Robust watermark embedding method and system based on multi-dimensional information embedding and texture guidance

Abstract

The present invention discloses a robust watermark embedding method and system based on multi-dimensional information embedding and texture guidance, and relates to the field of information security technology. The method comprises the following steps: receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the texture of the carrier image; inputting the watermark message block guided by the texture of the carrier image into a pre-established encoder, and outputting an encoded image; inputting the encoded image into a pre-established noise simulator, and outputting a noise image; and inputting the noise image into a pre-established decoder, and extracting one-dimensional watermark information.

Description

Technical Field

[0001] The invention relates to the technical field of information security, and in particular to a robust watermark embedding method and system based on multi-dimensional information embedding and texture guidance. Background Art

[0002] With the widespread application and dissemination of digital media, digital copyright protection faces unprecedented challenges. The original data in digital media can be easily copied, modified and tampered with, and the convenience of the Internet allows these modified digital media to be quickly disseminated and shared, which brings great risks to the copyright and content security of digital media. In order to solve this problem, digital watermarking technology was proposed. This technology can embed invisible identification information in digital media (images, videos, texts, sounds, etc.) to verify the integrity and authenticity of digital media, thereby improving the efficiency of digital copyright protection. By embedding digital watermarks into digital media, the owner of digital media can track its usage and protect its copyright, while preventing digital media from being tampered with and forged. The emergence of digital watermarking technology provides new means and technical support for copyright protection of digital media, and provides a strong guarantee for the healthy development of the digital media industry. An important issue of digital watermarking technology is its robustness, that is, whether the digital watermark can withstand various attacks and tampering without failure.

[0003] The representative digital watermarking algorithm in the early stage is LSB (Least Significant Bit), which embeds the watermark information into the pixels of the image based on the least significant bit (least significant bit) of the image. The LSB method embeds the binary data of the watermark information into the lowest bit of the image pixel. Because the lowest bit has little impact on the perception of the image, the appearance of the image after the watermark is embedded changes little, and the concealment of the watermark information can also be guaranteed. However, this method has two disadvantages: First, the robustness of the LSB method is weak. Due to the high sensitivity of the lowest bit, image processing operations (such as compression, rotation, filtering, etc.) may cause part of the lowest bit information to be lost or changed, making the watermark vulnerable to attack and tampering. Second. The LSB method can be easily detected by statistical measurement. Summary of the invention

[0004] In order to address the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide a robust watermark embedding method and system based on multi-dimensional information embedding and texture guidance, which can combine multi-dimensional watermark information with the texture information of the embedded image and use perceptual loss for training, thereby improving the embedding accuracy of watermark information and the quality of watermarked image.

[0005] In a first aspect, the purpose of the present invention can be achieved by the following technical solution: a robust watermark embedding method based on multi-dimensional information embedding and texture guidance, the method comprising the following steps:

[0006] Receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the carrier image texture;

[0007] Input the watermark message block guided by the carrier image texture into a pre-established encoder, and output the encoded image;

[0008] Input the coded image into a pre-established noise simulator and output a noise image;

[0009] The noisy image is input into a pre-established decoder to extract the one-dimensional watermark information.

[0010] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established message processor includes a multi-dimensional message processing module and a texture guidance module, wherein the multi-dimensional message processing module processes the watermark information into an initial watermark information block in three different ways, and then performs feature aggregation through convolution to form a final watermark information block;

[0011] The texture guidance module converts the carrier image I co Four images with different scales are obtained through SWT transformation, including low-frequency LL image, horizontal detail LH image, vertical detail HL image, and diagonal detail HH image. The images with different scales are integrated into the watermark information block of the multi-dimensional message processing module to achieve the effect of using the carrier image texture to guide the distribution of watermark information.

[0012] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the multi-dimensional message processing module includes in three different ways:

[0013] The first information processing method: The one-dimensional information M is transformed into a two-dimensional information graph through the MBRS information processing method, and the two-dimensional information block is expanded to a size of 128*128*64 through four deconvolution modules;

[0014] The second information processing method: diffuse the watermark information through the diffusion block, and expand the information to a watermark information block of 128*128*64 in size through the first method;

[0015] The third information processing method: the watermark information is enlarged to 16384, and then transformed into a two-bit information block of 128*128 size through four layers of SE blocks to process the information into a watermark information block of 128*128*64 size. Finally, the final watermark information block is formed by splicing three watermark information blocks and aggregating the watermark information through a layer of SE blocks.

