Anti-printing digital watermarking method and device based on U-Net network and DFT optimal quality radius

Abstract

The present invention discloses a method and device for resisting printing digital watermarking based on the combination of U-Net network and DFT optimal quality radius, the method comprising the following steps: (1) obtaining the Y channel spectrum of the carrier image; (2) selecting the domain to be embedded with the watermark; (3) reorganizing the domain to be embedded with the watermark, passing it into the Encoder network, sampling and outputting the dense matrix, replacing the corresponding parameters of the original spectrum, and generating a dense image; (4) distortion simulation; (5) constructing a Decoder network, adding a spatial transformation network to process the perspective distortion; (6) training multiple times to obtain a complete encoding and decoding network. The present invention aims at the problem that the quality of the dense image generated after the carrier image containing a large low-frequency area is poor after embedding the watermark. By limiting the Fourier domain embedding radius range operation before network training, the quality of the generated dense image is improved, the search space of the network is reduced, and the efficiency of network training is improved. At the same time, the distortion simulation such as perspective distortion and color transformation is used to improve the extraction rate and accuracy of the watermark after printing and scanning.

Description

Technical Field

[0001] The invention relates to an anti-printing digital watermarking method based on the combination of a U-Net network and a DFT optimal quality radius, and belongs to the technical field. Background Art

[0002] Digital watermarking is an important branch of information hiding technology research. It uses the characteristics of human auditory and visual systems to add some additional information to the information carrier to protect the copyright of digital works. In many cases, important documents, certificates, etc. need to be printed as paper products, so there is a demand for digital watermarking technology in printing and scanning scenarios. However, existing anti-printing watermarking technology faces great challenges: since image printing and scanning are subject to pixel distortion caused by digital halftone, dot gain, gamma correction and quantization, as well as a series of image geometric transformations during shooting and extraction, most digital watermarking algorithms cannot meet increasingly complex needs.

[0003] In recent years, the research and development of deep learning has entered a period of explosive growth, and many research results have been produced in fields such as computer vision. Compared with traditional methods that are manually designed based on prior knowledge, deep learning has a powerful feature learning ability. Researchers introduced deep learning into image steganography to allow the network to learn more covert steganographic behaviors. In 2019, Duan et al. used a fully convolutional network to construct a U-Net structured encoding network. Since the U-Net structured network can better integrate global features with local features, it can generate encrypted images with better image quality, but when the image has a large number of low-frequency areas, the effect is not ideal. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and provide an anti-printing digital watermarking method and device based on the combination of U-Net network and DFT optimal quality radius, which can solve the technical problem of embedding anti-printing watermarks into carrier images containing large low-frequency areas.

[0005] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0006] In a first aspect, the present invention provides an anti-printing digital watermarking method based on a combination of a U-Net network and a DFT optimal quality radius, comprising the following steps:

[0007] Obtain a carrier image, perform YCbCr and DFT transformation on the carrier image, move the zero-frequency component to the center of the spectrum, and obtain a spectrum diagram of the carrier image;

[0008] Obtain watermark information, and read the watermark information as a watermark 01 character string;

[0009] Selecting a region to be embedded with a watermark in the carrier image spectrum diagram, and recombining the values ​​in the region to be embedded with a watermark to generate a carrier matrix;

[0010] Combine the carrier matrix generated by reorganizing the values ​​in the area to be embedded in the watermark and the 01 string of the watermark to form an input tensor;

[0011] Input the input tensor into the Encoder network (encoding network) to obtain a dense matrix; the Encoder network is a watermark encoding network based on U-Net;

[0012] The corresponding carrier matrix in the carrier image spectrum is replaced by the dense matrix to generate a dense image spectrum, and the dense image spectrum is subjected to inverse amplitude, inverse Fourier and inverse YCbCr transformation to generate a dense image.

[0013] Further, a method of obtaining a carrier image, performing YCbCr and DFT transformation on the carrier image, and moving the zero-frequency component to the center of the spectrum to obtain a spectrum diagram of the carrier image includes:

[0014] Perform YCbCr transformation on the carrier image, take the Y channel for normalization, and obtain the normalized image Y channel matrix;

[0015] Perform Fourier transform on the normalized image Y channel matrix to obtain the spectrum diagram;

[0016] The zero-frequency component of the spectrum graph is moved to the center to obtain a carrier image spectrum graph.

