Automatic robust image watermark generation method and system based on deep neural network

Abstract

The invention provides an automatic robust image watermark generation method and system based on a deep neural network, and the method comprises the steps: carrying out the coding of watermark information through a coding function fitted by a convolutional neural network, and enabling the coding information to be embedded into a carrier image through an embedder fitted by the convolutional neural network; an invariance layer formed by a full-connection neural network is introduced to process distortion in the process of transmitting the marked image through a communication channel; extracting watermark information from the marked image through an extractor fitted by a convolutional neural network; and finally, decoding the coded information through a decoding function fitted by the convolutional neural network, reconstructing the watermark information, and overcoming the problem that robustness can be realized under the condition that distortion possibly occurring on the marked image needs to be known in advance in the prior art.

Description

Technical Field

[0001] The present application relates to the field of digital security and digital watermark technology, and in particular to an automated robust image watermark generation method and system based on a deep neural network. Background Art

[0002] Digital image watermarking is a technology that hides information in digital images, which is usually used for copyright protection, data verification and integrity protection. For different target scenarios, watermark information can be presented in different forms: for example, watermarks can be some random bits or electronic signatures for image protection and authentication; or some hidden messages for covert communication. In addition, watermark coding can also be used for different purposes, such as increasing security through encryption methods or restoring information integrity through error correction codes during network attacks.

[0003] The basic principle of digital watermarking is to embed some identification information directly into the image data, which can be copyright information, image source, date, etc. When the authenticity or copyright of the image needs to be verified, the watermark information can be extracted from the image for comparison.

[0004] Traditional image watermarking methods have shortcomings, for example, they can only tolerate certain types of distortions or can only resist a limited range of image analysis. To overcome these shortcomings, researchers began to introduce deep learning into digital image watermarking. By fitting and generalizing the ability of complex features, high-level and low-level watermarks of image watermarks are extracted from multiple instance big data to generate digital image watermarks in an adaptive manner. However, existing watermark generation methods based on deep learning neither fully apply the fitting ability to learn to achieve automatic embedding and extraction of digital watermarks, nor achieve the characteristics of robustness and blindness at the same time.

[0005] Therefore, there is an urgent need for a targeted automated robust image watermark generation method and system based on deep neural networks. Summary of the invention

[0006] The purpose of the present invention is to provide an automated robust image watermark generation method and system based on a deep neural network, which utilizes the fitting ability of the deep neural network to learn and automatically generate digital image watermarks. Its deep learning architecture is trained in an unsupervised manner and does not require any prior knowledge or adversarial samples of possible attacks to achieve its robustness.

[0007] In a first aspect, the present application provides an automated robust image watermark generation method based on a deep neural network, corresponding to the data sending side, the method comprising:

[0008] Given two input spaces, namely watermark image samples w i and carrier image sample ci ;

[0009] The watermark image sample w i First, it is processed by two identical convolution blocks to generate an output sample of 32×32×1, then a 1×1×24 convolution expansion is performed to expand it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×48 convolution expansion is performed again to expand it to 32×32×48 to obtain the encoded watermark information.

[0010] The encoded watermark information with a size of 32×32×48 Reshaped to 128×128×3, the embedder applies a convolutional block to extract the watermark information reshaped to 128×128×3 The feature map to be embedded;

[0011] The feature map to be embedded is parallel to the carrier image sample c along the channel dimension. i Connect to form a 128×6×6 connection result;

[0012] The connection result is processed by the convolution block and the 1×1×3 convolution to form a marked image with a size of 128×128×3 and embedded with watermark information;

[0013] The marked image with the embedded watermark information is sent.

[0014] In a second aspect, the present application provides an automated robust image watermark generation method based on a deep neural network, corresponding to the data receiving side, the method comprising:

[0015] receiving a tagged image via a communication channel;

[0016] Introducing the invariance layer composed of fully connected neural networks Dealing with distortion generated during transmission;

[0017] The extractor extracts the 128×3×3 watermark information from the labeled image by applying exactly the same two convolution blocks and a 1×1×3 convolution, which is the exact opposite process of the embedder.

