CN117635408B - Copyright protection-oriented image zero-watermark method, device and medium - Google Patents

Abstract

The present invention discloses a zero-watermark method, device and medium for copyright protection of images, the method comprising: constructing a zero-watermark generation network; training the zero-watermark generation network; completing copyright registration based on the trained zero-watermark generation network according to the acquired image to be protected and the corresponding copyright information; completing copyright detection based on the trained zero-watermark generation network according to the acquired copyright-in doubtful image and the corresponding copyright information. The present invention makes full use of the characteristics of self-supervised learning and deep convolutional networks in feature extraction, solves the zero-watermark collision problem caused by similar copyrights, makes the zero-watermarks generated for the same image by similar copyrights have large differences, and also ensures that the generated zero-watermark has strong robustness, and can determine the copyright ownership of images subjected to various attacks.

Description

Image zero watermark method, device and medium for copyright protection

Technical Field

The present invention belongs to the technical field of image copyright protection, and more specifically, relates to an image zero-watermark method, device and medium for copyright protection.

Background technique

Image copyright protection refers to a series of measures taken by individuals or institutions that own image copyrights to protect their image works from being illegally copied, disseminated or used. With the development of digital technology and the popularization of the Internet, the transmission and acquisition of images have become very convenient and fast. However, this also brings the risk of infringement of image copyrights. Illegal copying, reproduction and theft of image works not only damage the economic interests of the image copyright owner, but may also lead to infringement of the copyright owner's intellectual property rights and even endanger national security and national defense security. Therefore, protecting the security and intellectual property rights of image copyrights has become an issue that needs to be urgently addressed.

Image zero watermark technology is an emerging digital watermark technology that can achieve copyright protection and identity authentication for images without adding any redundant information. Unlike traditional digital watermark technology, image zero watermark technology does not need to embed any additional information in the image, nor does it increase the image's storage space and computational complexity, providing a practical solution for image copyright protection. Digital zero watermark technology has achieved many applications in security protection in the fields of images, vector maps, videos, audio, etc., and has also achieved certain research and experimental results in image-related areas in recent years.

There has been a considerable amount of research on zero watermark technology for images both at home and abroad.

For example, the document "Wenquan, Sun Bofeng, Wang Shuxun. Concept and Application of Zero Watermark [J]. Journal of Electronics, 2003(02):214-216.), and later some scholars made improvements to it (Jiang F, Gao T, Li D. Arobust zero-watermarking algorithm for color image based on tensor mode expansion [J]. Multimedia Tools and Applications, 2020, 79:7599-7614. " first proposed the image zero watermark algorithm.

There are relatively few studies on image zero-watermarking based on deep learning. "Fierro-Radilla A, Nakano-MiyatakeM, Cedillo-Hernandez M, et al. A robust image zero-watermarking using convolutional neural networks[C]//2019 7th International Workshop on Biometrics and Forensics(IWBF).IEEE, 2019:1-5." uses the fully connected layer tensor in a simple binary classification network as image robust information.

"Li Ximing, Cai Hexin, Chen Zhihao, et al. Zero watermark algorithm based on attention mechanism and autoencoder construction [J]. Computer Systems and Applications, 2022, 31(09): 257-264.), Liu G et al. used style transfer network to generate zero watermark information (Liu G, Xiang R, Liu J, et al. An invisible and robust watermarking scheme using convolutional neural networks [J]. Expert Systems with Applications, 2022, 210: 118529. " Further improvements were made using the autoencoder algorithm and attention mechanism.

However, the zero watermark method proposed by the above technology has the following problems:

1) Poor robustness of watermarking: Currently, most zero-watermarking algorithms have poor robustness under cropping attacks, rotation attacks, and high-intensity filtering attacks.

2) Low efficiency of watermark generation: Traditional watermark algorithms use various complex mathematical or graphic transformations, the algorithm complexity is high, and it is impossible to use graphics cards for parallel acceleration.

3) The watermark has low anti-collision ability: When the copyright information is similar but not identical, the zero watermarks generated by the existing methods tend to be the same, which can easily lead to zero watermark collisions and confusion of copyright records.

4) The generalization of the watermark generation network is poor: Some deep learning zero-watermark methods require not only training on the sample set, but also training on the images of the protected content. The effect is poor for untrained images, so the generalization is low and it is difficult to apply in practice.