[0017] Furthermore, the method of selecting the area to be embedded with the watermark in the carrier image spectrum diagram includes:

[0018] Taking the center of the carrier image spectrum as the center of the circle, radiating outwards watermark embeddable areas of different radii;

[0019] The row vector of the watermark 01 string is embedded into the corresponding matrix using formula (1) according to the radius:

[0020]

[0021] In the formula, W(x i ,y i ) is the watermark matrix, v(j) is the jth element of the row vector, M(x i ,y j ) is the pixel point of the carrier image, and the coordinate (x i ,y i ) is defined as:

[0022]

[0023]

[0024] Among them, m and n represent the size of the carrier image M matrix, and r represents the embedding radius.

[0025] The watermark matrix is ​​embedded into the carrier image using the following formula:

[0026] M W (x,y)=M(x,y)+α*W(x,y)

[0027] Among them, M(x,y) is the pixel point of the carrier image, W(x,y) is the corresponding value of the watermark matrix, and M w It is a secret image in the preprocessing stage.

[0028] At the same time, a search is performed within the range of ±10 of the radius of the watermark embedded in the encrypted image, and the amplitude coefficient is extracted as the row vector of the extracted watermark and the length is adjusted according to the formula:

[0029] l=(r+10)*π (4)

[0030] Normalize the row vector to the interval [0, 1] and calculate the cross covariance with the original watermark vector. It is defined as:

[0031]

[0032] Among them, C rv is the cross covariance, * is the complex conjugate, N is the length of the watermark vector, and |m| is the average value of m.

[0033] Find the radius of the encrypted image with a PSNR value higher than the average and a cross-covariance coefficient greater than 0.3 to form a circular ring area as the watermark area to be embedded.

[0034] Furthermore, the encoder network is based on U-Net, wherein the input parameter of U-Net is an input tensor, a dense matrix is ​​generated by seven downsampling and upsampling, and a discriminator network composed of multiple convolutional and maximum pooling layers is included;

[0035] The discriminator network uses Wasserstein loss as a supervisory signal to classify and train based on the mirflickr dataset. The training method includes:

[0036] Use L 2 Residual regularization loss L R , LPIPS perceptual loss L P , the loss of the discriminator is L C And the watermark information entropy loss L M The weighted sum of is used as the training loss:

[0037] L=λ R L R+λ P L P +λ C L C +λ M L M (6).

[0038] Furthermore, the method further comprises extracting watermark information from the encrypted image spectrum, comprising the following steps:

[0039] Obtaining a secret image to be identified;

[0040] The encrypted image to be identified is input into a trained decoding network (Decoder network) to obtain watermark information; the decoding network includes a spatial transformation network and a CNN network.

[0041] Furthermore, the training method of the decoding network includes:

[0042] Obtain multiple sets of generated encrypted image spectrograms and their watermark information;

[0043] Performing distortion simulation on the spectrum of the encrypted image to obtain a distorted encrypted image;

[0044] Pairing the distorted encrypted image and its watermark information to form a training set;

[0045] The decoding network is trained using the training set to obtain a trained decoding network.

[0046] Furthermore, the distortion transformation method includes one or more of the following methods:

[0047] Perspective distortion: randomly transform the four corner points of the image within a fixed range, find the homography from the original corner points to the new positions, perform bilinear resampling on the original image, and create a perspective distorted image;

[0048] Color transformation: The limited color gamut problem in the printing process is simulated by random affine color transformation, and the color of each pixel is shifted by [-0.1, 0.1];

[0049] JPEG compression: Emulate quantization in JPEG compression using piecewise functions of Shin and Song.

[0050] In a second aspect, the present invention provides an anti-printing digital watermarking device based on a combination of a U-Net network and a DFT optimal quality radius, comprising:

[0051] Spectrum conversion module: used to obtain the carrier image, perform YCbCr and DFT transformation on the carrier image, move the zero-frequency component to the center of the spectrum, and obtain the carrier image spectrum;

[0052] Watermark reading module: used to obtain watermark information and read the watermark information as a watermark 01 character string;

[0053] Carrier matrix module: used to select the area to be embedded with watermark in the carrier image spectrum diagram, and reorganize the values ​​in the area to be embedded with watermark to generate a carrier matrix;

[0054] Tensor combination module: used to combine the carrier matrix generated by reorganizing the values ​​in the area to be embedded in the watermark and the 01 string of the watermark to form an input tensor;

[0055] Encoding module: used to input the input tensor into the Encoder network to obtain a dense matrix; the Encoder network is a watermark encoding network based on U-Net;

[0056] Output module: used to replace the corresponding carrier matrix in the carrier image spectrum with the dense matrix to generate the dense image spectrum, and perform inverse amplitude, inverse Fourier and inverse YCbCr transformation on the dense image spectrum to generate the dense image.

