Copyright information steganography method and related equipment based on arbitrary image style transfer

Abstract

This application discloses a method and related equipment for steganography of copyright information based on arbitrary image style transfer. The method includes: acquiring a pre-set steganography network, a target image, and copyright information corresponding to the target image; the steganography network includes a feature extraction layer and an embedding layer; extracting multi-dimensional features of the target image through the feature extraction layer to obtain a target sequence of the target image; embedding the copyright information into the Y channel of the target image using the target sequence through the embedding layer of the steganography network to obtain a content image; acquiring a style transfer network and a style image; extracting the style sequence and content sequence through the style transfer network to obtain a position-encoded content sequence, forming a stylized content image. Clearly, this application can steganography of copyright information into a content image and perform arbitrary image style transfer on the content image using a style transfer network, which can reduce copyright disputes.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and more specifically, to a method and related equipment for steganography of copyright information based on arbitrary image style transfer. Background Technology

[0002] Style transfer refers to combining the content of one image with the style of another artistic image to generate an artistic image. Therefore, the input to style transfer is a content image and a style image, and the output is a stylized content image. With the development of internet technology, more and more people are publishing their stylized content images through digital media. Generally, content images are works created by creators with considerable time and effort, and creators should rightfully enjoy the intellectual property rights to their works. However, while digital media can help creators better disseminate their work and help people understand creators and their works, it can also easily lead to copyright disputes. Therefore, how to protect the content images created by creators, reduce copyright disputes, and perform style transfer on the content images created by creators has become a key focus. Summary of the Invention

[0003] In view of this, this application provides a method, apparatus, device and readable storage medium for steganography of copyright information based on arbitrary image style transfer, which can reduce copyright disputes and perform style transfer on content images created by creators.

[0004] To achieve the above objectives, the following solution is proposed:

[0005] A copyright information steganography method based on arbitrary image style transfer, comprising:

[0006] Obtain a pre-set steganography network, a target image, and copyright information corresponding to the target image. The steganography network includes a feature extraction layer and an embedding layer.

[0007] Through the feature extraction layers, multi-dimensional features of the target image are extracted to obtain the target sequence of the target image;

[0008] By using the embedding layers of the steganography network, the copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image;

[0009] Obtain a style transfer network and a style image, wherein the style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder;

[0010] The Transformer Encoder extracts the style sequence of the style image and the content sequence of the content image, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image;

[0011] The content sequence is position-encoded using the position encoder to obtain a position-encoded content sequence.

[0012] The Transformer Decoder uses the position-encoded content sequence and style sequence to form a stylized content image.

[0013] Optionally, obtaining the style transfer network includes:

[0014] Acquire sample images, style training images, and an initial network. The initial network includes an initial TransformerEncoder, an initial Transformer Decoder, and an initial position encoder. The sample images are image works with corresponding copyright information implicitly written into them.

[0015] The sample image and the style training image are input into the initial network to obtain the sample sequence and style sequence with detailed features extracted by the initial Transformer Encoder. The sample sequence contains various detailed features of the sample image, and the style sequence contains the artistic style features of the style training image.

[0016] The sample sequence is position-encoded using the initial position encoder to obtain a position-encoded sample sequence.

[0017] Using the initial Transformer Decoder, a stylized sample image is formed by utilizing the position-encoded sample sequence and the style sequence;

[0018] Based on the image artwork, the style training image, and the stylized sample image, calculate the total loss of the initial network;

[0019] Based on the total loss, the parameters of the initial network are adjusted until the total loss of the initial network reaches the minimum value that the initial network can achieve. The initial network obtained by the final training is used as the style transfer network, the initial Transformer Encoder obtained by the final training is used as the Transformer Encoder, the initial position encoder obtained by the final training is used as the position encoder, and the initial Transformer Decoder obtained by the final training is used as the Transformer Decoder.

[0020] Optionally, based on the image artwork, the style training images, and the stylized sample images, the total loss of the initial network is calculated, including:

[0021] Obtain the pre-trained convolutional neural network VGG-19;

[0022] The VGG-19 network is used to extract sample feature maps of the image works, style feature maps of the style training images, and transfer feature maps of the stylized sample images;

[0023] By comparing the differences between the sample feature map and the transfer feature map, the sample loss value corresponding to the initial network is obtained. The sample loss value indicates the feature information lost after the sample image is converted into a stylized sample image.

[0024] By comparing the differences between the style feature map and the transfer feature map, the style loss value corresponding to the initial network is obtained. The style loss value indicates the ability of the stylized sample image to express the artistic style of the style training image.

[0025] Based on the sample loss value and the style loss value, the total loss of the initial network is calculated.