[0016] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established encoder converts the carrier image I co After a ConvBNRelu block, the channels are expanded from 3 to 64 channels, and then the carrier image I is processed through four SE blocks. co The features of the image are extracted to obtain a 128*128*64 image feature block. The image feature block is concatenated with the watermark information block and fused with the watermark and image through a ConvBNRelu block to obtain a feature block. The feature block is then combined with the carrier image I co The concatenation is performed and the number of channels is reduced to 3 through a layer of convolution blocks to form the encoded image I en .

[0017] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established noise simulator simulates various noises to the encoded image I en Attack and form a noise image I no ,In the end-to-end training process, the influence of noise is propagated to the encoder through the back-propagation algorithm, making the embedded watermark information block robust.

[0018] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the carrier image I co and the coded image I en The discriminator distinguishes and labels the carrier image I co and the coded image I en The discriminator is trained to improve the identification ability of the discriminator and the encoding ability of the encoder through adversarial training.

[0019] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established decoder first converts the noise image I into no The channels are expanded to 64 channels, and then the height and width of the feature block are reduced through the SE downsampling block, and then the number of channels of the feature block is changed to 1 through the ConvBNRelu layer, and it is transformed into a one-dimensional watermark information M' through deformation and diffusion.

[0020] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the carrier image I co and the coded image I en By inputting it into the pre-trained VGG-19 model and outputting its middle layer feature image for comparison, the perceptual loss is obtained. The calculation process is as follows:

[0021]

[0022] Among them, λ jis the weight of layer j, J is the set of selected pre-trained network layers, C j H j W j Respectively represent the number of channels, height, and width of the selected layer, φ j (y), is the carrier image I co and the coded image I en Output through the network.

[0023] In a second aspect, in order to achieve the above-mentioned purpose, the present invention discloses a robust watermark embedding system based on multi-dimensional information embedding and texture guidance, comprising:

[0024] A message processing unit, used for receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the carrier image texture;

[0025] An encoding unit, used for inputting a watermark message block guided by the carrier image texture into a pre-established encoder, and outputting an encoded image;

[0026] Noise simulation unit: input the coded image into a pre-established noise simulator and output a noise image;

[0027] The decoding and extraction unit is used to input the noise image into a pre-established decoder to extract the one-dimensional watermark information.

[0028] In conjunction with the second aspect, in some implementations of the second aspect, the system further includes: a message processor pre-established in the message processing unit includes a multi-dimensional message processing module and a texture guidance module, wherein the multi-dimensional message processing module processes the watermark information into an initial watermark information block in three different ways, and then performs feature aggregation through convolution to form a final watermark information block;

[0029] The texture guidance module converts the carrier image I co Four decomposed images of different scales are obtained through SWT transformation, including low-frequency LL image, horizontal detail LH image, vertical detail HL image, and diagonal detail HH image. The images decomposed at each scale are fused into the watermark information block of the multi-dimensional message processing module to achieve the effect of using the carrier image texture to guide the distribution of watermark information;

[0030] Or the multi-dimensional message processing module in the message processing unit includes three different ways:

[0031] The first information processing method: The one-dimensional information M is transformed into a two-dimensional information graph through the MBRS information processing method, and the two-dimensional information block is expanded to a size of 128*128*64 through four deconvolution modules;

[0032] The second information processing method: diffuse the watermark information through the diffusion block, and expand the information to a watermark information block of 128*128*64 in size through the first method;

[0033] The third information processing method: the watermark information is enlarged to 16384, and then transformed into a two-bit information block of 128*128 size through four layers of SE blocks to process the information into a watermark information block of 128*128*64 size. Finally, the final watermark information block is formed by splicing three watermark information blocks and aggregating the watermark information through a layer of SE blocks.

