Y-Net based digital watermarking method, device and system

Abstract

The present invention discloses a method, apparatus and system for digital watermarking based on Y-Net. The method includes the steps of inputting a carrier image and a watermark image into a Y-Net generation network to generate a watermarked image, inputting the watermarked image into a pre-trained JPEG compression loss simulation network to obtain a simulated compressed watermarked image, using a pre-trained discrimination network to discriminate the simulated compressed watermarked image and obtain a watermark discrimination result, inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image, calculating the total loss function of the whole watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discrimination network, optimizing the watermark network with the goal of minimizing the loss function, and when the loss is reduced and maintained in a stable state, the training is considered to be completed. During the training process, the weights of the discrimination network and the compression loss simulation network are fixed, and the present invention has high watermarking capacity, high concealment, high security and strong robustness compared with traditional algorithms.

Description

[Technical field]

[0001] The present invention relates to the technical field of digital watermarking, and in particular to a method, apparatus and system for digital watermarking based on Y-Net. [Background technology]

[0002] With the rapid development of Internet technology, more and more digital works are being circulated over the network. While this makes communication easier, it also tests the originality of works, making copyright protection a very important issue for people.

[0003] Digital watermarking technology can protect the originality of a work according to the characteristics of the work itself. In recent years, a number of digital watermarking models based on deep learning have emerged, which can not only expand the watermark capacity but also effectively improve the security of transmission. However, these methods still have issues such as insufficient image quality and analysis resistance security of watermarked images, and weak robustness against JPEG compression attacks. Summary of the Invention [Problem to be solved by the invention]

[0004] In view of the above problems, the present invention proposes a method, apparatus and system for digital watermarking based on Y-Net, which has high watermark capacity, high concealment ability, high security and strong robustness. [Means for solving the problem]

[0005] In order to achieve the above technical object and achieve the above technical effects, the present invention is realized through the following technical means.

[0006] In a first aspect, the present invention provides a method for producing a composition comprising the steps of: inputting the carrier image and the watermark image into a Y-Net generating network to obtain a watermarked image in which the watermark image is embedded; inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result including identifying the watermarked image as a carrier image or a watermarked image; inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; Calculating a total loss function of the entire watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss is reduced and remains stable, the watermark network including a Y-Net generating network, a lossy simulation network, a discriminant network and an extracting network, and the weights of the discriminant network and the lossy simulation network are fixed during the optimization process; and generating a watermarked image based on the carrier image and the watermark image using the trained watermark network, compressing the watermarked image, and reconstructing the watermark image from the compressed watermarked image. We propose a digital watermarking method based on Y-Net, which includes

[0007] Alternatively, the Y-Net generating network is a skip connection structure, and includes: a first convolution operation group of carriers, a second convolution operation group of carriers, a third convolution operation group of carriers, a fourth convolution operation group of carriers, a fifth convolution operation group of carriers, a sixth convolution operation group of carriers, a first convolution operation group of watermarks, a second convolution operation group of watermarks, a third convolution operation group of watermarks, a fourth convolution operation group of watermarks, a fifth convolution operation group of watermarks, a sixth convolution operation group of watermarks, a seventh convolution operation group, a first deconvolution operation group, a second deconvolution operation group, a third deconvolution operation group, a fourth deconvolution operation group, a fifth deconvolution operation group, a sixth deconvolution operation group, a seventh deconvolution operation group, and an eighth convolution operation group; The output of the sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are fused with the output of the first deconvolution operation group; The output of the fifth convolution operation group of the carrier and the fifth convolution operation group of the watermark are fused with the output of the second deconvolution operation group; The output of the fourth convolution operation group of the carrier and the fourth convolution operation group of the watermark are fused with the output of the third deconvolution operation group; The outputs of the third convolution operation group of the carrier and the third convolution operation group of the watermark are fused with the output of a fourth deconvolution operation group; The outputs of the second convolution operation group of the carrier and the second convolution operation group of the watermark are fused with the output of a fifth deconvolution operation group; The outputs of the first convolution operation group of the carrier and the first convolution operation group of the watermark are fused with the output of a sixth deconvolution operation group; The sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are input to a seventh convolution operation group after channel stitching, and the output end of the seventh convolution operation group is connected to the first deconvolution operation group; The results after skip connection of the sixth deconvolution operation group, the first convolution operation group of the carrier, and the first convolution operation group of the watermark are input to a seventh deconvolution operation group; The input terminal of the eighth convolution operation group is input to the seventh deconvolution operation group, and the output is a watermarked image; A convolution operation group includes a convolution layer, an activation layer, and a batch normalization layer arranged in sequence, and a deconvolution operation group includes a deconvolution layer, an activation layer, and a batch normalization layer arranged in sequence.

