = 1, M' calculated according to formula (16) ww (i,j) may be wrong, for the random matrix difference D M We discuss two different cases and derive the correct M based on equation (16). ww (i,j) calculation method.

[0077]

[0078] From formula (21), we can see that in feature F” FEA In the case of (i,j) error, the copyright certification obtains M' ww (i,j) is not all wrong, when D M (i,j)=0, that is, when there is no difference between two random matrices, M' ww (i,j)=M ww (i,j), that is, M' ww (i,j) has no error; when D M (i,j)=1, that is, when there is a difference in a certain point between the two random matrices, That is M' ww (i,j) is wrong. It can be concluded that when D F (i,j) and D M (i,j) are both 1, the copyright authentication result is M' ww (i,j) is wrong. Since M1 and M2 are randomly generated binary matrices, Indicates the occurrence of event D M (i,j) = 1, so the probability that the values ​​of an element in M1 and M2 are different is Let the feature matrix F' FEA and F” FEA The bit error rate is μ, Indicates the occurrence of event D F(i,j)=1, feature index matrix F' FEA and F” FEA The probability that an element has a different value is It can be concluded that when the watermark image is XORed into two regions, the probability of numerical error in the copyright authentication watermark is That is, the bit error rate of the copyright authentication watermark is about 0.5×μ. In the application of the zero watermark algorithm, the watermark can be divided into multiple regions, corresponding to a certain number of random matrices. The generation and authentication process of the zero watermark is shown in equations (22) and (23).

[0079]

[0080] Where h is the number of regions the watermark is divided into, and the bit error rate of the watermark obtained by copyright authentication is approximately As h increases, the bit error rate of the copyright watermark approaches the bit error rate μ between the features. Since it can be expanded to multiple random matrices M h , so the chaotic binary matrix generated by multiple keys can be used as M h , distribute the key to multiple people to achieve multi-key sharing; the redundant carrier features can also be used as the matrix M h , to amplify the differences between similar features.

[0081] Extract features from the image using two feature extraction algorithms, and use the two extracted carrier features as index matrices F' FEA The matrix M1 and the chaotic binary matrix generated by the key are used as M2. In the copyright authentication process, the extracted matrix F" FEA and M'1 may be related to the matrix F' FEA There is a certain deviation from the matrix M1, as previously proved, when the characteristic matrix F' FEA and F” FEA When the bit error rate of the feature matrix M1 and M'1 is μ', and the matrix F' FEA =F” FEA , index matrix F' FEA When the number of binary values ​​is balanced, the bit error rate of the copyright authentication watermark is about 0.5×μ'.

[0082] The false alarm problem of zero watermark is mainly manifested in the high similarity of image features. The feature index region XOR algorithm proposed in this invention can construct two features of the image into zero watermark. When the uniqueness of a certain feature is insufficient, resulting in a false alarm problem, the uniqueness of another feature can be used to improve the copyright authentication watermark bit error rate value. Combined with the three-value quantization algorithm in Section 2.3, the false alarm problem can be avoided. False alarm problems will only occur when all features of the two images are highly similar, and can be avoided by increasing the types of features.

[0083] The low false alarm zero watermark method based on regional XOR of the present invention includes a zero watermark construction process and a copyright authentication step:

[0084] The zero watermark construction includes the following:

[0085] In the zero watermark construction, frequency domain features and spatial domain features are first extracted from the image, and then the spatial domain features, frequency domain features and copyright watermark redundancy are constructed into a matrix of size 128×128. The frequency domain features are used as the index matrix to select the pixels of the copyright watermark, and are XORed with the spatial domain features and the key matrix to obtain the zero watermark.

[0086] 1.1 Extract image feature sequence.

[0087] The brightness component of the color image of size 512×512×3 is divided into non-overlapping blocks. To increase the stability of the feature, the size of each sub-block is set to 16×16. The feature is constructed based on the size relationship between the sub-block mean and the overall mean of the brightness component, and the spatial domain feature f of size 32×32 is obtained. SP Then extract the frequency domain features of the image, reduce the dimension of the color image features, extract the brightness component into non-overlapping blocks, each sub-block size is 2×2, calculate the sub-block mean to construct a matrix of size 256×256, use discrete cosine transform (DCT) to extract the frequency domain features of the image, select a coefficient matrix of size 4×8 in the low-frequency part, and convert it into a 32-bit binary sequence f according to the positive and negative coefficients. FR as frequency domain features.