[0026] Optionally, calculating the total loss of the initial network based on the sample loss value and the style loss value includes:

[0027] Stylized image works are formed by using the initial network, the style training images, and the image works;

[0028] By comparing the stylized image artwork with the stylized sample image, the encoding loss value corresponding to the initial network is determined. The encoding loss value indicates the similarity between the stylized sample image and the stylized image artwork.

[0029] The stylized sample image is decoded to obtain the decoded copyright information;

[0030] By comparing the decoded copyright information with the copyright information corresponding to the image work, the decoding loss value corresponding to the initial network is determined;

[0031] Based on the ability of the encoding loss value, the decoding loss value, the sample loss value, and the style loss value to represent each parameter of the initial network, determine the weights of the encoding loss value, the decoding loss value, the sample loss value, and the style loss value in the total loss of the initial network;

[0032] The total loss of the initial network is calculated based on the encoding loss value and its weight in the total loss of the initial network, the decoding loss value and its weight in the total loss of the initial network, the sample loss value and its weight in the total loss of the initial network, and the style loss value and its weight in the total loss of the initial network.

[0033] Optionally, after forming a stylized content image using the position-encoded content sequence and style sequence through the Transformer Decoder, the method further includes:

[0034] The target sequence is position-encoded using the position encoder to obtain a position-encoded target sequence.

[0035] The Transformer Decoder uses the position-encoded target sequence and the style sequence to form a stylized target image.

[0036] Obtain the Y-channel information of the stylized content image and the Y-channel information of the stylized target image;

[0037] Subtract the Y-channel information of the stylized target image from the Y-channel information of the stylized content image to obtain a difference image;

[0038] The Y-channel information of the difference image and the Y-channel information of the stylized content image are added together to obtain a summed image, which is then used as the final stylized content image.

[0039] Optionally, before extracting multi-dimensional features of the target image through the feature extraction layer to obtain the target sequence of the target image, the method further includes:

[0040] The target image and the copyright information are sized and aligned to obtain the aligned copyright information.

[0041] By embedding the copyright information into the Y channel of the target image using the target sequence through the embedding layer of the steganography network, a content image is obtained, including:

[0042] By using the embedding layers of the steganography network, aligned copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image.

[0043] A copyright information steganography device based on arbitrary image style transfer, comprising:

[0044] A steganalysis network acquisition unit is used to acquire a preset steganalysis network, a target image, and copyright information corresponding to the target image. The steganalysis network includes a feature extraction layer and an embedding layer.

[0045] The target sequence extraction unit is used to extract multi-dimensional features of the target image through the feature extraction layer to obtain the target sequence of the target image;

[0046] The copyright information utilization unit is used to embed the copyright information into the Y channel of the target image through the embedding layer of the steganography network and the target sequence to obtain a content image;

[0047] A style transfer network acquisition unit is used to acquire a style transfer network, a style image, and a content image. The style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder.

[0048] The content sequence extraction unit is used to extract the style sequence of the style image and the content sequence of the content image through the Transformer Encoder, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image;

[0049] A position encoding unit is used to perform position encoding on the content sequence through the position encoder to obtain a position-encoded content sequence;

[0050] The style forming unit is used to form a stylized content image by using the position-encoded content sequence and style sequence through the Transformer Decoder.

[0051] A copyright information steganography device based on arbitrary image style transfer, comprising a memory and a processor;

[0052] The memory is used to store programs;

[0053] The processor is used to execute the program to implement the various steps of the copyright information steganography method based on arbitrary image style transfer described above.

[0054] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the copyright information steganography method based on arbitrary image style transfer as described above.

[0055] As can be seen from the above technical solution, the copyright information steganography method based on arbitrary image style transfer provided in this application can obtain a preset steganography network, a target image, and copyright information corresponding to the target image. The steganography network includes a feature extraction layer and an embedding layer. Through the feature extraction layer, multi-dimensional features of the target image are extracted to obtain a target sequence of the target image. Through the embedding layer of the steganography network, the copyright information is embedded into the Y channel of the target image using the target sequence to obtain a content image. Through the above steps, copyright information can be steganographically written into the target image to obtain a content image. Subsequently, a style transfer network, a style image, and a content image can be obtained. The style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder. Through the Transformer Encoder, the style sequence of the style image and the content sequence of the content image are extracted. The style sequence indicates the artistic style of the style image, and the content sequence contains various detailed features of the content image. Through the position encoder, the content sequence is positionally encoded to obtain a positionally encoded content sequence. Through the Transformer... The decoder uses the position-encoded content sequence and style sequence to form a stylized content image. Therefore, this application can embed copyright information into the content image and utilize the Transformer Encoder, Transformer Decoder, and position encoder of the style transfer network to perform style transfer on the content image. Thus, the stylized content image contains copyright information, reducing copyright disputes.