[0008] Alternatively, the discrimination network is obtained through the following training steps: collecting a plurality of image samples and determining a label for each sample image, the label comprising a probability of a watermark image being contained in a corresponding sample image; A step of training Zhu-Net using each image sample as input and the label of each image sample as output to obtain a discrimination network.

[0009] Alternatively, the compression loss simulation network is a U-Net network.

[0010] Alternatively, the extraction network is a U-Net++ network.

[0011] Alternatively, the discrimination network is a discriminator, the discriminator including a softmax function.

[0012] Alternatively, the total loss function of the watermark network is:

[0013]

number

[0014] Alternatively, the formula for the Y-Net generated network loss is:

[0015]

number

number

[0016] 5. Extract network loss L s The calculation formula is as follows:

[0017]

number

number

[0018] 2. The discrimination network loss L d The calculation formula is as follows:

[0019]

number

[0020] In a second aspect, the present invention provides a method for producing a composition comprising the steps of: a first processing unit for inputting the carrier image and the watermark image into a Y-Net generating network to obtain a watermarked image in which the watermark image is embedded; a second processing unit for inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; a discrimination unit for discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result, the discrimination result including identifying the watermarked image as a carrier image or a watermarked image; a reconstruction unit for inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; a training unit for calculating a total loss function of the watermark network according to the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss is reduced and maintained in a stable state (the watermark network includes a Y-Net generating network, a loss simulation network, a discriminant network and an extracting network, and during the optimization process, the weights of the discriminant network and the loss simulation network are fixed); a watermarking unit for generating a watermarked image based on a carrier image and a watermark image using the trained watermarking network, compressing the watermarked image, and reconstructing the watermarked image from the compressed watermarked image; We provide a robust watermarking device based on Y-Net equipped with

[0021] In a third aspect, the present invention provides a strong and robust watermarking system based on Y-Net, comprising: a storage medium; and a processor; the storage medium is adapted to store instructions; The processor is adapted to operate in accordance with the instructions to carry out the method according to any one of the first aspects. Effect of the Invention

[0022] Compared with the prior art, the advantages of the present invention include: We design a feature-rich branch generating network, Y-Net, to improve the visual quality of watermarked images. For the first time, the U-Net++ structure is introduced in the field of image watermarking, and it can be used as an extraction network to improve the visual quality of the reconstructed image of the watermark. Considering the image quality loss caused by JPEG compression during network transmission, we use a U-Net structure network to simulate JPEG compression loss and improve the watermark network's resistance to JPEG compression. The Zhu-Net network structure is used as the discrimination network to monitor the security of watermarked images in real time. When training the watermark network, the weights of the discriminant network and the compression loss simulation network are fixed, and a one-way adversarial training method is used to accelerate the convergence speed of the watermark network, which not only improves the image quality but also ensures the security of the watermarked image.

[0023] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS In order to make the present invention easier to understand clearly, the present invention will be described in more detail below based on specific embodiments in conjunction with the drawings. [Brief description of the drawings]

[0024] [Figure 1] 1 is a flowchart of a digital watermarking method based on Y-Net according to an embodiment. [Diagram 2] FIG. 1 is a schematic diagram of a Y-Net generating network according to one embodiment. [Diagram 3] FIG. 1 is a schematic diagram of a compression loss simulation network according to an embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0025] In order to make the purpose, technical means and advantages of the present application clearer, the present application will be described in more detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to interpret the present application, and do not limit the present application.

[0026] An "embodiment" referred to in this specification means that the particular features, structures, or characteristics depicted in the embodiment may be combined and included in at least one embodiment of the present application. The appearance of the phrase "in one embodiment" in various places in this specification does not necessarily all refer to the same embodiment, and separate or alternative embodiments are not necessarily mutually exclusive with other embodiments. Those skilled in the art will understand, both explicitly and implicitly, that the embodiments depicted in this specification can be combined with other embodiments.