[0088] 1.2 Redundant construction carrier features.

[0089] The feature f of size 32×32 SP Redundancy is constructed into a feature F of size 128×128 SP , according to the 32-bit signature sequence f FR Redundant Structure F FR , the construction method is shown in formula (24-27).

[0090]

[0091] c mn =f SP (25)

[0092] F FR =(c1 … c k ) (26)

[0093] c k (i,j)=f FR (1,k) (27)

[0094] where ck To construct F FR The size of each sub-block is 128×4, i∈{1,2,3,...,128} and j∈{1,2,3,4}. k The subscript index of the internal element, composed of 32 c k Together they form a feature matrix F of size 128×12 FR ;c mn To construct F SP The sub-blocks, m∈{1,2,3}, n∈{1,2,3,4}, are composed of 16 c mn Together they form a feature matrix F of size 128×128 SP .

[0095] 1.3 Redundantly construct copyright watermarks.

[0096] Divide the 32×32 copyright logo w into non-overlapping blocks, with a sub-block size of 32 / p×32 / q, i∈{1,2,3,...,128}. Generate a random number matrix L of size (4×p)×(4×q), quantize the values ​​into p*q types, with the same number of each type of value, and save them as the sub-block arrangement key K1. Arrange the sub-blocks redundantly according to the method as shown in formula (29).

[0097]

[0098] Where p and q are the sub-block arrangement position indices, i∈{1,2,3,...,4×p}, j∈{1,2,3,...,4×q}, and the redundantly constructed W matrix is ​​composed of b pq The total size of the redundant copyright logo W is 128×128.

[0099] 1.4 Regional XOR of Feature Index

[0100] The frequency domain feature F after redundant construction FR As the index matrix, the redundant copyright logo W is divided into two regions, which are respectively connected with the redundant constructed spatial domain features F SP XOR with the randomly generated binary matrix key K2 to obtain the zero watermark Z, as shown in formula (30).

[0101]

[0102] Where i∈{1,2,3,...,128}, j∈{1,2,3,...,128} are pixel position indices.

[0103] The copyright certification steps include the following:

[0104] Using the feature extraction method in 1.1 and the redundant construction method in 1.2 and 1.3, extract the spatial domain feature F' of size 128×128 from the image that needs to verify copyright SP and frequency domain feature F' FR , select the pixel points in the zero watermark Z according to the frequency domain characteristics, and compare them with the spatial domain features F' SP XOR with the key K2 to obtain the redundant copyright watermark W', as shown in formula (31).

[0105]

[0106] Where i∈{1,2,3,...,12}8 and j∈{1,2,3,...,128} are pixel position indices. The image is attacked by Gaussian noise (attack intensity 0.2), and redundant frequency domain features F' are extracted. FR And the spatial characteristics after F' SP , redundant copyright watermark W' extraction step.

[0107] Watermark redundancy removal and ternary quantization:

[0108] 3.1 Watermarking to remove redundancy

[0109] The redundant copyright watermark W' is non-overlapping, and the sub-block size is consistent with the previous section, which is 32 / p×32 / q. The sub-block is marked as b' m,n , where m∈{1,2,3,...,4×p}, n∈{1,2,3,...,4×q}, and the key K1 is de-redundant according to the arrangement method, as shown in formula (32).

[0110]

[0111] Where m∈{1,2,3,...,×4p}, n∈{1,2,3,...,4×q} are sub-block position indices, and a 11 ,a 12 ,…,a pq is a matrix of size 32 / p×32 / q with an initial value of 0. m,n After the operation of formula (33), the watermark w' with a size of 32×32 is obtained.

[0112]

[0113] 3.2 Ternary Quantization

[0114] Since the 128×128 size W' is de-redundanted into a 32×32 size w', the value range of each pixel in w' is 0-16. According to formula (34), the watermark w' is quantized into a matrix w with only three values. final , where the value is {0,0.5,1}, wfinal As the final extraction result, copyright authentication is completed.