[0057] Furthermore, the device also includes a watermark extraction module: used to obtain a secret image to be identified, and input the secret image to be identified into a trained decoding network to obtain watermark information; the decoding network includes a spatial transformation network and a CNN network.

[0058] In a third aspect, the present invention provides an anti-printing digital watermarking device based on a combination of a U-Net network and a DFT optimal quality radius, including a processor and a storage medium;

[0059] The storage medium is used to store instructions;

[0060] The processor is used to operate according to the instructions to execute the steps of the method described in the first aspect.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. The method of limiting the frequency domain embedding range greatly improves the quality of encrypted images, and the preprocessing reduces the network's search space and training time;

[0063] 2. Use the U-Net network structure as the generating network, so that the generated image can keep the characteristics of the carrier image as much as possible, and further improve the quality of the encrypted image;

[0064] 3. Add perspective distortion, color transformation, JPEG compression and other distortion simulations to improve the robustness of encrypted images;

[0065] 4. A spatial transformation network is added to the decoder to solve the problem of perspective distortion during shooting and scanning, further improving the robustness of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 is a flow chart of the present invention;

[0067] Figure 2 is a spectrum diagram of the carrier image in the present invention;

[0068] Figure 3 Demonstration of the steps of the present invention;

[0069] Figure 4 is the Encoder network model of the present invention;

[0070] Figure 5 is the Decoder network model of the present invention;

[0071] Figure 6 It is a comparison chart of the experimental results of the present invention. DETAILED DESCRIPTION

[0072] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0073] Embodiment 1:

[0074] This embodiment provides an anti-printing digital watermark method based on the combination of U-Net network and DFT optimal quality radius, including the following steps:

[0075] Obtain a carrier image, perform YCbCr and DFT transformation on the carrier image, move the zero-frequency component to the center of the spectrum, and obtain a spectrum diagram of the carrier image;

[0076] Obtain watermark information, and read the watermark information as a watermark 01 character string;

[0077] Selecting a region to be embedded with a watermark in the carrier image spectrum diagram, and recombining the values ​​in the region to be embedded with a watermark to generate a carrier matrix;

[0078] Combine the carrier matrix generated by reorganizing the values ​​in the area to be embedded in the watermark and the 01 string of the watermark to form an input tensor;

[0079] Input the input tensor into the Encoder network (encoding network) to obtain a dense matrix; the Encoder network is a watermark encoding network based on U-Net;

[0080] The corresponding carrier matrix in the carrier image spectrum is replaced by the dense matrix to generate a dense image spectrum, and the dense image spectrum is subjected to inverse amplitude, inverse Fourier and inverse YCbCr transformation to generate a dense image.

[0081] This method uses Fourier spectrum, PSNR and covariance coefficient to select the area to be embedded with watermark. Based on U-Net, the encoding network is constructed to generate a dense matrix, replace the corresponding parameters in the carrier image spectrum, and generate a dense image after inverse transformation. By simulating perspective distortion, color transformation, JPEG compression and other distortion simulations in the dense image, the robustness of anti-printing watermark is improved. Based on CNN and spatial transformation network, the decoding network is constructed to further improve the robustness of anti-printing watermark.

[0082] like Figure 1 Shown is a flow chart of the present invention, and the detailed steps are as follows:

[0083] (1) Perform YCbCr and DFT transformation on the carrier image to move the zero-frequency component to the center of the spectrum

[0084] In order to solve the problem of poor quality of dense images in existing deep learning networks when the carrier image contains a large number of low-frequency areas, this method first performs YCbCr transformation on the carrier image before inputting it into the U-Net network, and takes its Y channel for DFT transformation. Since embedding in the low-frequency area will have a greater impact on the quality of the carrier image, the zero-frequency component is moved to the center. Figure 2 This is the spectrum diagram of the carrier image in this embodiment.