Summary of the invention

The present invention is provided to solve the above problems existing in the prior art. Therefore, a copyright protection-oriented image zero watermark method, device and medium are needed. In view of the problems of poor robustness, low efficiency of zero watermark generation, low anti-collision and poor generalization of current image zero watermark technology, a combination of self-supervision and supervised classification is adopted for training, which effectively improves robustness, efficiency and generalization, and reduces collision.

According to a first technical solution of the present invention, a copyright protection-oriented image zero watermark method is provided, the method comprising:

Construct a zero watermark generation network;

Train the zero-watermark generation network;

Based on the trained zero watermark generation network, copyright registration is completed according to the acquired image to be protected and the corresponding copyright information;

Based on the trained zero-watermark generation network, copyright detection is completed according to the acquired copyright-questionable images and corresponding copyright information.

Furthermore, the zero watermark generation network includes a weak noise layer, a strong noise layer, two deep convolutional networks and two multi-layer perceptrons, the two deep convolutional networks are respectively a first deep convolutional network and a second deep convolutional network, and the two multi-layer perceptrons are respectively a first multi-layer perceptron and a second multi-layer perceptron.

Further, the weak noise layer and the strong noise layer are used to perform random modification or transformation operations on the image content in sequence, wherein the degree of random modification or transformation operations performed by the strong noise layer on the image content is greater than the degree of random modification or transformation operations performed by the weak noise layer on the image content, and the random modification includes one of cropping the image, rotating the image, salt and pepper noise, and filtering, and a combination thereof;

Furthermore, the input of the zero watermark generation network includes an image X and copyright information W, wherein the image is an image content in any format, and the copyright information is any copyright content that can identify the copyright source and owner information, and the copyright information includes any form of a coding sequence, text, icon, image, video, and voice;

The image X is passed through the weak noise layer to obtain a first image X _w , and then the first image X _w has two branches;

The first branch inputs the first image _Xw into the first deep convolutional network to obtain a first feature map _Xw -N, inputs the copyright information W into the second deep convolutional network to obtain a second feature map WN, performs tensor connection on the first feature map _Xw -N and the second feature map WN, and then inputs the first multi-layer perceptron to obtain a third feature map _Pw , and then inputs the third feature map _Pw into the second multi-layer perceptron to obtain a copyright code _Yw ;

The second branch inputs the first image _Xw into the strong noise layer to obtain a second image _Xs , inputs the second image _Xs into the first deep convolutional network to obtain a fourth feature map _Xs -N, performs tensor connection on the fourth feature map _Xs -N and the second feature map WN, and then inputs them into the first multi-layer perceptron to obtain a fifth feature map _Ps , and then inputs the fifth feature map _Ps into the second multi-layer perceptron to obtain the copyright code _Ys .

Furthermore, the training of the zero watermark generation network specifically includes:

Obtain image datasets and copyright datasets;

Based on the image dataset and the copyright dataset, the zero watermark generation network is trained using the following loss function:

Loss＝Loss1*weight1+Loss2*weight2+Loss3*weight3

Among them, Loss1, Loss2 and Loss3 represent three loss functions, weight1, weight2 and weight3 are the weights of the three loss functions, which can be adjusted arbitrarily under the following constraints:

weight1+weight2+weight3=1

The three loss functions are expressed as:

Where i is the subscript of the tensor, N is the total length of the tensor, Y _w , Y _s , and Y represent copyright codes and are of equal length, and P _w and P _s are of equal length.

Furthermore, the zero watermark generation network based on the training completes the copyright registration according to the acquired image to be protected and the corresponding copyright information, specifically including:

Obtain the images and copyright information to be protected;

Inputting the image to be protected into a first deep convolutional network to obtain a first tensor, inputting the copyright information into a second deep convolutional network to obtain a second tensor, connecting the first tensor and the second tensor, and inputting the connected first tensor and the second tensor into a first multi-layer perceptron to obtain a feature map;

Binarizing the feature map, wherein a binarization threshold is set to an average value of the feature map;

The binary feature map is scrambled, the scrambling key is set to key, and the scrambled result ZW is the generated zero watermark;

Store key and ZW as copyright records to complete the copyright registration of the image.