[0115]

[0116] where i∈{1,2,3,...,3}2, j∈{1,2,3,...,32} is w final Inner pixel index, s1 and s2 are two quantization thresholds s1 = α × R', s2 = (1-α) × R', satisfying 0≤s1≤s2≤16, α is the three-value quantization parameter, and R' is the redundancy of the copyright watermark. When the quantization threshold is s1 = 1 and s2 = 15, the redundant copyright watermark W' is obtained, and the final three-value matrix w is obtained after de-redundant three-value quantization. final .

[0117] The above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and changes can be made without departing from the principle of the present invention, and these improvements and changes are also regarded as the protection scope of the present invention.

Claims

1. A low false alarm zero watermark method based on region-specific XOR, characterized in that: Includes zero watermark construction and copyright authentication: Zero watermark construction includes the following steps: 1.1 Extracting image feature sequence: The brightness component of the color image of size 512×512×3 is divided into non-overlapping blocks. To increase the stability of the feature, the size of each sub-block is set to 16×16. The feature is constructed based on the size relationship between the sub-block mean and the overall mean of the brightness component, and the spatial domain feature f of size 32×32 is obtained. SP , then extract the frequency domain features of the image, reduce the dimension of the color image features, extract the brightness component into non-overlapping blocks, each sub-block size is 2×2, calculate the sub-block mean to construct a matrix of size 256×256, use discrete cosine transform to extract the frequency domain features of the image, select a coefficient matrix of size 4×8 in the low-frequency part, and convert it into a 32-bit binary sequence f according to the positive and negative coefficients. FR As frequency domain features; F FR and f FR is the frequency domain feature, F SP and f SP is the airspace feature; 1.2 Redundant construction of carrier features: The spatial domain features f of size 32×32 SP Redundancy is constructed into a spatial feature F of size 128×128 SP , according to the 32-bit frequency domain feature sequence f FR Redundancy is constructed into a 128×128 frequency domain feature F FR ; 1.3 Redundant construction of copyright watermark: Divide the 32×32 copyright logo w into non-overlapping blocks, with a sub-block size of 32 / p×32 / q, i∈{1,2,3,...,128}, generate a random number matrix L of size (4×p)×(4×q), quantize the values ​​into p*q types, with the same number of each type of value, save them as the sub-block arrangement key K1, and arrange the sub-blocks redundantly; p and q are the sub-block arrangement position indexes; 1.4 Regional XOR of feature index: The frequency domain feature F after redundant construction FR As the index matrix, the redundant copyright logo W is divided into two regions, which are respectively connected with the redundant constructed spatial domain features F SP XOR with the randomly generated binary matrix key K2 to obtain a zero watermark; Copyright certification includes the following steps: 2.1 Extract spatial features F' of size 128×128 from the image to be verified for copyright SP and frequency domain feature F' FR , select the pixel points in the zero watermark Z according to the frequency domain characteristics, and compare them with the spatial domain features F' SP XOR with key K2 to obtain redundant copyright watermark; In the copyright authentication process, the same feature extraction and regional XOR method is used to obtain a redundantly constructed copyright watermark image, which is quantized into a ternary matrix after redundancy removal to achieve copyright authentication that can distinguish similar images.

2. The low false alarm zero watermark method based on region-based XOR as claimed in claim 1, characterized in that: It also includes watermark de-redundancy and three-value quantization; Watermark redundancy removal: Redundant copyright watermarks are not overlapped in blocks, the sub-block size is 32 / p×32 / q, and the sub-block is marked as b' m,n , where m∈{1,2,3,...,4×p}, n∈{1,2,3,...,4×q}, and the key K1 is de-redundant according to the arrangement.

3. The low false alarm zero watermark method based on region-based XOR as claimed in claim 2, characterized in that: Set the quantization thresholds to s1 = 1, s2 = 15, and the value range of each position in w' is 0-16. After the quantization thresholds are set, compare the value in w' with s1 = 1, s2 = 15 to obtain the redundant copyright watermark W'. After de-redundant three-value quantization, the final three-value matrix w is obtained. final .

Images