[0085] (2) Select the area to be embedded with watermark

[0086] First, embed the watermark into the carrier image using the traditional method, obtain the position with better effect and record and filter it. Finally, only a radius range is obtained, that is, the area to be embedded with the watermark. The specific steps are as follows:

[0087] The row vector of the watermark 01 string is embedded into the corresponding matrix using formula (1) according to the radius:

[0088]

[0089] In the formula, W(x i ,y i ) is the watermark matrix, v(j) is the jth element of the row vector, M(x i ,y j ) is the pixel point of the carrier image, and the coordinate (x i ,y i ) is defined as:

[0090]

[0091]

[0092] Among them, m and n represent the size of the carrier image M matrix, and r represents the embedding radius.

[0093] The watermark matrix is ​​embedded into the carrier image using the following formula:

[0094] M W (x,y)=M(x,y)+α*W(x,y)

[0095] Among them, M(x,y) is the pixel point of the carrier image, W(x,y) is the corresponding value of the watermark matrix, and M w It is a secret image in the preprocessing stage.

[0096] At the same time, a search is performed within the range of ±10 of the radius of the watermark embedded in the encrypted image, and the amplitude coefficient is extracted as the row vector of the extracted watermark and the length is adjusted according to formula (5):

[0097] l=(r+10)*π (4)

[0098] Normalize the row vector to the interval [0, 1] and calculate the cross covariance with the original watermark vector. It is defined as:

[0099]

[0100] Among them, C rv is the cross covariance and * is the complex conjugate.

[0101] Find the radius of the encrypted image with a PSNR value higher than the average and a cross-covariance coefficient greater than 0.3 to form a ring domain as the embedding domain.

[0102] (3) Construct an Encoder network based on U-Net and inversely transform the output parameters to generate a encrypted image

[0103] Take the ring embedding domain value to generate an m*n matrix cover (carrier matrix), transform the watermark row vector into an m*n matrix, and build an Encoder network based on U-Net, where the input parameter of U-Net is an m×n×2 tensor. Through seven downsampling and upsampling, a dense matrix with richer cover features is generated. In order to obtain the perceptual loss loss of the encoder / decoder channel, a discriminator network composed of a series of convolutional pools and maximum pools is introduced. The network uses Wasserstein loss as a supervision signal to classify the Encoder input and output for training. In order to minimize the perceptual loss of the Encoder process, L 2 Residual regularization loss L R , LPIPS perceptual loss L P , the loss of the discriminator is L C And the watermark information entropy loss L M The weighted sum of is used as the training loss:

[0104] L=λ R LR +λ P L P +λ C L C +λ M L M (6)

[0105] The inverse transformation of the dense matrix generates a ring, replaces the corresponding amplitude value in the original carrier image, and performs inverse amplitude value and inverse Fourier transform to generate a dense image.

[0106] Figure 4 It is the Encoder network model of this method; Figure 5 It is the Decoder network model of this method.

[0107] (5) Perform distortion simulations such as perspective distortion, color transformation, and JPEG compression on encrypted images

[0108] In order to enhance the robustness of anti-printing watermark, a series of distortion transformations simulating the actual printing scanning process are added between the Encoder network and the Decoder network, including

[0109] Perspective distortion: Randomly transform the four corner points of the image within a fixed range, find the homography from the original corner points to the new positions, perform bilinear resampling on the original image, and create a perspective distorted image.

[0110] Color transformation: Random affine color transformation is used to simulate the limited color gamut problem in printing. Each pixel is shifted by [-0.1, 0.1].

[0111] JPEG compression: Emulate quantization in JPEG compression using piecewise functions of Shin and Song.

[0112]

[0113] (6) Build a decoding network based on CNN and spatial transformation network to extract watermark information

[0114] Due to the good image reconstruction ability of CNN, CNN is used as the decoding network. At the same time, in order to solve the perspective distortion in the scanning process, a spatial transformation network is added to perform perspective transformation.

[0115] (7) Finally, the trained network embeds secret information in a limited frequency domain to generate a secret image.

[0116] After the network training is completed, the generated network is used to obtain the secret image. On the extraction side, the secret information is losslessly extracted.

[0117] The overall data processing process of the present invention is as follows Figure 3shown.

[0118] Figure 6 This is a comparison chart of the experimental results of this method. (The secret carrier in the figure is the secret image).