[0027] The principle of use of the present invention will be described in detail below with reference to the drawings.

[0028] (First embodiment) The embodiment of the present invention provides a digital watermarking method based on Y-Net, as shown in FIG. 1, and specifically includes the following steps:

[0029] S10: inputting the carrier image and the watermark image into a Y-Net generating network to obtain a watermarked image with the watermark image embedded therein, i.e., embedding the watermark image into the carrier image to generate a watermarked image; In a specific embodiment of the present invention, as shown in FIG. 3, the Y-Net generating network is a skip connection structure, including a first convolution operation group Conv1_c of carriers, a second convolution operation group Conv2_c of carriers, a third convolution operation group Conv3_c of carriers, a fourth convolution operation group Conv4_c of carriers, a fifth convolution operation group Conv5_c of carriers, a sixth convolution operation group Conv6_c of carriers, a first convolution operation group Conv1_s of watermarks, a second convolution operation group Conv2_s of watermark ... fifth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_c of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks, a sixth convolution operation group Conv6_s of watermarks the first convolution operation group Conv1_s, the second convolution operation group Conv2_s, the third convolution operation group Conv3_s, the fourth convolution operation group Conv4_s, the fifth convolution operation group Conv5_s, the sixth convolution operation group Conv6_s, the seventh convolution operation group Conv7, the first deconvolution operation group ConvT1, the second deconvolution operation group ConvT2, the third deconvolution operation group ConvT3, the fourth deconvolution operation group ConvT4, the fifth deconvolution operation group ConvT5, the sixth deconvolution operation group ConvT6, the seventh deconvolution operation group ConvT7, and the eighth convolution operation group Conv8; The output of the sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are fused with the output of the first deconvolution operation group; The output of the fifth convolution operation group of the carrier and the fifth convolution operation group of the watermark are fused with the output of the second deconvolution operation group; The output of the fourth convolution operation group of the carrier and the fourth convolution operation group of the watermark are fused with the output of the third deconvolution operation group; The outputs of the third convolution operation group of the carrier and the third convolution operation group of the watermark are fused with the output of a fourth deconvolution operation group; The outputs of the second convolution operation group of the carrier and the second convolution operation group of the watermark are fused with the output of a fifth deconvolution operation group; The outputs of the first convolution operation group of the carrier and the first convolution operation group of the watermark are fused with the output of a sixth deconvolution operation group; The sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are input to a seventh convolution operation group after channel stitching, and the output end of the seventh convolution operation group is connected to the first deconvolution operation group; The results after skip connection of the sixth deconvolution operation group, the first convolution operation group of the carrier, and the first convolution operation group of the watermark are input to a seventh deconvolution operation group; The input terminal of the eighth convolution operation group is input to the seventh deconvolution operation group, and the output is a watermarked image; One convolution operation group includes a convolution layer Conv, an activation layer LeakyReLu, and a batch normalization layer BN arranged in sequence, and one deconvolution operation group includes a deconvolution layer ConvT, an activation layer LeakyReLu, and a batch normalization layer BN arranged in sequence.

[0030] It can be seen that the Y-Net generating network uses all carrier convolution operation groups to extract the features of carrier image, and all watermark convolution operation groups to extract the features of watermark image, and the operation of two branches to extract the features of carrier image and watermark image independently can effectively prevent the confusion of the features of two kinds of images. In addition, in the subsequent deconvolution process, a skip connection structure is provided to fuse the multi-dimensional features extracted in the early stage, so that the watermarked image can retain as much original image information as possible, which is very convenient for the subsequent watermark image reconstruction work.

[0031] S20: inputting the watermarked image into a pre-trained compression loss simulation network to obtain a simulated compressed watermarked image; In the concrete implementation process, in order to improve the transmission rate during network transmission, a JPEG compression operation is usually performed on the image. JPEG compression removes the information of the high frequency components of the image, so that the watermark image embedded in the high frequency components is seriously damaged and cannot be reconstructed. Therefore, it is necessary to introduce JPEG compression loss into the watermark network, but the quantization operation of JPEG compression is discontinuous and cannot participate in the backpropagation process of the network. Therefore, the compression loss simulation network in the implementation of the present invention can select the U-Net network to be used as the simulation of JPEG compression loss and effectively improve the image quality of the reconstructed image of the watermark.