[0056] Furthermore, the style transfer network of this application combines a Transformer Encoder, a Transformer Decoder, and a positional encoder, possessing a self-attention mechanism. This allows it to more easily learn global information from both the style and content images, and it exhibits strong feature representation capabilities, preventing the loss of detailed features in both the style and content images, thus generating more stylized content images. Additionally, this application embeds copyright information into the Y channel of the target image, effectively reducing the embedding rate required for embedding copyright information, thereby increasing the embedding capacity, reducing distortion, and improving the visual experience. Moreover, the Y channel contains a high amount of semantic information, further improving the distortion rate of the embedded information, generating better content images, and enhancing the reliability and accuracy of this application. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0058] Figure 1 This is a flowchart of a copyright information steganography method based on arbitrary image style transfer disclosed in an embodiment of this application;

[0059] Figure 2 This is a schematic diagram of a copyright information steganography device based on arbitrary image style transfer disclosed in an embodiment of this application;

[0060] Figure 3 This is a hardware structure block diagram of a copyright information steganography device based on arbitrary image style transfer disclosed in an embodiment of this application. Detailed Implementation

[0061] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0062] The copyright information steganography method based on arbitrary image style transfer provided in this application can be applied to the field of image processing technology. This method can be applied to various system platforms, and the hardware device supporting this system platform can be a PC, terminal, or other processing device. The execution entity of the method can be the processor of the system platform. Furthermore, testing has verified that the copyright information steganography method based on arbitrary image style transfer of this application can be deployed on a Raspberry Pi. After connecting to a network, the Raspberry Pi carrying the copyright information steganography method based on arbitrary image style transfer of this application can be deployed as a mobile server. The Raspberry Pi (RPi) is a microcomputer only the size of a credit card.

[0063] Next, combine Figure 1 The copyright information steganography method based on arbitrary image style transfer proposed in this application is described in detail, including the following steps:

[0064] S1. Obtain a preset steganalysis network, a target image, and copyright information corresponding to the target image. The steganalysis network includes a feature extraction layer and an embedding layer.

[0065] Specifically, a steganalysis network can be pre-trained to embed copyright information corresponding to the target image into the target image, forming a content image. Its inputs are the copyright information and the target image, and its output is the content image. Before inputting the copyright information and the target image into the steganalysis network, the copyright information and the target image can be aligned.

[0066] The steganography network can include feature extraction layers and embedding layers.

[0067] The feature extraction layer is used to extract the target sequence of the target image, and the embedding layer is used to embed copyright information into the target sequence to form a content image.

[0068] The target image can be an artwork created by the creator. By embedding copyright information that reflects the creator's identity into the target image, a content image can be obtained.

[0069] S2. Through the feature extraction layer, extract multi-dimensional features of the target image to obtain the target sequence of the target image.

[0070] Specifically, the feature extraction layer can include downsampling and upsampling. Downsampling is composed of UNet, in which the BackBone is replaced with ResNet, which has a stronger learning ability, making the downsampling process similar to ResNet-50. The pooling layer in the initial layer is replaced with a convolutional layer to better preserve features and improve the appearance. The ReLU activation function is replaced with the LeakyReLU activation function. A CBAM layer is added at the end of each residual block.

[0071] ResNet is also known as a residual network. ResNet is constructed from residual building blocks.

[0072] LeakyReLU is a commonly used activation function in deep learning.

[0073] CBAM is a simple yet effective attention module for convolutional neural networks. Given an intermediate feature map in the convolutional neural network, CBAM injects the attention map along both the channel and spatial dimensions of the feature map, and then multiplies the attention by the input feature map, adaptively refining the input feature map. CBAM is a general-purpose end-to-end module.

[0074] The upsampling process includes five deconvolutional modules, each consisting of a deconvolutional layer, an IN layer, a LeakyReLU activation function layer, and a CBAM layer. Finally, a Skip Connection method similar to UNet is used to merge the residual blocks and deconvolutional blocks in the same layer, resulting in more comprehensive information.

[0075] The spatial attention and channel attention can be improved by downsampling and upsampling at the feature extraction level, so as to extract multi-dimensional features of the target image and restore the size of the target image to obtain the target sequence of the target image.

[0076] S3. Through the embedding layer of the steganography network, the copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image.

[0077] Specifically, the embedding layer allows copyright information to be embedded into the Y channel of the target image to obtain a content image.

[0078] Copyright information corresponding to the content image can be embedded in the content image.

[0079] The content image can be a work created by a creator, and the copyright information is information used by the creator to indicate that the work was created by the copyright holder. A steganography network can be used to embed the copyright information into the work; therefore, the content image can have its corresponding copyright information embedded through a steganography network.

[0080] In the process of writing copyright information into a work using a steganography network, the copyright information and the work can be converted to a 256*256 format. The information contained in the Y channel of the work can then be extracted and concatenated with the copyright information matrix to obtain the content image. This step allows the copyright information and the work to be displayed without size limitations.