[0119] In summary, the present invention uses Fourier spectrum, PSNR and covariance coefficient to select the area to be embedded with watermark. Based on U-Net, the encoding network is constructed to generate a dense matrix, replace the corresponding parameters in the carrier image spectrum, and generate a dense image after inverse transformation. By simulating perspective distortion, color transformation, JPEG compression and other distortion simulations in the dense image, the robustness of anti-printing watermark is improved. Based on CNN and spatial transformation network, the decoding network is constructed to further improve the robustness of anti-printing watermark.

[0120] Embodiment 2:

[0121] This embodiment provides an anti-printing digital watermarking device based on a combination of a U-Net network and a DFT optimal quality radius, including:

[0122] Spectrum conversion module: used to obtain the carrier image, perform YCbCr and DFT transformation on the carrier image, move the zero-frequency component to the center of the spectrum, and obtain the carrier image spectrum;

[0123] Watermark reading module: used to obtain watermark information and read the watermark information as a watermark 01 character string;

[0124] Carrier matrix module: used to select the area to be embedded with watermark in the carrier image spectrum diagram, and reorganize the values ​​in the area to be embedded with watermark to generate a carrier matrix;

[0125] Tensor combination module: used to combine the carrier matrix generated by reorganizing the values ​​in the area to be embedded in the watermark and the 01 string of the watermark to form an input tensor;

[0126] Encoding module: used to input the input tensor into the Encoder network to obtain a dense matrix; the Encoder network is a watermark encoding network based on U-Net;

[0127] Output module: used to replace the corresponding carrier matrix in the carrier image spectrum with the dense matrix to generate the dense image spectrum, and perform inverse amplitude, inverse Fourier and inverse YCbCr transformation on the dense image spectrum to generate the dense image.

[0128] Watermark extraction module: used to obtain the encrypted image to be identified, and input the encrypted image to be identified into a trained decoding network to obtain watermark information; the decoding network includes a spatial transformation network and a CNN network.

[0129] The device of this embodiment can be used to implement the method described in the first embodiment.

[0130] Embodiment three:

[0131] This embodiment provides an anti-printing digital watermarking device based on a combination of a U-Net network and a DFT optimal quality radius, including a processor and a storage medium;

[0132] The storage medium is used to store instructions;

[0133] The processor is used to operate according to the instructions to execute the steps of the method described in embodiment 1.

[0134] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0135] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0136] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0137] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0138] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An anti-printing digital watermarking method based on the combination of U-Net network and DFT optimal quality radius, It is characterized in that The following steps are involved: Obtain a carrier image, perform YCbCr and DFT transformation on the carrier image, move the zero-frequency component to the center of the spectrum, and obtain a spectrum diagram of the carrier image; Obtain watermark information, and read the watermark information as a watermark 01 character string; Selecting a region to be embedded with a watermark in the carrier image spectrum diagram, and recombining the values ​​in the region to be embedded with a watermark to generate a carrier matrix; Combine the carrier matrix generated by reorganizing the values ​​in the area to be embedded in the watermark and the 01 string of the watermark to form an input tensor; Input the input tensor into the Encoder network to obtain a dense matrix; the Encoder network is a watermark encoding network based on U-Net; The corresponding carrier matrix in the carrier image spectrum is replaced by the dense matrix to generate a dense image spectrum, and the dense image spectrum is subjected to inverse amplitude, inverse Fourier and inverse YCbCr transformation to generate a dense image.

2. The anti-printing digital watermark method according to claim 1, It is characterized in that The method of obtaining a carrier image, performing YCbCr and DFT transformation on the carrier image, moving the zero-frequency component to the center of the spectrum, and obtaining a spectrum diagram of the carrier image includes: Perform YCbCr transformation on the carrier image, take the Y channel for normalization, and obtain the normalized image Y channel matrix; Perform Fourier transform on the normalized image Y channel matrix to obtain the spectrum diagram; The zero-frequency component of the spectrum graph is moved to the center to obtain a carrier image spectrum graph.