Furthermore, the copyright detection is completed based on the trained zero watermark generation network according to the acquired copyright-questionable image and corresponding copyright information, specifically including:

Get copyright-questionable images and copyright information W;

Inputting the copyright-questionable image into the first deep convolutional network to obtain a third tensor, inputting the copyright information into the second deep convolutional network to obtain a fourth tensor, connecting the third tensor and the fourth tensor, and then inputting the connected third tensor and fourth tensor into the first multi-layer perceptron to obtain a feature map;

Binarizing the feature map, wherein a binarization threshold is set to an average value of the feature map;

The binarized feature map is scrambled, and the scrambling key uses the key stored during copyright registration. The scrambled result ZW2 is the generated zero watermark.

Calculate the correlation coefficient NC between the zero watermark ZW in the copyright annotation record and the generated zero watermark ZW2;

Compare NC with a preset threshold. If NC is greater than or equal to the preset threshold, determine that the copyright of the copyright-questionable image belongs to copyright information W. If NC is less than the preset threshold, determine that the copyright of the copyright-questionable image does not belong to copyright information W.

Furthermore, the correlation coefficient NC between the zero watermark ZW in the copyright annotation record and the generated zero watermark ZW2 is calculated by the following formula:

Among them, XNOR is the exclusive OR operation, i is the subscript of the tensor, and N is the total length of the tensor.

According to a second technical solution of the present invention, there is provided an image zero watermark device for copyright protection, the device comprising:

A network construction module, configured to construct a zero-watermark generation network;

A network training module, configured to train a zero-watermark generation network;

The copyright registration module is configured to complete copyright registration based on the trained zero watermark generation network and the acquired image to be protected and the corresponding copyright information;

The copyright detection module is configured to complete copyright detection based on the trained zero-watermark generation network and the acquired copyright-questionable images and corresponding copyright information.

Furthermore, the zero watermark generation network includes a weak noise layer, a strong noise layer, two deep convolutional networks and two multi-layer perceptrons, the two deep convolutional networks are respectively a first deep convolutional network and a second deep convolutional network, and the two multi-layer perceptrons are respectively a first multi-layer perceptron and a second multi-layer perceptron.

Furthermore, the weak noise layer and the strong noise layer are used to perform random modification or transformation operations on the image content in sequence, wherein the degree to which the strong noise layer performs random modification or transformation operations on the image content is greater than the degree to which the weak noise layer performs random modification or transformation operations on the image content, and the random modification includes one of cropping the image, rotating the image, salt and pepper noise, and filtering, and a combination thereof.

Furthermore, the input of the zero watermark generation network includes an image X and copyright information W, wherein the image is an image content in any format, and the copyright information is any copyright content that can identify the copyright source and owner information, and the copyright information includes any form of a coding sequence, text, icon, image, video, and voice;

The network building module is further configured to:

The image X is passed through the weak noise layer to obtain a first image X _w , and then the first image X _w has two branches;

The first branch inputs the first image _Xw into the first deep convolutional network to obtain a first feature map _Xw -N, inputs the copyright information W into the second deep convolutional network to obtain a second feature map WN, performs tensor connection on the first feature map _Xw -N and the second feature map WN, and then inputs the first multi-layer perceptron to obtain a third feature map _Pw , and then inputs the third feature map _Pw into the second multi-layer perceptron to obtain a copyright code _Yw ;

The second branch inputs the first image _Xw into the strong noise layer to obtain a second image _Xs , inputs the second image _Xs into the first deep convolutional network to obtain a fourth feature map _Xs -N, performs tensor connection on the fourth feature map _Xs -N and the second feature map WN, and then inputs them into the first multi-layer perceptron to obtain a fifth feature map _Ps , and then inputs the fifth feature map _Ps into the second multi-layer perceptron to obtain the copyright code _Ys .

Furthermore, the network training module is further configured as follows:

Obtain image datasets and copyright datasets;

Based on the image dataset and the copyright dataset, the zero watermark generation network is trained using the following loss function:

Loss＝Loss1*weight1+Loss2*weight2+Loss3*weight 3

Among them, Loss1, Loss2 and Loss3 represent three loss functions, weight1, weight2 and weight3 are the weights of the three loss functions, which can be adjusted arbitrarily under the following constraints:

weight1+weight2+weight3=1

The three loss functions are expressed as:

Where i is the subscript of the tensor, N is the total length of the tensor, Y _w , Y _s , and Y represent copyright codes and are of equal length, and P _w and P _s are of equal length.