[0018] The extracted watermark information of size 128×128×3 Reshaped to 32×32×48;

[0019] Reshape the watermark information into 32×32×48 As the input sample, it is first processed by two identical convolution blocks to generate an output sample of 32×32×48, then a 1×1×24 convolution is performed to tighten it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×1 convolution is performed to tighten it to 32×32×1, and the watermark information restored after decoding is obtained.

[0020] Verification is performed based on the restored watermark information to obtain a verification result.

[0021] In a third aspect, the present application provides an automated robust image watermark generation system based on a deep neural network, the system comprising:

[0022] The input space unit is used to give two input spaces, namely the watermark image sample w i and carrier image sample c i ;

[0023] The encoding unit is used to encode the watermark image sample w i First, it is processed by two identical convolution blocks to generate an output sample of 32×32×1, then a 1×1×24 convolution expansion is performed to expand it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×48 convolution expansion is performed again to expand it to 32×32×48 to obtain the encoded watermark information.

[0024] The feature extraction unit is used to convert the encoded watermark information of size 32×32×48 Reshape to 128×128×3, call the embedder, apply the convolutional block to extract the watermark information reshaped to 128×128×3 The feature map to be embedded;

[0025] A connection unit is used to connect the feature map to be embedded with the carrier image sample c along the channel dimension i Connect to form a 128×6×6 connection result;

[0026] An image generating unit, used for processing the connection result through a convolution block and a 1×1×3 convolution to form a marked image with a size of 128×128×3 and embedded with watermark information;

[0027] The sending unit is used to send the marked image embedded with the watermark information.

[0028] In a fourth aspect, the present application provides an automated robust image watermark generation system based on a deep neural network, the system comprising:

[0029] A receiving unit, configured to receive a marked image via a communication channel;

[0030] Rectification unit, used to introduce the invariance layer of the fully connected neural network Dealing with distortion generated during transmission;

[0031] The watermark extraction unit is used to call the extractor to extract the 128×3×3 watermark information from the marked image by applying the same two convolution blocks and a 1×1×3 convolution, that is, the process is completely opposite to the embedder.

[0032] The watermark reshaping unit is used to reshape the extracted watermark information into a size of 128×128×3 Reshaped to 32×32×48;

[0033] The decoding unit is used to reshape the watermark information into 32×32×48 As the input sample, it is first processed by two identical convolution blocks to generate an output sample of 32×32×48, then a 1×1×24 convolution is performed to tighten it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×1 convolution is performed to tighten it to 32×32×1, and the watermark information restored after decoding is obtained.

[0034] The verification unit is used to perform verification according to the restored watermark information to obtain a verification result.

[0035] In a fifth aspect, the present application provides an automated robust image watermark generation system based on a deep neural network, the system comprising a processor and a memory:

[0036] The memory is used to store program code and transmit the program code to the processor;

[0037] The processor is used to execute any one of the four possible methods of the first aspect or the second aspect according to the instructions in the program code.

[0038] In a sixth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store program code, and the program code is used to be executed by a processor to implement any one of the four possible methods in the first aspect or the second aspect.

[0039] Beneficial Effects

[0040] The present invention provides an automated robust image watermark generation method and system based on a deep neural network. The watermark information is encoded by a coding function fitted by a convolutional neural network, and then the encoded information is embedded into a carrier image by an embedder fitted by the convolutional neural network. An invariance layer composed of a fully connected neural network is introduced to handle the distortion in the process of transmitting the marked image through a communication channel. The watermark information is extracted from the marked image by an extractor fitted by the convolutional neural network. Finally, the encoded information is decoded by a decoding function fitted by the convolutional neural network to reconstruct the watermark information, thereby overcoming the problem in the prior art that robustness can only be achieved when the distortion that may occur on the marked image needs to be known in advance.