[0034] Preferably, the encoder pre-established in the encoding unit converts the carrier image I co After a ConvBNRelu block, the channels are expanded from 3 to 64 channels, and then the carrier image I is processed through four SE blocks. co The features of the image are extracted to obtain a 128*128*64 image feature block. The image feature block is concatenated with the watermark information block and fused with the watermark and image through a ConvBNRelu block to obtain a feature block. The feature block is then combined with the carrier image I co The concatenation is performed and the number of channels is reduced to 3 through a layer of convolution blocks to form the encoded image I en ;

[0035] Preferably, the noise simulator pre-established in the noise simulation unit simulates various noises to the coded image I en Attack and form a noise image I no ,In the end-to-end training process, the influence of noise is propagated to the encoder through the back-propagation algorithm, making the embedded watermark information block robust;

[0036] Among them, the carrier image I co and the coded image I en The discriminator distinguishes and labels the carrier image I co and the coded image I en Train the discriminator and improve the discriminative ability of the discriminator and the encoding ability of the encoder through adversarial training;

[0037] Preferably, the decoder pre-established in the decoding extraction unit first converts the noise image I no The channel is expanded to 64 channels, and then the height and width of the feature block are reduced through the SE downsampling block, and then the number of channels of the feature block is changed to 1 through the ConvBNRelu layer, and it is transformed into a one-dimensional watermark information M' through deformation and diffusion;

[0038] Furthermore, the carrier image I co and the coded image Ien By inputting into the pre-trained VGG-19 model and outputting its intermediate layer feature image for comparison, the perceptual loss is obtained. The role of the perceptual loss is to help train the encoder to generate encoded images with higher human visual quality. The calculation process is as follows:

[0039]

[0040] Among them, λ j is the weight of layer j, J is the set of selected pre-trained network layers, C j H j W j Respectively represent the number of channels, height, and width of the selected layer, φ j (y), is the carrier image I co and the coded image I en Output through the network.

[0041] Beneficial effects of the present invention:

[0042] The present invention uses a robust watermark embedding method guided by information multi-dimensional embedding and texture to embed multi-dimensional information into the host image by fusing the host image texture information, and further enhances the generated watermark image features through training by perceptual loss, wherein the use of information multi-dimensional embedding can increase the redundancy of the information embedding method and improve the robustness of the embedded watermark. Using the texture-guided method, the information can be guided to be embedded into the image texture to improve the quality of the generated watermark image. The use of perceptual loss to train the deep learning framework improves the visual quality of the generated image. Compared with the traditional robust watermark embedding method, this patent can greatly improve the image quality of the embedded watermark image and the accuracy of the embedded watermark. For the same noise, our method and other deep learning robust watermark embedding algorithms, the error rate (Error_rate) and image quality (PSNR and SSIM) obtained are compared for the image quality after the watermark information is embedded. This shows that compared with other deep learning robust watermark embedding methods, this method can embed watermark information more accurately and the embedded image is more imperceptible, and has greater use value. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a schematic flow chart of the method of the present invention;

[0045] Figure 2It is a model framework structure diagram of the robust watermark embedding method guided by multi-dimensional information embedding and texture of the present invention;

[0046] Figure 3 It is a framework diagram of the multidimensional message processor of the present invention;

[0047] Figure 4 It is a structural diagram of the texture guided processor framework of the present invention;

[0048] Figure 5 The original carrier image of the present invention is transformed by SWT to obtain four decomposed images of different scales;

[0049] Figure 6 It is a schematic diagram of the system structure of the present invention;

[0050] Figure 7 It is a detailed implementation flow chart of the method of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] Embodiment 1:

[0053] The following is an introduction to the relevant terms involved in the embodiments of the present application:

[0054] End-to-end training: is a machine learning training method that allows the model to automatically learn feature representations and optimal solutions for tasks from input data. We use end-to-end training of encoders and decoders for watermark embedding and extraction.

[0055] Encoder: An encoder is a component in machine learning and deep learning models that converts input data into representation vectors or feature representations. The encoder accepts input data and maps it to a low-dimensional representation space through a series of transformations.

[0056] In natural language processing, an encoder usually refers to a model that converts a text sequence into a fixed-length vector representation. Common encoders include recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), etc. These models model the entire sequence and generate a final representation vector by traversing the input sequence and updating the hidden state at each time step.