[0032] In a specific embodiment of the present invention, the U-Net compression loss simulation network has a skip connection structure, and includes a first convolution operation group, a second convolution operation group, a third convolution operation group, a fourth convolution operation group, a first deconvolution operation group, a second deconvolution operation group, a third deconvolution operation group and a fifth convolution operation group; the output of the third convolution operation group is fused with the output of the first deconvolution operation group; the output of the second convolution operation group is fused with the output of the second deconvolution operation group; the output of the first convolution operation group is fused with the output of a third deconvolution operation group; A convolution operation group includes a convolution layer, an activation layer, and a batch normalization layer set in sequence, and a deconvolution operation group includes a deconvolution layer, an activation layer, and a batch normalization layer set in sequence.

[0033] The U-Net compression loss simulation network is shown in Figure 3, where the first convolution operation group, the second convolution operation group, the third convolution operation group, the fourth convolution operation group, the first deconvolution operation group, the second deconvolution operation group, the third deconvolution operation group, and the fifth convolution operation group correspond to Conv1, Conv2, Conv3, Conv4, ConvT1, ConvT2, ConvT3, and Conv5 shown in Figure 3. The convolution layer, activation layer, batch normalization layer, and deconvolution layer correspond to Conv, LeakyReLU, BN, and ConvT in Figure 3, respectively.

[0034] The watermarked image is input to the compression loss simulation network to simulate image quality loss during the JPEG compression process, and the watermarked image after the simulated compression, i.e., the simulated compressed watermarked image, is output.

[0035] S30: discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result, including identifying the watermarked image as a carrier image or a watermarked image; After obtaining the simulated compressed watermarked image, in order to improve security and judge whether the embedding position of the watermark image is reasonable, if the optimal steganalysis device Zhu-Net is selected as the discriminator, the simulated compressed watermarked image and the carrier image can be effectively distinguished. Therefore, in the specific implementation process, the discrimination network selects the Zhu-Net network to learn the watermarked image generated by the watermarking algorithms in different spatial domains such as S-UNIWARD, HUGO, etc., and the compressed simulated watermarked image is input to the discrimination network for discrimination, and a binary classification result is output, that is, the watermarked image is identified as a carrier image or a watermarked image.

[0036] In a specific embodiment of the present invention, the discrimination network is obtained through the following training steps: collecting a plurality of image samples and determining a label for each sample image, said label comprising a probability of a watermark image being contained in the corresponding sample image; A step of training Zhu-Net using each image sample as input and the label of each image sample as output to obtain a discrimination network.

[0037] S40: inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; In a specific implementation process, the extraction network can select a U-Net++ network to input the compressed simulated watermarked image into the extraction network and output a reconstructed image of the watermark. The extraction network constructed with a U-Net++ network structure can capture the features of different levels of the feature map through a series of nested jump connection operations, and ensure high accuracy of watermark image extraction.

[0038] S50: Calculate the total loss function of the watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discriminant network, and optimize the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss is reduced and maintained in a stable state (the watermark network includes a Y-Net generating network, a compression loss simulation network, a discriminant network and an extracting network, and the weights of the discriminant network and the compression loss simulation network are fixed during the optimization process); Using the one-way adversarial training method, a pre-trained compression loss simulation network and a discriminant network are added to the watermark network, and the weights of the compression loss simulation network and the discriminant network are fixed. In the network training process, only the Y-Net generating network and the extracting network are continuously changed, which greatly improves the stability of the network, accelerates the convergence speed, and improves the image quality while ensuring the security of the image watermark.

[0039] In a specific implementation of an embodiment of the present invention, the total loss function of the watermark network was:

[0040]

number

[0041] The formula for the Y-Net generative network loss (i.e., the image quality loss of the watermarked image) is as follows:

[0042]

number

number

[0043] The extraction network loss (i.e., the image quality loss of the reconstructed image of the watermark) L s The calculation formula was as follows:

[0044]

number

number

[0045] 2. The discrimination network loss L d The calculation formula was as follows:

[0046]

number

[0047] During the training process of the watermark network, the weights of the pre-trained lossy network and discriminant network are fixed, and only the weights of the generative network and the extractive network are continuously changed, which effectively improves the stability of the network training.