Furthermore, the copyright registration module is further configured to:

Obtain the images and copyright information to be protected;

Inputting the image to be protected into a first deep convolutional network to obtain a first tensor, inputting the copyright information into a second deep convolutional network to obtain a second tensor, connecting the first tensor and the second tensor, and inputting the connected first tensor and the second tensor into a first multi-layer perceptron to obtain a feature map;

Binarizing the feature map, wherein a binarization threshold is set to an average value of the feature map;

The binary feature map is scrambled, the scrambling key is set to key, and the scrambled result ZW is the generated zero watermark;

Store key and ZW as copyright records to complete the copyright registration of the image.

Furthermore, the copyright detection module is further configured to:

Get copyright-questionable images and copyright information W;

Inputting the copyright-questionable image into the first deep convolutional network to obtain a third tensor, inputting the copyright information into the second deep convolutional network to obtain a fourth tensor, connecting the third tensor and the fourth tensor, and then inputting the connected third tensor and fourth tensor into the first multi-layer perceptron to obtain a feature map;

Binarizing the feature map, wherein a binarization threshold is set to an average value of the feature map;

The binarized feature map is scrambled, and the scrambling key uses the key stored during copyright registration. The scrambled result ZW2 is the generated zero watermark.

Calculate the correlation coefficient NC between the zero watermark ZW in the copyright annotation record and the generated zero watermark ZW2;

Compare NC with a preset threshold. If NC is greater than or equal to the preset threshold, determine that the copyright of the copyright-questionable image belongs to copyright information W. If NC is less than the preset threshold, determine that the copyright of the copyright-questionable image does not belong to copyright information W.

Furthermore, the copyright detection module is further configured to calculate the correlation coefficient NC between the zero watermark ZW in the copyright annotation record and the generated zero watermark ZW2 by the following formula:

Among them, XNOR is the exclusive OR operation, i is the subscript of the tensor, and N is the total length of the tensor.

According to a third technical solution of the present invention, a readable storage medium is provided, wherein the readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method as described above.

The present invention has at least the following beneficial effects:

(1) Zero watermark has high robustness. It uses a multi-layer convolutional neural network as the backbone network to extract features, adds a variety of noises for training, and adopts a self-supervised learning scheme to greatly improve the watermark's anti-attack ability.

(2) Fast zero watermark generation speed. With CUDA acceleration, the time to generate a zero watermark for a single image is only 20ms on a single A100-40G card, which reduces time overhead.

(3) Low zero-watermark collision probability. Introducing hash-generated copyright codes during training makes the zero-watermarks generated by the model have low collision probability, effectively avoiding the situation where zero-watermarks generated by similar copyrights are easily confused.

(4) The zero watermark storage space is small. The zero watermark generated by the present invention is only 256 bits, which greatly reduces the storage overhead.

(5) The zero watermark generation network has strong generalization. The present invention uses a method combining self-supervised learning with supervised classification to train the model. It does not need to train the image to be protected. The output can be generated by directly inputting the network, which has strong generalization.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 shows an overall flow chart of an image zero watermark method for copyright protection according to an embodiment of the present invention;

FIG2 is a schematic diagram showing an image to be protected according to an embodiment of the present invention;

FIG3 shows a schematic diagram of copyright information according to an embodiment of the present invention;

FIG4 is a schematic diagram showing an image subjected to a rotation attack according to an embodiment of the present invention;

FIG5 is a schematic diagram showing an image attacked by salt and pepper noise during a test according to an embodiment of the present invention;

FIG6 is a schematic diagram showing an image attacked by Gaussian noise during a test according to an embodiment of the present invention;

FIG7 is a schematic diagram showing an image subjected to a random cropping attack during a test according to an embodiment of the present invention;

FIG8 is a schematic diagram showing an image subjected to an upper left corner cropping attack during a test according to an embodiment of the present invention;

FIG9 is a schematic diagram showing an image subjected to a mean filter attack during a test according to an embodiment of the present invention;

FIG10 is a schematic diagram showing an image attacked by Gaussian filtering during a test according to an embodiment of the present invention;

FIG. 11 shows a zero-watermark NC result diagram generated by similar copyright according to an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and specific embodiments, but are not intended to limit the present invention. For the various steps described herein, if there is no necessity for a causal relationship between each other, the order in which they are described as examples herein should not be regarded as a limitation, and those skilled in the art should know that they can be adjusted in order, as long as the logic between them is not destroyed, resulting in the inability to implement the entire process.