3. The anti-printing digital watermark method according to claim 1, It is characterized in that The method for selecting a region to be embedded with a watermark in the carrier image spectrum diagram comprises: Taking the center of the carrier image spectrum as the center of the circle, radiating outwards watermark embeddable areas of different radii; The row vector of the watermark 01 string is embedded into the corresponding matrix using formula (1) according to the radius: In the formula, W(x i ,y i ) is the watermark matrix, v(j) is the jth element of the row vector, M(x i ,y j ) is the pixel point of the carrier image, and the coordinate (x i ,y i ) is defined as: Among them, m and n represent the size of the carrier image M matrix, and r represents the embedding radius; The watermark matrix is ​​embedded into the carrier image using the following formula: M W (x,y)=M(x,y)+α*W(x,y) Among them, M(x,y) is the pixel point of the carrier image, W(x,y) is the corresponding value of the watermark matrix, and M w It is the encrypted image in the pre-processing stage; At the same time, a search is performed within the range of ±10 of the radius of the watermark embedded in the encrypted image, and the amplitude coefficient is extracted as the row vector of the extracted watermark and the length l is adjusted according to the formula: l＝(r+10)*π (4) Normalize the row vector to the interval [0, 1] and calculate the cross covariance with the original watermark vector; it is defined as: Among them, C rv is the cross covariance, * is the complex conjugate, N is the length of the watermark vector, |m| is the average value of m; Find the radius of the encrypted image with a PSNR value higher than the average and a cross-covariance coefficient greater than 0.3 to form a circular ring area as the watermark area to be embedded.

4. The anti-printing digital watermark method according to claim 1, It is characterized in that The encoder network is based on U-Net, where the input parameter of U-Net is the input tensor, and a dense matrix is ​​generated by seven downsampling and upsampling, including a discriminator network composed of multiple convolution pools and maximum pools; The discriminator network uses Wasserstein loss as a supervisory signal to classify and train based on the mirflickr dataset. The training method includes: Use L 2 Residual regularization loss L R , LPIPS perceptual loss L P , the loss of the discriminator is L C And the watermark information entropy loss L M The weighted sum of is used as the training loss: L=λ R L R +λ P L P +λ C L C +λ M L M (6)。 5. The anti-printing digital watermark method according to claim 1, It is characterized in that The method further comprises extracting watermark information of the encrypted image spectrum, comprising the following steps: Obtaining a secret image to be identified; The encrypted image to be identified is input into a trained decoding network to obtain watermark information; the decoding network includes a spatial transformation network and a CNN network.

6. The anti-printing digital watermark method according to claim 5, It is characterized in that The training method of the decoding network includes: Obtain multiple sets of generated encrypted image spectrograms and their watermark information; Performing distortion simulation on the spectrum of the encrypted image to obtain a distorted encrypted image; Pairing the distorted encrypted image and its watermark information to form a training set; The decoding network is trained using the training set to obtain a trained decoding network.

7. The anti-printing digital watermark method according to claim 6, It is characterized in that The distortion transformation method includes one or more of the following methods: Perspective distortion: randomly transform the four corner points of the image within a fixed range, find the homography from the original corner points to the new positions, perform bilinear resampling on the original image, and create a perspective distorted image; Color transformation: The limited color gamut problem in the printing process is simulated by random affine color transformation, and the color of each pixel is shifted by [-0.1, 0.1]; JPEG compression: Emulate quantization in JPEG compression using piecewise functions of Shin and Song. 8.An anti-printing digital watermarking device based on the combination of U-Net network and DFT optimal quality radius, It is characterized in that include: Spectrum conversion module: used to obtain the carrier image, perform YCbCr and DFT transformation on the carrier image, move the zero-frequency component to the center of the spectrum, and obtain the carrier image spectrum; Watermark reading module: used to obtain watermark information and read the watermark information as a watermark 01 character string; Carrier matrix module: used to select the area to be embedded with watermark in the carrier image spectrum diagram, and reorganize the values ​​in the area to be embedded with watermark to generate a carrier matrix; Tensor combination module: used to combine the carrier matrix generated by reorganizing the values ​​in the area to be embedded in the watermark and the 01 string of the watermark to form an input tensor; Encoding module: used to input the input tensor into the Encoder network to obtain a dense matrix; the Encoder network is a watermark encoding network based on U-Net; Output module: used to replace the corresponding carrier matrix in the carrier image spectrum with the dense matrix to generate the dense image spectrum, and perform inverse amplitude, inverse Fourier and inverse YCbCr transformation on the dense image spectrum to generate the dense image.

9. The anti-printing digital watermark device according to claim 8, It is characterized in that The device also includes a watermark extraction module: used for acquiring a secret image to be identified, and inputting the secret image to be identified into a trained decoding network to obtain watermark information; the decoding network includes a space transformation network and a CNN network. 10.An anti-printing digital watermarking device based on the combination of U-Net network and DFT optimal quality radius, It is characterized in that including processor and storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 7.

Images