[0041] The method and system of the present invention have the following advantages and effects:

[0042] We exploit the fitting capabilities of deep neural networks to generalize image watermarking algorithms, demonstrating an architecture for training watermarking tasks in an unsupervised manner and achieving robustness without requiring prior knowledge of possible distortions on the labeled image.

[0043] The image watermarking algorithm is automatically learned by using the fitting ability of deep neural networks to achieve system automation.

[0044] It not only demonstrates good performance against individual common attacks, but also has considerable capabilities and potential in cutting-edge and challenging camera surveillance applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 is a system architecture diagram of the present invention;

[0048] Figure 3 The present invention provides an architecture design for the convolutional block according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0050] The present application provides an automated robust image watermark generation method based on a deep neural network, such as Figure 1As shown, the method includes:

[0051] The data sending side is given two input spaces, namely the watermark image sample w i and carrier image sample c i ;

[0052] The watermark image sample w i First, it is processed by two identical convolution blocks to generate an output sample of 32×32×1, then a 1×1×24 convolution expansion is performed to expand it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×48 convolution expansion is performed again to expand it to 32×32×48 to obtain the encoded watermark information.

[0053] The encoded watermark information with a size of 32×32×48 Reshaped to 128×128×3, the embedder applies a convolutional block to extract the watermark information reshaped to 128×128×3 The feature map to be embedded;

[0054] The feature map to be embedded is parallel to the carrier image sample c along the channel dimension. i Connect to form a 128×6×6 connection result;

[0055] The connection result is processed by the convolution block and the 1×1×3 convolution to form a marked image with a size of 128×128×3 and embedded with watermark information;

[0056] The marked image with the embedded watermark information is sent.

[0057] In some preferred embodiments, all of the convolution blocks have the same structure, consisting of a 1×1, a 3×3, and a 5×5 convolution and a residual connection; in the structure of the convolution block, each convolution has 32 filters; the 32-channel feature maps generated by different convolution paths are connected in series along the channel dimension to form a 96-channel feature, and a 1×1 convolution is applied to convert the 96-channel feature back to the original input channel size, and summed in the residual connection.

[0058] In some preferred embodiments, the first convolutional block used by the embedder requires intermediate results to calculate the loss, specifically:

[0059] The loss function is expressed as:

[0060]

[0061] where λ i ,i=1,2,3 are weight factors, is the restored watermark image, w i is the watermark image sample, m i is the marked image with embedded watermark information, c i is the carrier image sample, is the encoded watermark information, and ψ is a function for calculating correlation, specifically:

[0062]

[0063] Among them, g represents the Gram matrix containing all possible inner products, and B1 and B2 are convolution blocks with the same structure;

[0064] Through regularization, the proposed scheme objective is expressed as L(θ)+λ4P, where P is the penalty term for achieving robustness and λ4 is the weight that controls the strength of the regularization term; the deep neural network of the convolutional block needs to learn the parameter θ that minimizes L(θ)+λ4P. * :

[0065] θ * =argmin θ [L(θ)+λ4P].

[0066] The data receiving side receives the marked image through the communication channel;

[0067] Introducing the invariance layer composed of fully connected neural networks Dealing with distortion generated during transmission;

[0068] The extractor extracts the 128×3×3 watermark information from the labeled image by applying exactly the same two convolution blocks and a 1×1×3 convolution, which is the exact opposite process of the embedder.

[0069] The extracted watermark information of size 128×128×3 Reshaped to 32×32×48;

[0070] Reshape the watermark information into 32×32×48 As the input sample, it is first processed by two identical convolution blocks to generate an output sample of 32×32×48, then a 1×1×24 convolution is performed to tighten it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×1 convolution is performed to tighten it to 32×32×1, and the watermark information restored after decoding is obtained.

[0071] Verification is performed based on the restored watermark information to obtain a verification result.