[0048] S60: using the trained watermark network to generate a watermarked image based on the carrier image and the watermark image, JPEG compressing the watermarked image, and reconstructing the watermark image from the compressed watermarked image; In practical applications, the carrier image and the watermark image are input into the trained Y-Net generation network to obtain a high-quality watermarked image, the watermarked image is input into the trained compression loss simulation network to obtain a simulated JPEG compressed watermarked image, and the compressed watermarked image is input into the trained extraction network to extract the watermark image hidden in the watermarked image.

[0049] In summary, the Y-Net-based digital watermarking method in the embodiment of the present invention inputs a carrier image and a watermark image into a Y-Net generation network to obtain a watermarked image, inputs the watermarked image into a pre-trained compression loss simulation network to obtain a simulated compressed watermarked image, and uses a pre-trained discrimination network to determine whether the watermark image is hidden in the simulated compressed watermarked image, and at the same time inputs the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image, so that it has high concealment ability, high security and strong robustness.

[0050] To verify the effectiveness of the present invention, we first trained the proposed watermark model based on the public dataset PASCAL-VOC200 and tested it on the public dataset PASCAL-VOC2012. The experimental results on image quality are shown in Table 1. Among them, the U-Net [Duan XT, Jia K, Li BX, et al. Reversible image steganography schemebased on a U-Net structure[J]. IEEE Access, 2019,2019(7): 9314-9323] model is currently the best model for generating watermarked images and reconstructing watermarked images. ISGAN [Zhang R, Dong SQ, Liu JY. Invisible steganography via generative adversarial networks. Multimedia Tools and Applications,2019,78(7): 8559-8575] is the best model for generating watermarked images and reconstructing watermarked images for the grayscale watermark embedding surface. The experimental results on safety are shown in Table 2. SRNet [Boroumand M, Chen M, Fridrich J. Deep residual network for steganalysis of digital images[J]. IEEE Transactions on Information Forensics and Security, 2018, 14(5): 1181-1193] and Zhu-Net [Zhang R, Zhu F, Liu J, et al. Depth-wise separable convolutions and multi-level pooling for an efficient spatial CNN-based steganalysis[J]. IEEE Transactions on Information Forensics and Security, 2019, 2019(15): 1138-1150] are currently the two steganalysis models with better effectiveness.

[0051] Table 1 [Table 1]

[0052] Table 2 [Table 2]

[0053] Second Embodiment Based on the same technical idea of ​​the invention as the first embodiment, the embodiment of the present invention is as follows: a first processing unit for inputting the carrier image and the watermark image into a Y-Net generating network to obtain a watermarked image in which the watermark image is embedded; a second processing unit for inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; a discrimination unit for discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result, the discrimination result including identifying the watermarked image as a carrier image or a watermarked image; a reconstruction unit for inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; a training unit for calculating a total loss function of the watermark network according to the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss is reduced and maintained in a stable state (the watermark network includes a Y-Net generating network, a loss simulation network, a discriminant network and an extracting network, and during the optimization process, the weights of the discriminant network and the loss simulation network are fixed); a watermarking unit for generating a watermarked image based on a carrier image and a watermark image using the trained watermarking network, compressing the watermarked image, and reconstructing the watermarked image from the compressed watermarked image; A strong and robust watermarking system based on Y-Net with eigenvalues ​​is presented.

[0054] The other parts are the same as those in the first embodiment.

[0055] (Third embodiment) Based on the same inventive technical idea as the first embodiment, the embodiment of the present invention provides a strong and robust digital watermarking system based on Y-Net, which includes a storage medium and a processor; the storage medium is adapted to store instructions; The processor is adapted to operate according to the instructions to perform the method according to any one of the first embodiments.

[0056] As will be appreciated by one skilled in the art, embodiments of the present application may be embodied as a system, method, or computer program product. Thus, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied in one or more computer readable recording media (including, but not limited to, magnetic disk storage devices, CD-ROMs, optical memory, etc.) having computer readable program code thereon.