[0057] In image processing, the encoder is usually part of a convolutional neural network (CNN). CNN extracts local features of an image through a series of convolutional layers and pooling layers, and gradually reduces the spatial dimensions of the feature map. Finally, the image features are mapped to a low-dimensional representation space through a fully connected layer or other methods.

[0058] The goal of an encoder is to transform input data into representations with good expressive power so that subsequent tasks or models can better utilize these representations. The selection and design of an encoder usually depends on the specific task and data type, as well as performance requirements and computing resource constraints.

[0059] In this task, the role of the encoder is to use a convolutional neural network to extract features from the image and fuse the image feature blocks with the message feature blocks to generate components of the encoded image.

[0060] Decoder: A decoder is a component in machine learning and deep learning models that converts the representation vector or feature representation processed by the encoder back to the original data or representation in the target domain. The decoder takes the representation vector generated by the encoder and maps it back to the original data or target domain through a series of transformations.

[0061] In natural language processing, a decoder usually refers to a model that converts the vector representation generated by an encoder into a text sequence. Common decoders include recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), etc. These models gradually build the final text sequence by traversing the output sequence and generating a word or character at each time step.

[0062] In image processing, the decoder is usually part of a convolutional neural network (CNN). CNN maps image features back to the original image space through a series of deconvolution layers, upsampling layers, or other methods. Finally, the image can be restored to its original state through some post-processing steps (such as activation functions or thresholds).

[0063] The goal of the decoder is to convert the representation vector generated by the encoder into a representation in the original data or target domain in order to complete a specific task or meet a requirement. The choice and design of the decoder usually depends on the specific task and data type, as well as the performance requirements and computing resource constraints.

[0064] In this task, the decoder accepts the coded image attacked by noise as input and obtains the watermark message by performing feature extraction and analysis on the noisy image.

[0065] Robust watermark: Robust watermark algorithm is one of the most widely used digital image processing technologies. It is a "hidden information" technology that can insert one or more information into an image. This information can affect the quality of the image without loss, and the original image can be perfectly reconstructed when the watermark is removed. This technology can effectively prevent the theft of pictures and effectively protect the copyright of pictures, thus greatly ensuring the security of pictures.

[0066] like Figure 1 As shown, the robust watermark embedding method based on multi-dimensional information embedding and texture guidance is characterized in that the method comprises the following steps:

[0067] Receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the carrier image texture;

[0068] The pre-established message processor includes a multi-dimensional message processing module and a texture guidance module, wherein the multi-dimensional message processing module processes the watermark information into an initial watermark information block through three different methods, and then performs feature aggregation through convolution to form a final watermark information block;

[0069] The texture guidance module converts the carrier image I co Four images with different scales are obtained through SWT transformation, including low-frequency LL image, horizontal detail LH image, vertical detail HL image, and diagonal detail HH image. The images with different scales are integrated into the watermark information block of the multi-dimensional message processing module to achieve the effect of using the carrier image texture to guide the distribution of watermark information.

[0070] Furthermore, the multi-dimensional message processing module includes three different methods:

[0071] The first information processing method: The one-dimensional information M is transformed into a two-dimensional information graph through the MBRS information processing method, and the two-dimensional information block is expanded to a size of 128*128*64 through four deconvolution modules;

[0072] The second information processing method: diffuse the watermark information through the diffusion block, and expand the information to a watermark information block of 128*128*64 in size through the first method;

[0073] The third information processing method: the watermark information is enlarged to 16384, and then transformed into a two-bit information block of 128*128 size through four layers of SE blocks to process the information into a watermark information block of 128*128*64 size. Finally, the final watermark information block is formed by splicing three watermark information blocks and aggregating the watermark information through a layer of SE blocks.