[0072] In some preferred embodiments, the invariance layer composed of the fully connected neural network is introduced as follows:

[0073] By using a fully connected neural network, this invariance layer will learn a transformation from the space M to the over-complete space T, where the neurons are sparsely activated, i.e., the most important information in M ​​is redundantly projected into T, and the neural connections on the regions of M that are not related to the watermark are deactivated;

[0074] Based on the contraction autoencoder, the fully connected neural network uses a regularization term, which is obtained by the Frobenius norm of the Jacobian matrix of the output of one layer relative to the input. The regularization term P is:

[0075]

[0076] Among them, X i is the ith input, h j is the output of the jth hidden unit of the fully connected layer; the Jacobian matrix is ​​written as:

[0077]

[0078] Among them, A is the activation function, ω ji Yes j and X i The weight between them; setting A to a hyperbolic tangent tanh function with strong gradient and bias avoidance, P is finally calculated as:

[0079]

[0080] In order to better embed the watermark into the original carrier image to obtain a robust watermark image, auxiliary information used to restore the original carrier image, that is, the residual, is embedded to finally obtain the watermark image.

[0081] The original carrier image can be subtracted from the encoded image, and Huffman lossless compression can be performed to reduce zero pixel values ​​in the difference image. The obtained difference image and the encoded image are passed through a reversible neural network model of forward mapping, so as to embed the difference image into the encoded image.

[0082] The forward output of the carrier image branch of the forward-mapped reversible neural network model is used as the final embedded watermark image, and the forward output of the difference image branch of the forward-mapped reversible neural network model is defined as a constant matrix containing no valid information.

[0083] The final embedded watermark image and the constant matrix are passed through a reversible neural network model with reverse mapping; utilizing the reverse mapping of the reversible neural network, the reverse output of the carrier image branch of the reverse mapped reversible neural network model is set as the restored coding image, the reverse output of the difference image branch of the reverse mapped reversible neural network model is set as the output information, and the restored coding image is added to the output information to obtain the final restored carrier image.

[0084] If the watermarked image is attacked and the information is tampered with, the restored encoded image is passed through a convolutional neural network, a pooling layer, and a linear layer to obtain a restored watermark sequence again.

[0085] The method of the present invention adopts a reversible neural network to realize the reversibility of the watermark, so that the original carrier image can be restored.

[0086] Figure 2 An architectural diagram of an automated robust image watermark generation system based on a deep neural network provided in this application, the system comprising: a sending side and a receiving side.

[0087] The sending side includes:

[0088] The input space unit is used to give two input spaces, namely the watermark image sample w i and carrier image sample c i ;

[0089] The encoding unit is used to encode the watermark image sample w i First, it is processed by two identical convolution blocks to generate an output sample of 32×32×1, then a 1×1×24 convolution expansion is performed to expand it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×48 convolution expansion is performed again to expand it to 32×32×48 to obtain the encoded watermark information.

[0090] The feature extraction unit is used to convert the encoded watermark information of size 32×32×48 Reshape to 128×128×3, call the embedder, apply the convolutional block to extract the watermark information reshaped to 128×128×3 The feature map to be embedded;

[0091] A connection unit is used to connect the feature map to be embedded with the carrier image sample c along the channel dimension i Connect to form a 128×6×6 connection result;

[0092] An image generating unit, used for processing the connection result through a convolution block and a 1×1×3 convolution to form a marked image with a size of 128×128×3 and embedded with watermark information;

[0093] The sending unit is used to send the marked image embedded with the watermark information.

[0094] The receiving side includes:

[0095] A receiving unit, configured to receive a marked image via a communication channel;

[0096] Rectification unit, used to introduce the invariance layer of the fully connected neural network Dealing with distortion generated during transmission;

[0097] The watermark extraction unit is used to call the extractor to extract the 128×3×3 watermark information from the marked image by applying the same two convolution blocks and a 1×1×3 convolution, that is, the process is completely opposite to the embedder.