[0057] The present application will be described with reference to flowcharts and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the present application. Each flow and / or block in the flowcharts and / or block diagrams and combinations of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that the instructions, executed via a processor of the computer or other programmable data processing apparatus, generate means for implementing the functions defined in the flow or flows of the flowcharts and / or the block or blocks of the block diagrams.

[0058] These computer program instructions may also be stored in a computer-readable memory, where the instructions stored in the computer-readable memory may direct a computer, other programmable data processing apparatus, or other device to function in a particular manner to produce an article of manufacture that includes instructions that implement the functions defined in a flow or flows of the flowchart and / or a block or blocks of the block diagram.

[0059] The computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide a process for implementing the functions defined in the flow or flows of the flowcharts and / or the block or blocks of the block diagrams.

[0060] Finally, it should be noted that the above embodiments are merely for clarifying the technical means of the present invention, and the present invention is not limited by these embodiments. Although the present invention has been described in detail with reference to the above embodiments, it will be understood by those skilled in the art that the specific embodiments of the present invention can be modified or replaced with equivalents, and any modifications or replacements that do not depart from the spirit and scope of the present invention should fall within the scope of the claims of the present invention.

[0061] (Additional Note) (Appendix 1) A digital watermarking method based on Y-Net, comprising: inputting the carrier image and the watermark image into a Y-Net generating network to obtain a watermarked image in which the watermark image is embedded; inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result including identifying the watermarked image as the carrier image or the watermarked image; inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; calculating a total loss function of the entire watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss decreases and remains stable, the watermark network including the Y-Net generating network, the loss simulation network, the discriminant network and the extraction network, and during the optimization process, the weights of the discriminant network and the loss simulation network are fixed; and generating the watermarked image based on the carrier image and the watermark image using a trained watermark network, compressing the watermarked image, and reconstructing the watermark image from the compressed watermarked image; A digital watermarking method based on Y-Net, comprising:

[0062] (Appendix 2) The Y-Net generating network is a skip connection structure, and includes: a first convolution operation group of carrier, a second convolution operation group of carrier, a third convolution operation group of carrier, a fourth convolution operation group of carrier, a fifth convolution operation group of carrier, a sixth convolution operation group of carrier, a first convolution operation group of watermark, a second convolution operation group of watermark, a third convolution operation group of watermark, a fourth convolution operation group of watermark, a fifth convolution operation group of watermark, a sixth convolution operation group of watermark, a seventh convolution operation group, a first deconvolution operation group, a second deconvolution operation group, a third deconvolution operation group, a fourth deconvolution operation group, a fifth deconvolution operation group, a sixth deconvolution operation group, a seventh deconvolution operation group, and an eighth convolution operation group; The output of the sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are fused with the output of the first deconvolution operation group; The output of the fifth convolution operation group of the carrier and the fifth convolution operation group of the watermark are fused with the output of the second deconvolution operation group; the output of the fourth convolution operation group of the carrier and the fourth convolution operation group of the watermark are fused with the output of the third deconvolution operation group; the outputs of the third convolution operation group of the carrier and the third convolution operation group of the watermark are fused with the output of the fourth deconvolution operation group; the outputs of the second convolution operation group of the carrier and the second convolution operation group of the watermark are fused with the output of the fifth deconvolution operation group; The outputs of the first convolution operation group of the carrier and the first convolution operation group of the watermark are fused with the output of the sixth deconvolution operation group; The sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are input to the seventh convolution operation group after channel stitching, and the output end of the seventh convolution operation group is connected to the first deconvolution operation group; The results after skip connection of the sixth deconvolution operation group, the first convolution operation group of the carrier, and the first convolution operation group of the watermark are input to the seventh deconvolution operation group; The input terminal of the eighth convolution operation group is input to the seventh deconvolution operation group, and the output is the watermarked image; A convolution operation group includes a convolution layer, an activation layer, and a batch normalization layer arranged in sequence, and a deconvolution operation group includes a deconvolution layer, an activation layer, and a batch normalization layer arranged in sequence; 2. The Y-Net-based watermarking method according to claim 1,

[0063] (Appendix 3) The discrimination network comprises: collecting a plurality of image samples and determining a label for each of the sample images, the label comprising a probability of the watermark image being included in the corresponding sample image; and training Zhu-Net using each of the image samples as input and the labels of each of the image samples as output to obtain the discrimination network; 2. The digital watermarking method based on Y-Net described in Appendix 1, characterized in that the Y-Net is obtained through a training step of:

[0064] (Appendix 4) The Y-Net-based watermarking method according to claim 1, characterized in that the compression loss simulation network is a U-Net network.