[0074] Input the watermark message block guided by the carrier image texture into a pre-established encoder, and output a coded image; input the coded image into a pre-established noise simulator, and output a noise image;

[0075] The pre-established encoder transforms the carrier image I co After a ConvBNRelu block, the channels are expanded from 3 to 64 channels, and then the carrier image I is processed through four SE blocks. co The features of the image are extracted to obtain a 128*128*64 image feature block. The image feature block is concatenated with the watermark information block and fused with the watermark and image through a ConvBNRelu block to obtain a feature block. The feature block is then combined with the carrier image I co The concatenation is performed and the number of channels is reduced to 3 through a layer of convolution blocks to form the encoded image I en .

[0076] The pre-built noise simulator simulates various noises on the coded image I en Attack and form a noise image I no ,In the end-to-end training process, the influence of noise is propagated to the encoder through the back-propagation algorithm, making the embedded watermark information block robust.

[0077] Specifically, the present invention is further described below by way of embodiments:

[0078] Noise simulator: However, this noise simulation method through back propagation requires the simulated noise to be differentiable. In JPE compression, it is not differentiable due to the quantization operation. The gradient cannot be propagated to the encoder through the real JPEG compression method, so this patent chooses to use the JPEG-mask method to simulate the noise of JPEG compression. The specific method is that JPEG-Mask applies a fixed mask, retaining only 25 low-frequency DCT coefficients in the Y channel and 9 in the U, V channels (following JPEG, it also retains more information in the Y channel), and the other coefficients are set to 0.

[0079] Carrier Image I co and the coded image I en The discriminator distinguishes and labels the carrier image I co and the coded image I en The discriminator is trained to improve the identification ability of the discriminator and the encoding ability of the encoder through adversarial training.

[0080] Specifically, the present invention is further described below by way of embodiments:

[0081] Discriminator: Applying adversarial learning to improve the quality of generated watermark images is an important method for deep learning robust watermarking. The discriminator is a classifier used to distinguish between the encoded image I en and carrier image I co . Through the labeled encoded image I en and carrier image I co To train the discriminator, so that its resolution ability is continuously improved. At the same time, it is hoped that the encoder will generate the encoded image I en To confuse the discriminator, so that the discriminator cannot distinguish the encoded image I en and carrier image I co Through adversarial training, the discriminator's ability to identify will continue to improve, and the encoded image I generated by the encoder en The quality of the generated encoded image I is also getting higher and higher. en image quality and increases the imperceptibility of watermark information.

[0082] The noisy image is input into a pre-established decoder to extract the one-dimensional watermark information.

[0083] The pre-built decoder first passes the noise image I through a ConvBNRelu layer no The channels are expanded to 64 channels, and then the height and width of the feature block are reduced through the SE downsampling block, and then the number of channels of the feature block is changed to 1 through the ConvBNRelu layer, and it is transformed into a one-dimensional watermark information M' through deformation and diffusion.

[0084] Among them, the carrier image I co and the coded image I en The perceptual loss is obtained by inputting it into the pre-trained VGG-19 model and outputting its middle layer feature image for comparison;

[0085] Specifically, the present invention is further described below by way of embodiments:

[0086] VGG Perceptron: VGG Perceptron mainly uses VGG-19 to complete the calculation of perceptual loss. Recent work has focused on minimizing the mean square reconstruction error. The resulting image has a high peak signal-to-noise ratio, but the generated image usually lacks high-frequency details and is perceptually unsatisfactory. Specifically, this patent uses a loss function based on the Euclidean distance between feature maps extracted from the VGG-19 network. Perceptually, the generated image obtains more convincing results. Perceptual loss is a loss function commonly used in image style transfer methods based on deep learning. Compared with the traditional mean square error loss function (MSE), perceptual loss pays more attention to the perceptual quality of the image and is more in line with the human eye's perception of image quality. The perceptual loss is calculated by passing two pictures through the same neural network and comparing the output layers in the middle to measure their differences. Pre-trained convolutional neural networks (CNNs) are usually used. These networks have been trained on large-scale datasets and have strong performance and generalization capabilities. The VGG-19 network has begun to be used for classification tasks, and higher classification accuracy has been achieved through a deeper network structure. The convolutional layer in the VGG-19 network can extract the texture and structural information of the image, while the fully connected layer of the network can extract the semantic information of the image. The perceptual loss is usually calculated by passing the input image and the target image through a pre-trained neural network respectively to obtain their feature representations in the network. These feature representations are then used as inputs to the loss function to calculate the Euclidean distance between them. The goal of the perceptual loss is to minimize the distance between the input image and the target image in the feature space to enhance the feature fitting effect of the encoded image. The method of improving image quality by performing perceptual loss on a pre-trained network was initially used in style transfer and super-resolution tasks. This patent uses this method to achieve great results in improving the quality of encoded images. By passing the encoded image I en and carrier image I co At the same time, it is put into the pre-trained VGG-19 model and its intermediate layer feature image is output for comparison to obtain the perceptual loss. The role of perceptual loss is to help train the encoder to generate coded images with higher human visual quality. The perceptual loss calculation formula based on the VGG-19 network structure is as follows:

[0087]

[0088] Among them, λ j is the weight of layer j, J is the set of selected pre-trained network layers, C j H j W j Respectively represent the number of channels, height, and width of the selected layer, φ j (y), is the carrier image I co and the coded image Ien Output features of the j-th layer network.

[0089] Table 1: Comparison results of this patented method with other watermark embedding methods

[0090]

[0091] Embodiment 2: In the second aspect, as Figure 6 As shown, in order to achieve the above purpose, the present invention discloses a robust watermark embedding system based on multi-dimensional information embedding and texture guidance, comprising:

[0092] A message processing unit, used for receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the carrier image texture;

[0093] An encoding unit, used for inputting a watermark message block guided by the carrier image texture into a pre-established encoder, and outputting an encoded image;

[0094] Noise simulation unit: input the coded image into a pre-established noise simulator and output a noise image;

[0095] The decoding and extraction unit is used to input the noise image into a pre-established decoder to extract the one-dimensional watermark information.

[0096] In conjunction with the second aspect, in some implementations of the second aspect, the system further includes: a message processor pre-established in the message processing unit includes a multi-dimensional message processing module and a texture guidance module, wherein the multi-dimensional message processing module processes the watermark information into an initial watermark information block in three different ways, and then performs feature aggregation through convolution to form a final watermark information block;

[0097] The texture guidance module converts the carrier image I co Four decomposition images of different scales are obtained through SWT transformation, including low-frequency LL image, horizontal detail LH image, vertical detail HL image and diagonal detail HH image; the images decomposed at each scale are fused into the watermark information block of the multi-dimensional message processing module to achieve the effect of using the carrier image texture to guide the distribution of watermark information.

[0098] Or the multi-dimensional message processing module in the message processing unit includes three different ways:

[0099] The first information processing method: The one-dimensional information M is transformed into a two-dimensional information graph through the MBRS information processing method, and the two-dimensional information block is expanded to a size of 128*128*64 through four deconvolution modules;

[0100] The second information processing method: diffuse the watermark information through the diffusion block, and expand the information to a watermark information block of 128*128*64 in size through the first method;

[0101] The third information processing method: the watermark information is enlarged to 16384, and then transformed into a two-bit information block of 128*128 size through four layers of SE blocks to process the information into a watermark information block of 128*128*64 size. Finally, the final watermark information block is formed by splicing three watermark information blocks and aggregating the watermark information through a layer of SE blocks.

[0102] The encoder pre-established in the coding unit converts the carrier image I co After a ConvBNRelu block, the channels are expanded from 3 to 64 channels, and then the carrier image I is processed through four SE blocks. co The features of the image are extracted to obtain a 128*128*64 image feature block. The image feature block is concatenated with the watermark information block and fused with the watermark and image through a ConvBNRelu block to obtain a feature block. The feature block is then combined with the carrier image I co The concatenation is performed and the number of channels is reduced to 3 through a layer of convolution blocks to form the encoded image I en ;

[0103] The noise simulator pre-established in the noise simulation unit simulates various noises on the coded image I en Attack and form a noise image I no ,In the end-to-end training process, the influence of noise is propagated to the encoder through the back-propagation algorithm, making the embedded watermark information block robust to the imposed noise;

[0104] Among them, the carrier image I co and the coded image I en The discriminator distinguishes and labels the carrier image I co and the coded image I en Train the discriminator and improve the discriminant ability and the encoder encoding ability through adversarial training between the encoder and the discriminator.