The embodiment of the present invention provides an image zero watermark method for copyright protection. As shown in FIG1 , the overall process includes four steps: constructing a zero watermark generation network, training the zero watermark generation network, copyright registration, and copyright detection.

In this example, Lena is selected as the image to be protected, as shown in Figure 2, and the emblem of Nanjing University of Finance and Economics is selected as the copyright image, as shown in Figure 3. With respect to the entire process of constructing a zero watermark generation network, training a zero watermark generation network, copyright registration, and copyright detection of the present invention, an embodiment of the present invention is given to further explain the present invention in detail.

1) Construct a zero watermark generation network

(1) A weak noise layer is set, including cropping 0.0625 around, randomly rotating 10 degrees, randomly cropping 0.05-0.1 parts of the image content, 0.02 salt and pepper noise, 0.01 Gaussian noise, 5*5 mean filter and Gaussian filter, and each noise has the same weight.

(2) Setting a strong noise layer, including cropping 0.25 around, randomly rotating 180 degrees, randomly cropping 0.2-0.35 parts of the image content, 0.3 salt and pepper noise, 0.3 Gaussian noise, 13*13 mean filter and Gaussian filter, each noise weight is the same;

(3) The architecture of the deep convolutional network Net1 is set as shown in Table 1.

Table 1Net1 network structure

(4) Set the network architecture of Net2 as shown in Table 2.

Table 2 Net2 network structure

(5) Assume that both MLP1 and MLP2 are neural networks with two fully connected layers.

2) Training the zero watermark generation network

(1) In terms of the process, the image to be protected X passes through a parameter-free weak noise layer and is randomly enhanced with some weak noise to obtain X _w . X _w passes through Net1 to generate a 1×768 tensor. The copyright information W is input into Net2 to obtain a 1×500 tensor. The two tensors are merged into a 1×1268 tensor and input into MLP1 to output a 1×256 tensor P _W . P _W is input into MLP2 to output a 1×15 tensor Y _W . The hash value of the copyright information W is calculated using MD5 and reconstructed into a 1×15 tensor Y. X _w is input into a strong noise layer to obtain X _s . A similar process as shown in Figure 1 is used to obtain a 1×256 tensor P _s and a 1×15 tensor Y _S .

(2) The objective function Loss of the training is:

Where i is the subscript of the tensor, N is the total length of the tensor, _Yw , _Ys , and Y are of equal length, _Pw and _Ps are of equal length, and the weights of the three losses are all set to 1/3.

(3) The network uses 30,801 images from miniImageNet for training and 1,284 images for evaluation and testing. The main hardware environment is the NVDIA A100-40G graphics card.

(4) The AdaMax optimizer is used for training, with the learning rate set to 0.00001, the number of epochs set to 200, and the batch size set to 20.

3) Copyright registration

Copyright registration includes the following steps:

Step 1: The user submits the image X to be protected (Figure 2) and copyright information W (Figure 3);

Step 2: Input X into Net1, input W into Net2, concatenate the two output tensors, and then input them into MLP1 to obtain a 1×256 tensor P _w

Step 3: Binarize P _w , where the binarization threshold can be set to the average value of P _w ;

Step 4: Perform Arnold scrambling, set the scrambling key to key, and the scrambling result ZW is the generated zero watermark.

Step 5: Store key and ZW as copyright records to complete the copyright registration of image X (Figure 2).

4) Copyright detection

Copyright detection includes the following steps:

Step 1: The user submits the attacked and copyright-questionable image X′ (Figure 4) and copyright information W (Figure 3);

Step 2: Input X′ into Net1, input W into Net2, concatenate the two output tensors, and then input them into MLP1 to obtain P _w 2;

Step 3: Binarize P _w 2, where the binarization threshold can be set to the average value of P _w 2;

Step 4: Perform Arnold scrambling. The scrambling key uses the key stored during copyright registration. The scrambled result ZW2 is the generated zero watermark.