[0098] The watermark reshaping unit is used to reshape the extracted watermark information into a size of 128×128×3 Reshaped to 32×32×48;

[0099] The decoding unit is used to reshape the watermark information into 32×32×48 As the input sample, it is first processed by two identical convolution blocks to generate an output sample of 32×32×48, then a 1×1×24 convolution is performed to tighten it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×1 convolution is performed to tighten it to 32×32×1, and the watermark information restored after decoding is obtained.

[0100] The verification unit is used to perform verification according to the restored watermark information to obtain a verification result.

[0101] The present application provides an automated robust image watermark generation system based on a deep neural network, the system comprising: the system comprising a processor and a memory:

[0102] The memory is used to store program code and transmit the program code to the processor;

[0103] The processor is used to execute the method described in any one of all embodiments of the first aspect according to the instructions in the program code.

[0104] The present application provides a computer-readable storage medium, wherein the computer-readable storage medium is used to store program code, and the program code is used to be executed by a processor to implement the method described in any one of all the embodiments of the first aspect.

[0105] In a specific implementation, the present invention further provides a computer storage medium, wherein the computer storage medium may store a program, and when the program is executed, the program may include some or all of the steps in each embodiment of the present invention. The storage medium may be a disk, an optical disk, a read-only storage memory (abbreviated as: ROM) or a random access memory (abbreviated as: RAM), etc.

[0106] Those skilled in the art can clearly understand that the technology in the embodiments of the present invention can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution in the embodiments of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention or some parts of the embodiments.

[0107] The same and similar parts between the various embodiments of this specification can be referred to each other. In particular, for the embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the description in the method embodiment.

[0108] The above-described embodiments of the present invention do not limit the protection scope of the present invention.

Claims

1. An automated robust image watermark generation method based on a deep neural network, corresponding to the data sending side, characterized in that: The method comprises: Given two input spaces, namely watermark image samples w i and carrier image sample c i ; The watermark image sample w i First, it is processed by two identical convolution blocks to generate an output sample of 32×32×1, then a 1×1×24 convolution expansion is performed to expand it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×48 convolution expansion is performed again to expand it to 32×32×48 to obtain the encoded watermark information. The encoded watermark information with a size of 32×32×48 Reshaped to 128×128×3, the embedder applies a convolutional block to extract the watermark information reshaped to 128×128×3 The feature map to be embedded; The feature map to be embedded is parallel to the carrier image sample c along the channel dimension. i Connect to form a 128×6×6 connection result; The connection result is processed by the convolution block and the 1×1×3 convolution to form a marked image with a size of 128×128×3 and embedded with watermark information; The marked image with the embedded watermark information is sent.

2. The method according to claim 1, characterized in that: All of the convolution blocks have the same structure, consisting of a 1×1, a 3×3, and a 5×5 convolution and a residual connection; in the structure of the convolution block, each convolution has 32 filters; the 32-channel feature maps generated by different convolution paths are connected in series along the channel dimension to form a 96-channel feature, and a 1×1 convolution is applied to convert the 96-channel feature back to the original input channel size and summed in the residual connection.

3. The method according to claim 1, characterized in that: The first convolutional block used by the embedder requires intermediate results to calculate the loss, specifically: The loss function is expressed as: where λ i ,i=1,2,3 are weight factors, is the restored watermark image, w i is the watermark image sample, m i is a marked image with embedded watermark information, c i is the carrier image sample, is the encoded watermark information, and ψ is a function for calculating correlation, specifically: Among them, g represents the Gram matrix containing all possible inner products, and B1 and L2 are convolution blocks with the same structure; Through regularization, the proposed scheme objective is expressed as B(θ)+λ4P, where P is the penalty term for achieving robustness and λ4 is the weight that controls the strength of the regularization term; the deep neural network of the convolutional block needs to learn the parameter θ that minimizes L(θ)+λ4P. * : i * = argminθ[L(θ)+λ4P].