[0105] Preferably, the decoder pre-established in the decoding extraction unit first converts the noise image I no The channel is expanded to 64 channels, and then the height and width of the feature block are reduced through the SE block downsampling block, and then the number of feature block channels is changed to 1 through the ConvBNRelu layer, and it is transformed into a one-dimensional watermark information M' through deformation and diffusion;

[0106] Furthermore, the carrier image I co and the coded image I enBy inputting into the pre-trained VGG-19 model and outputting its intermediate layer feature image for comparison, the perceptual loss is obtained. The role of the perceptual loss is to help train the encoder to generate encoded images with higher human visual quality. The calculation process is as follows:

[0107]

[0108] Among them, λ j is the weight of layer j, J is the set of selected pre-trained network layers, C j H j W j Respectively represent the number of channels, height, and width of the selected layer, φ j (y), is the carrier image I co and the coded image I en Output features of the j-th layer network.

[0109] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0110] It needs to be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by a processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0111] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0112] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. A robust watermark embedding method based on multi-dimensional information embedding and texture guidance, characterized in that: The method comprises the following steps: Receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the carrier image texture; The pre-established message processor includes a multi-dimensional message processing module and a texture guiding module, wherein the multi-dimensional message processing module processes the watermark information into an initial watermark information block in three different ways, and then performs feature aggregation through convolution to form a final watermark information block; The texture guidance module converts the carrier image I co Four decomposition images of different scales are obtained through SWT transformation, including low-frequency LL image, horizontal detail LH image, vertical detail HL image, and diagonal detail HH image; The multi-dimensional message processing module includes three different methods: The first information processing method: The one-dimensional information M is transformed into a two-dimensional information graph through the MBRS information processing method, and the two-dimensional information block is expanded to a size of 128*128*64 through four deconvolution modules; The second information processing method: diffuse the watermark information through the diffusion block, and expand the information to a watermark information block of 128*128*64 in size through the first method; The third information processing method: the watermark information is enlarged to 16384, and then transformed into a two-bit information block of 128*128 size through four layers of SE blocks to process the information into a watermark information block of 128*128*64 size. Finally, the final watermark information block is formed by splicing three watermark information blocks and aggregating the watermark information through a layer of SE blocks. Input the watermark message block guided by the carrier image texture into a pre-established encoder, and output the encoded image; Input the coded image into a pre-established noise simulator and output a noise image; The noisy image is input into a pre-established decoder to extract the one-dimensional watermark information.

2. The robust watermark embedding method based on multi-dimensional information embedding and texture guidance according to claim 1 is characterized in that: The pre-established encoder transforms the carrier image I co After a ConvBNRelu block, the channels are expanded from 3 to 64 channels, and then the carrier image I is processed through four SE blocks. co The features of the image are extracted to obtain a 128*128*64 image feature block. The image feature block is concatenated with the watermark information block and fused with the watermark and image through a ConvBNRelu block to obtain a watermark image fusion feature block. The feature block is concatenated with the carrier image I co The concatenation is performed and the number of channels is reduced to 3 through a layer of convolution blocks to form the encoded image I en .

3. The robust watermark embedding method based on multi-dimensional information embedding and texture guidance according to claim 1 is characterized in that: The pre-established noise simulator simulates various noises on the coded image I en Attack and form a noise image I no ,During the end-to-end training process, the effects of noise are propagated to the encoder through the back-propagation algorithm.

4. The robust watermark embedding method based on multi-dimensional information embedding and texture guidance according to claim 1 is characterized in that: The carrier image I co and the coded image I en The discriminator distinguishes and labels the carrier image I co and the coded image I en The discriminator is trained and the trained discriminator is obtained through adversarial training.

5. The robust watermark embedding method based on multi-dimensional information embedding and texture guidance according to claim 1 is characterized in that: The pre-built decoder first transforms the noise image I through a ConvBNRelu layer no The channels are expanded to 64 channels, and then the height and width of the feature block are reduced through the SE downsampling block, and then the number of channels of the feature block is changed to 1 through the ConvBNRelu layer, and it is transformed into a one-dimensional watermark information M' through deformation and diffusion.