Step 5: Calculate the correlation coefficient NC between ZW and ZW2. The calculation formula is as follows:

Wherein, XNOR is an exclusive OR operation, and in this embodiment, the calculation result of NC is 0.995.

Step 5: Compare whether NC is higher than a preset threshold to determine the copyright ownership. In this embodiment, the threshold is set to 0.8. Since the calculated NC is higher than 0.8, it is considered that the copyright of the copyright-questionable image X′ (FIG. 4) belongs to W (FIG. 3).

The method proposed in the present invention is a zero watermark method based on a convolutional network. To further verify the performance of the method, the following robustness test and anti-collision test are performed on the method. The image to be protected used in the test is Figure 2, and the copyright information is Figure 3. In the robustness test, a zero watermark is first generated for the original image, i.e., Figure 2. Then, a rotation attack, salt and pepper noise attack, Gaussian noise attack, cropping attack, filtering attack and other attack methods are performed on Figure 2. The difference between the zero watermark generated by the image after the attack and the original zero watermark is calculated to verify the robustness. In the anti-collision test, the anti-collision is verified by comparing the difference in zero watermarks generated by similar copyrights. The detailed experimental results are as follows.

(1) Spin attack

Different degrees of rotation attacks are performed on Figure 2. An example of an attack with a rotation of 10 degrees is shown in Figure 4. The experimental results are shown in Table 3.

Table 3 Rotation attack results

As can be seen from Table 3, the correlation coefficients of watermark detection are all higher than 0.995. This shows that this method can effectively resist various angle rotation attacks.

(2) Salt and pepper noise attack

Different degrees of salt and pepper noise attacks are applied to Figure 2. An attack example with a noise parameter of 0.05 is shown in Figure 5, and the experimental results are shown in Table 4.

Table 4 Salt and pepper attack results

As can be seen from Table 4, the correlation coefficients of watermark detection are all higher than 0.98, indicating that this method can resist salt and pepper noise attacks.

(3) Gaussian noise attack

Different degrees of Gaussian noise attacks are applied to Figure 2. An attack example with a Gaussian noise parameter of 0.05 is shown in Figure 6. The experimental results are shown in Table 5.

Table 5 Gaussian noise attack results

As can be seen from Table 5, the correlation coefficients of watermark detection are all higher than 0.97, indicating that this method is effective in resisting Gaussian noise attacks.

(4) Clipping attack

Random cropping attacks of different ratios are applied to Figure 2. An attack example with a cropping ratio of 0.25 is shown in Figure 7. The experimental results are shown in Table 6.

Table 6 Random cropping attack results

In addition, cropping attacks of different proportions are applied to the four corners of the image. An attack example of cropping 0.0625 in the upper left corner is shown in Figure 8. The experimental results are shown in Table 7.

Table 7. Cropping attack results

As can be seen from Tables 6 and 7, in experiments with different degrees of random cropping attacks or cropping attacks on four corners, the correlation coefficient of the extracted watermark information is still higher than 0.97, and the correlation coefficient of most cropping attack watermarks is 1. This shows that after the cropping attack, the watermark detection still has high reliability. It can be seen that this method has strong resistance to cropping attacks.

(5) Filter Attack

The test image is subjected to a mean filter attack. An example of an attack with a mean filter of 5 is shown in Figure 9, and the results are shown in Table 8.

Table 8 Mean filter attack results

The test image is subjected to a Gaussian filter attack. An attack example with a Gaussian filter of 5 is shown in Figure 10, and the results are shown in Table 9.

Table 9 Gaussian filter attack results

It can be seen from Tables 8 and 9 that in the mean filter attack or Gaussian filter attack experiments of different degrees, the correlation coefficient of the extracted watermark information is higher than 0.99, and the correlation coefficient of most attack watermarks is 1. This shows that the method is highly robust to filter attacks.

(6) Collision resistance test

In order to verify the anti-collision ability of this method, eight similar copyright information are registered with zero watermarks together with Figure 2, and the correlation coefficients of the generated zero watermarks are compared to judge the similarity between them. The experimental results are shown in Figure 11.

As shown in Figure 11, in the anti-collision experiment, although the styles of the copyright information are similar, the correlation coefficients of the generated zero watermark information are all lower than 0.8, indicating that the correlation between the zero watermarks is low, that is, it is difficult to generate repeated and very similar zero watermarks, thus avoiding copyright confusion. It can be seen that the zero watermark generated by this method has good anti-collision performance.