4. An automated robust image watermark generation method based on a deep neural network, corresponding to the data receiving side, characterized in that: The method comprises: receiving a tagged image via a communication channel; Introducing the invariance layer composed of fully connected neural networks Dealing with distortion generated during transmission; The extractor extracts the 128×3×3 watermark information from the labeled image by applying exactly the same two convolution blocks and a 1×1×3 convolution, which is the exact opposite process of the embedder. The extracted watermark information of size 128×128×3 Reshaped to 32×32×48; Reshape the watermark information into 32×32×48 As the input sample, it is first processed by two identical convolution blocks to generate an output sample of 32×32×48, then a 1×1×24 convolution is performed to tighten it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×1 convolution is performed to tighten it to 32×32×1, and the watermark information restored after decoding is obtained. Verification is performed based on the restored watermark information to obtain a verification result.

5. The method according to claim 4, characterized in that: The invariance layer composed of the fully connected neural network is specifically: By using a fully connected neural network, this invariance layer will learn a transformation from the space M to the over-complete space T, where the neurons are sparsely activated, i.e., the most important information in M ​​is redundantly projected into T, and the neural connections on the regions of M that are not related to the watermark are deactivated; Based on the contraction autoencoder, the fully connected neural network uses a regularization term, which is obtained by the Frobenius norm of the Jacobian matrix of the output of one layer relative to the input. The regularization term P is: Among them, X i is the ith input, h j is the output of the jth hidden unit of the fully connected layer; the Jacobian matrix is ​​written as: Among them, A is the activation function, ω ji Yes j and X i The weight between them; setting A to a hyperbolic tangent tanh function with strong gradient and bias avoidance, P is finally calculated as:

6. An automated robust image watermark generation system based on deep neural network, characterized in that: The system comprises: The input space unit is used to give two input spaces, namely the watermark image sample w i and carrier image sample c i ; The encoding unit is used to encode the watermark image sample w i First, it is processed by two identical convolution blocks to generate an output sample of 32×32×1, then a 1×1×24 convolution expansion is performed to expand it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×48 convolution expansion is performed again to expand it to 32×32×48 to obtain the encoded watermark information. The feature extraction unit is used to convert the encoded watermark information of size 32×32×48 Reshape to 128×128×3, call the embedder, apply the convolutional block to extract the watermark information reshaped to 128×128×3 The feature map to be embedded; A connection unit is used to connect the feature map to be embedded with the carrier image sample c along the channel dimension i Connect to form a 128×6×6 connection result; An image generating unit, used for processing the connection result through a convolution block and a 1×1×3 convolution to form a marked image with a size of 128×128×3 and embedded with watermark information; The sending unit is used to send the marked image embedded with the watermark information.

7. An automated robust image watermark generation system based on deep neural network, characterized in that: The system comprises: A receiving unit, configured to receive a marked image via a communication channel; Rectification unit, used to introduce the invariance layer of the fully connected neural network Dealing with distortion generated during transmission; The watermark extraction unit is used to call the extractor to extract the 128×3×3 watermark information from the marked image by applying the same two convolution blocks and a 1×1×3 convolution, that is, the process is completely opposite to the embedder. The watermark reshaping unit is used to reshape the extracted watermark information into a size of 128×128×3 Reshaped to 32×32×48; The decoding unit is used to reshape the watermark information into 32×32×48 As the input sample, it is first processed by two identical convolution blocks to generate an output sample of 32×32×48, then a 1×1×24 convolution is performed to tighten it to 32×32×24, and then a 32×32×24 output sample is generated by two identical convolution blocks as above, and then a 1×1×1 convolution is performed to tighten it to 32×32×1, and the watermark information restored after decoding is obtained. The verification unit is used to perform verification according to the restored watermark information to obtain a verification result.

8. An automated robust image watermark generation system based on deep neural network, characterized in that: The system comprises a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the method according to any one of claims 1 to 5 according to the instructions in the program code.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store program codes, and the program codes are used to be executed by a processor to implement the method according to any one of claims 1 to 5.

Images