[0065] (Appendix 5) 2. The Y-Net based watermarking method according to claim 1, wherein the extraction network is a U-Net++ network.

[0066] (Appendix 6) The Y-Net-based watermarking method according to claim 1, characterized in that the discrimination network is a discriminator, and the discriminator includes a softmax function.

[0067] (Appendix 7) The total loss function of the watermarking network is

number

[0068] (Appendix 8) The formula for calculating the Y-Net generated network loss is:

number

number

number

number

number

[0069] (Appendix 9) A robust watermarking device based on Y-Net, a first processing unit for inputting a carrier image and a watermark image into a Y-Net generating network to obtain a watermarked image in which the watermark image is embedded; a second processing unit for inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; a discrimination unit for discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result, the discrimination result including identifying the watermarked image as the carrier image or the watermarked image; a reconstruction unit for inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; a training unit for calculating a total loss function of a watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss is reduced and maintained in a stable state (the watermark network includes the Y-Net generating network, the compression loss simulation network, the discriminant network and the extraction network, and during the optimization process, the weights of the discriminant network and the compression loss simulation network are fixed); a watermarking unit for generating the watermarked image based on the carrier image and the watermark image using a trained watermarking network, compressing the watermarked image, and reconstructing the watermark image from the compressed watermarked image; A strong and robust digital watermarking device based on Y-Net, characterized by including:

[0070] (Appendix 10) A robust Y-Net based watermarking system, comprising: a storage medium; and a processor; the storage medium is adapted to store instructions; The processor is adapted to operate in accordance with the instructions to perform the method according to any one of claims 1 to 8. A strong and robust watermarking system based on Y-Net, characterized by:

Claims

1. A computer-implemented Y-Net-based watermarking method, comprising: inputting the carrier image and the watermark image into a Y-Net generation network to obtain a watermarked image in which the watermark image is embedded; inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result including identifying the watermarked image as the carrier image or the watermarked image; inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; calculating a total loss function of the entire watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss decreases and remains stable, the watermark network including the Y-Net generating network, the loss simulation network, the discriminant network and the extraction network, and during the optimization process, the weights of the discriminant network and the loss simulation network are fixed; and generating the watermarked image based on the carrier image and the watermark image using a trained watermark network, compressing the watermarked image, and reconstructing the watermark image from the compressed watermarked image; A Y-Net-based digital watermarking method, comprising:

2. The Y-Net generating network is a skip connection structure, and includes: a first convolution operation group of a carrier, a second convolution operation group of a carrier, a third convolution operation group of a carrier, a fourth convolution operation group of a carrier, a fifth convolution operation group of a carrier, a sixth convolution operation group of a carrier, a first convolution operation group of a watermark, a second convolution operation group of a watermark, a third convolution operation group of a watermark, a fourth convolution operation group of a watermark, a fifth convolution operation group of a watermark, a sixth convolution operation group of a watermark, a seventh convolution operation group, a first deconvolution operation group, a second deconvolution operation group, a third deconvolution operation group, a fourth deconvolution operation group, a fifth deconvolution operation group, a sixth deconvolution operation group, a seventh deconvolution operation group, and an eighth convolution operation group; the outputs of the sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are fused with the output of the first deconvolution operation group; the output of the fifth convolution operation group of the carrier and the fifth convolution operation group of the watermark are fused with the output of the second deconvolution operation group; the outputs of the fourth convolution operation group of the carrier and the fourth convolution operation group of the watermark are fused with the output of the third deconvolution operation group; the outputs of the third convolution operation group of the carrier and the third convolution operation group of the watermark are fused with the output of the fourth deconvolution operation group; the outputs of the second convolution operation group of the carrier and the second convolution operation group of the watermark are fused with the output of the fifth deconvolution operation group; the outputs of the first convolution operation group of the carrier and the first convolution operation group of the watermark are fused with the output of the sixth deconvolution operation group; The sixth convolution operation group of the carrier and the sixth convolution operation group of the watermark are input to the seventh convolution operation group after channel stitching, and the output end of the seventh convolution operation group is connected to the first deconvolution operation group; The results after skip connection of the sixth deconvolution operation group, the first convolution operation group of the carrier, and the first convolution operation group of the watermark are input to the seventh deconvolution operation group; The input of the eighth convolution operation group is input to the seventh deconvolution operation group, and the output is the watermarked image; A convolution operation group includes a convolution layer, an activation layer, and a batch normalization layer arranged in sequence, and a deconvolution operation group includes a deconvolution layer, an activation layer, and a batch normalization layer arranged in sequence. The Y-Net based digital watermarking method according to claim 1.