In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present invention with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. The elements in the claims will be interpreted broadly based on the language adopted in the claims, and are not limited to the examples described in this specification or during the implementation of this application, and the examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. For example, those of ordinary skill in the art can use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features can be grouped together to simplify the present invention. This should not be interpreted as a feature of an invention that is not claimed for protection being necessary for any claim. On the contrary, the subject matter of the present invention may be less than all the features of the embodiments of a specific invention. Thus, the following claims are incorporated into the specific embodiments as examples or embodiments, wherein each claim is independently used as a separate embodiment, and it is considered that these embodiments can be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the full scope of the equivalent forms of the attached claims and these claims.

Claims (7)

1. A copyright protection-oriented image zero watermark method, characterized in that the method comprises: Construct a zero watermark generation network; Train the zero-watermark generation network; Based on the trained zero watermark generation network, copyright registration is completed according to the acquired image to be protected and the corresponding copyright information; Based on the trained zero-watermark generation network, copyright detection is completed according to the acquired copyright-questionable images and corresponding copyright information; The zero watermark generation network includes a weak noise layer, a strong noise layer, two deep convolutional networks and two multi-layer perceptrons, the two deep convolutional networks are respectively a first deep convolutional network and a second deep convolutional network, and the two multi-layer perceptrons are respectively a first multi-layer perceptron and a second multi-layer perceptron; The input of the zero watermark generation network includes an image X and copyright information W, wherein the image is an image content in any format, and the copyright information is any copyright content that can identify the copyright source and owner information, and the copyright information includes any form of a coding sequence, text, icon, image, video, and voice; The image X passes through the weak noise layer to obtain a first image _Xw , and then the first image _Xw has two branches; the first branch inputs the first image _Xw into the first deep convolutional network to obtain a first feature map _Xw -N, inputs the copyright information W into the second deep convolutional network to obtain a second feature map WN, performs tensor connection on the first feature map _Xw -N and the second feature map WN, and then inputs them into the first multi-layer perceptron to obtain a third feature map _Pw , and then inputs the third feature map _Pw into the second multi-layer perceptron to obtain a copyright code _Yw ; the second branch inputs the first image _Xw into the strong noise layer to obtain a second image _Xs , inputs the second image _Xs into the first deep convolutional network to obtain a fourth feature map _Xs -N, performs tensor connection on the fourth feature map _Xs -N and the second feature map WN, and then inputs them into the first multi-layer perceptron to obtain a fifth feature map _Ps , and then inputs the fifth feature map _Ps into the second multi-layer perceptron to obtain a copyright code _Ys ; The method of generating a zero watermark based on the trained network completes the copyright registration according to the acquired image to be protected and the corresponding copyright information, specifically including: Acquire an image to be protected and copyright information; input the image to be protected into a first deep convolutional network to obtain a first tensor, input the copyright information into a second deep convolutional network to obtain a second tensor, connect the first tensor and the second tensor, and input the connected first tensor and second tensor into a first multi-layer perceptron to obtain a feature map; binarize the feature map, wherein the binarization threshold is set to the average value of the feature map; scramble the binarized feature map, set the scrambling key to key, and the scrambled result ZW is the generated zero watermark; store key and ZW as copyright records to complete the copyright registration of the image. 2. The method according to claim 1 is characterized in that the weak noise layer and the strong noise layer are used to perform random modification or transformation operations on the image content in sequence, wherein the degree to which the strong noise layer performs random modification or transformation operations on the image content is greater than the degree to which the weak noise layer performs random modification or transformation operations on the image content, and the random modification includes one of cropping the image, rotating the image, salt and pepper noise, and filtering, and a combination thereof. 3. The method according to claim 1, characterized in that the training of the zero watermark generation network specifically comprises: Obtain image datasets and copyright datasets; Based on the image dataset and the copyright dataset, the zero watermark generation network is trained using the following loss function: Loss＝Loss1*weight1+Loss2*weight2+Loss3*weight 3 Among them, Loss1, Loss2 and Loss3 represent three loss functions, weight1, weight2 and weight3 are the weights of the three loss functions, which can be adjusted arbitrarily under the following constraints: weight1+weight2+weight3=1 The three loss functions are expressed as: Where i is the subscript of the tensor, N is the total length of the tensor, Y _w , Y _s , and Y represent copyright codes and are of equal length, and P _w and P _s are of equal length. 