6. The robust watermark embedding method based on multi-dimensional information embedding and texture guidance according to claim 4 is characterized in that: The carrier image I co and the coded image I en By inputting it into the pre-trained VGG-19 model and outputting its middle layer feature image for comparison, the perceptual loss is obtained. The calculation process is as follows: Among them, λ j is the weight of layer j, J is the set of selected pre-trained network layers, C j , H j、 W j Respectively represent the number of channels, height, and width of the selected layer, φ j (y), is the carrier image I co and the coded image I en The output of layer j of the VGG-19 network.

7. Robust watermark embedding system based on multi-dimensional information embedding and texture guidance, characterized by: include: A message processing unit, used for receiving watermark information and a carrier image, inputting the watermark information and the carrier image into a pre-established message processor, and outputting a watermark message block guided by the carrier image texture; The message processor pre-established in the message processing unit includes a multi-dimensional message processing module and a texture guidance module, wherein the multi-dimensional message processing module processes the watermark information into an initial watermark information block in three different ways, and then performs feature aggregation through convolution to form a final watermark information block; The texture guidance module converts the carrier image I co Four decomposition images of different scales are obtained through SWT transformation, including low-frequency LL image, horizontal detail LH image, vertical detail HL image, and diagonal detail HH image; Or the multi-dimensional message processing module in the message processing unit includes three different ways: The first information processing method: The one-dimensional information M is transformed into a two-dimensional information graph through the MBRS information processing method, and the two-dimensional information block is expanded to a size of 128*128*64 through four deconvolution modules; The second information processing method: diffuse the watermark information through the diffusion block, and expand the information to a watermark information block of 128*128*64 in size through the first method; The third information processing method: the watermark information is enlarged to 16384, and then transformed into a two-bit information block of 128*128 size through four layers of SE blocks to process the information into a watermark information block of 128*128*64 size. Finally, the final watermark information block is formed by splicing three watermark information blocks and aggregating the watermark information through a layer of SE blocks. An encoding unit, used for inputting a watermark message block guided by the carrier image texture into a pre-established encoder, and outputting an encoded image; Noise simulation unit: input the coded image into a pre-established noise simulator and output a noise image; The decoding and extraction unit is used to input the noise image into a pre-established decoder to extract the one-dimensional watermark information.

8. The robust watermark embedding system based on multi-dimensional information embedding and texture guidance according to claim 7 is characterized in that: The encoder pre-established in the encoding unit converts the carrier image I co After a ConvBNRelu block, the channels are expanded from 3 to 64 channels, and then the carrier image I is processed through four SE blocks. co The features of the image are extracted to obtain a 128*128*64 image feature block. The image feature block is concatenated with the watermark information block and fused with the watermark and image through a ConvBNRelu block to obtain a feature block. The feature block is then combined with the carrier image I co The concatenation is performed and the number of channels is reduced to 3 through a layer of convolution blocks to form the encoded image I en ; The noise simulator pre-established in the noise simulation unit simulates various noises on the coded image I en Attack and form a noise image I no ,During the end-to-end training process, the influence of noise is propagated to the encoder through the back-propagation algorithm; Among them, the carrier image I co and the coded image I en The discriminator distinguishes and labels the carrier image I co and the coded image I en Train the discriminator and obtain the trained discriminator through adversarial training; The decoder pre-established in the decoding extraction unit first transforms the noise image I through a ConvBNRelu layer no The channel is expanded to 64 channels, and then the height and width of the feature block are reduced through the SE downsampling block, and then the number of feature block channels is changed to 1 through the ConvBNRelu layer, and it is transformed into a one-dimensional watermark information M′ through deformation and diffusion; Furthermore, the carrier image I co and the coded image I en By inputting it into the pre-trained VGG-19 model and outputting its middle layer feature image for comparison, the perceptual loss is obtained. The calculation process is as follows: Among them, λ j is the weight of layer j, J is the set of selected pre-trained network layers, C j , H j , W j Respectively represent the number of channels, height, and width of the selected layer, φ j (y), is the carrier image I co and the coded image I en The output of layer j of the VGG-19 network.

Images