3. The discrimination network comprises: collecting a plurality of image samples and determining a label for each said image sample, said label comprising a probability of said watermark image being contained in the corresponding image sample; and training Zhu-Net using each of the image samples as input and the labels of each of the image samples as output to obtain the discrimination network; The Y-Net-based watermarking method according to claim 1, characterized in that the Y-Net-based watermarking method is obtained through a training step of:

4. The Y-Net-based digital watermarking method according to claim 1, wherein the compression loss simulation network is a U-Net network.

5. The Y-Net based watermarking method according to claim 1, characterized in that the extraction network is a U-Net++ network.

6. The Y-Net-based watermarking method according to claim 1, wherein the discrimination network is a discriminator, and the discriminator includes a softmax function.

7. The total loss function of the watermarking network is [0010] [In the formula, L ｃ is the Y-Net generated network loss, L ｓ is the extracted network loss, L ｄ is the discriminative network loss, and α, β, δ are weights for controlling the Y-Net generative network loss, the extractive network loss and the discriminative network loss.

8. The formula for calculating the Y-Net generation network loss is: [0025] [where n is the total number of pixels in the image, c ｉ is the original carrier image, [0030] represents the watermarked image.] The extracted network loss L ｓ Formula for calculation: [0045] [In the formula, s ｉ is the original watermark image, [0050] represents the reconstructed image of the watermark.] The discrimination network loss L ｄ Formula for calculation: [006] [In the formula, y ｉ ′ represents the true label, i.e., the true result of whether the image input to the discrimination network is a carrier image or a watermarked image, and y ｉ represents the output of the discrimination network.] The Y-Net-based digital watermarking method according to claim 7, wherein:

9. A robust watermarking device based on Y-Net, comprising: a first processing unit for inputting a carrier image and a watermark image into a Y-Net generating network to obtain a watermarked image in which the watermark image is embedded; a second processing unit for inputting the watermarked image into a pre-trained lossy simulation network to obtain a simulated compressed watermarked image; a discrimination unit for discriminating the simulated compressed watermarked image using a pre-trained discrimination network to obtain a discrimination result, the discrimination result including identifying the watermarked image as the carrier image or the watermarked image; a reconstruction unit for inputting the simulated compressed watermarked image into an extraction network to obtain a watermarked reconstructed image; a training unit for calculating a total loss function of a watermark network based on the watermarked image, the watermarked reconstructed image and the result of the discriminant network, optimizing the watermark network with the goal of minimizing the loss function, and the training is considered to be completed when the loss is reduced and maintained in a stable state (the watermark network includes the Y-Net generating network, the compression loss simulation network, the discriminant network and the extraction network, and the weights of the discriminant network and the compression loss simulation network are fixed during the optimization process); a watermarking unit for generating the watermarked image based on the carrier image and the watermark image using a trained watermarking network, compressing the watermarked image, and reconstructing the watermark image from the compressed watermarked image; A strong and robust digital watermarking device based on Y-Net, comprising:

10. A robust Y-Net based watermarking system, comprising: a storage medium; and a processor; the storage medium is adapted to store instructions; The processor is adapted to operate in accordance with the instructions to carry out the method of any one of claims 1 to 8. A strong and robust digital watermarking system based on Y-Net, characterized in that:

Images