4. The method according to claim 1 is characterized in that the copyright detection is completed based on the trained zero watermark generation network according to the acquired copyright-questionable image and corresponding copyright information, specifically comprising: Get copyright-questionable images and copyright information W; Inputting the copyright-questionable image into the first deep convolutional network to obtain a third tensor, inputting the copyright information into the second deep convolutional network to obtain a fourth tensor, connecting the third tensor and the fourth tensor, and then inputting the connected third tensor and fourth tensor into the first multi-layer perceptron to obtain a feature map; Binarizing the feature map, wherein a binarization threshold is set to an average value of the feature map; The binary feature map is scrambled. The scrambling key uses the key stored during copyright registration. The scrambled result ZW2 is the generated zero watermark. Calculate the correlation coefficient NC between the zero watermark ZW in the copyright annotation record and the generated zero watermark ZW2; Compare NC with a preset threshold. If NC is greater than or equal to the preset threshold, determine that the copyright of the copyright-questionable image belongs to copyright information W. If NC is less than the preset threshold, determine that the copyright of the copyright-questionable image does not belong to copyright information W. 5. The method according to claim 4, characterized in that the correlation coefficient NC between the zero watermark ZW in the copyright annotation record and the generated zero watermark ZW2 is calculated by the following formula: Among them, XNOR is the exclusive OR operation, i is the subscript of the tensor, and N is the total length of the tensor. 6. An image zero watermark device for copyright protection, characterized in that the device comprises: A network construction module, configured to construct a zero-watermark generation network; A network training module, configured to train a zero-watermark generation network; The copyright registration module is configured to complete copyright registration based on the trained zero watermark generation network and the acquired image to be protected and the corresponding copyright information; The copyright detection module is configured to complete copyright detection based on the acquired copyright-questionable image and corresponding copyright information based on the trained zero-watermark generation network; The zero watermark generation network includes a weak noise layer, a strong noise layer, two deep convolutional networks and two multi-layer perceptrons, the two deep convolutional networks are respectively a first deep convolutional network and a second deep convolutional network, and the two multi-layer perceptrons are respectively a first multi-layer perceptron and a second multi-layer perceptron; The input of the zero watermark generation network includes an image X and copyright information W, wherein the image is an image content in any format, and the copyright information is any copyright content that can identify the copyright source and owner information, and the copyright information includes any form of a coding sequence, text, icon, image, video, and voice; The image X passes through the weak noise layer to obtain a first image _Xw , and then the first image _Xw has two branches; the first branch inputs the first image _Xw into the first deep convolutional network to obtain a first feature map _Xw -N, inputs the copyright information W into the second deep convolutional network to obtain a second feature map WN, performs tensor connection on the first feature map _Xw -N and the second feature map WN, and then inputs them into the first multi-layer perceptron to obtain a third feature map _Pw , and then inputs the third feature map _Pw into the second multi-layer perceptron to obtain a copyright code _Yw ; the second branch inputs the first image _Xw into the strong noise layer to obtain a second image _Xs , inputs the second image _Xs into the first deep convolutional network to obtain a fourth feature map _Xs -N, performs tensor connection on the fourth feature map _Xs -N and the second feature map WN, and then inputs them into the first multi-layer perceptron to obtain a fifth feature map _Ps , and then inputs the fifth feature map _Ps into the second multi-layer perceptron to obtain a copyright code _Ys ; The method of generating a zero watermark based on the trained network completes the copyright registration according to the acquired image to be protected and the corresponding copyright information, specifically including: Acquire an image to be protected and copyright information; input the image to be protected into a first deep convolutional network to obtain a first tensor, input the copyright information into a second deep convolutional network to obtain a second tensor, connect the first tensor and the second tensor, and input the connected first tensor and second tensor into a first multi-layer perceptron to obtain a feature map; binarize the feature map, wherein the binarization threshold is set to the average value of the feature map; scramble the binarized feature map, set the scrambling key to key, and the scrambled result ZW is the generated zero watermark; store key and ZW as copyright records to complete the copyright registration of the image. 7. A readable storage medium, characterized in that the readable storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the method according to any one of claims 1 to 5.