[0081] `concat` is used to join two or more arrays. It does not modify the existing arrays, but only returns a copy of the joined arrays.

[0082] Step S4: Obtain the style transfer network and style image. The style transfer network includes TransformerEncoder, TransformerDecoder and position encoder.

[0083] Specifically, the style transfer network may include a Transformer Encoder, a Transformer Decoder, and a position encoder.

[0084] The style images can be various publicly available classic paintings.

[0085] Generally, style images can be obtained from the iCartoonFace dataset. Among them, iCartoonFace is currently the largest cartoon media dataset in the field of image recognition, and it is of high quality, richly annotated, and comprehensive in content, including similar images, occluded images, and images with changes in appearance.

[0086] Transformer Encoder and Transformer Decoder are relational modeling structures where different layers can extract similar structural information. Therefore, after adopting Transformer Encoder and Transformer Decoder, the style transfer network in this application has a strong feature representation capability.

[0087] Step S5: Extract the style sequence of the style image and the content sequence of the content image using the Transformer Encoder.

[0088] Specifically, a style transfer network can contain two Transformer Encoders: one encoding the content sequence of the content image and the other encoding the style sequence of the style image. The style sequence can be features that reflect the artistic style of the style image, such as spatial features, color features, and channel features. The content sequence is a sequence of various structural features, object features, and channel features of the content image that reflects multiple aspects of the content image's features.

[0089] The Transformer Encoder encodes style or content images through a self-attention mechanism and a feedforward neural network, which can improve channel attention and spatial attention, resulting in style and content sequences containing more detailed features and global information.

[0090] Step S6: The content sequence is position-encoded using the position encoder to obtain the position-encoded content sequence.

[0091] Specifically, a position encoder can be used to encode the content sequence to obtain a position-encoded content sequence, so as to maintain the structural features of the content image during the process of converting the content image into a stylized content image.

[0092] The positional encoder can be CAPE (Content-Aware Positional Encoding), which is suitable for visual tasks with varying sizes and has scaling invariance. CAPE generates positional codes based on the content sequence. During the encoding process, CAPE first uses adaptive average pooling to fix the size of the input content sequence to [n,n], then uses a 1x1 convolution to model and learn the positional codes, and finally interpolates and scales the positional codes back to the size of the input content sequence.

[0093] Step S7: Using the Transformer Decoder, a stylized content image is formed by utilizing the position-encoded content sequence and style sequence.

[0094] Specifically, the Transformer Decoder of this application comprises two layers: two MSA (Measurement Systems Analysis) layers and one FFN (feed-forward network) layer. The input includes a position-encoded content sequence Ys and a style sequence Yc. Yc is used to generate Q, and Ys is used to generate K and V. The output X can be obtained through the following formula:

[0095]

[0096]

[0097]

[0098] The output X is then refined using a three-layer CNN decoder to obtain a stylized content image.

[0099] As can be seen from the above technical solution, the stylized content images formed by this application contain copyright information, which can reduce copyright disputes and make the works created by the creators more artistic and artistic.

[0100] Furthermore, the style transfer network of this application combines Transformer Encoder, Transformer Decoder and position encoder with a self-attention mechanism, which makes it easier to learn the global information of style network and content image, and has a strong feature representation capability. It can avoid the loss of detailed features of style image and content image, and can generate stylized content image better.

[0101] Furthermore, by embedding copyright information into the Y channel of the target image, this application effectively reduces the embedding rate required for embedding copyright information, thereby increasing the embedding capacity and reducing distortion for a better viewing experience. Moreover, the Y channel contains a high amount of semantic information, which further improves the distortion rate of the embedded information, resulting in better content image generation and enhanced reliability and accuracy.

[0102] In some embodiments of this application, the process of obtaining the style transfer network in step S4 is described in detail, and the steps are as follows:

[0103] S40. Obtain sample images, style training images, and an initial network. The initial network includes an initial Transformer Encoder, an initial Transformer Decoder, and an initial position encoder. The sample images are image works with corresponding copyright information implicitly written into them.

[0104] Specifically, image works can be selected from datasets such as ImageNet, ALASKA, BOSSBase, DIV2K, and Flickr2k. The source of the image work can be used as copyright information. This copyright information can be text information, such as when the image work comes from ImageNet, the copyright information can be ImageNet, or it can be an image that contains ImageNet.

[0105] Copyright information can be embedded into images selected from a dataset using a steganography network. The sample images are images with copyright information embedded in them through the steganography network.

[0106] ImageNet is the name of a computer vision system recognition project and is currently the world's largest image recognition database; ALASKA is an image steganography dataset; BOSSBase is a standardized dataset containing 10,000 carrier images; DIV2K contains 1,000 high-resolution 2K images; Flickr2K contains images of people, animals, and landscapes.

[0107] Generally, style training images can be obtained from the WikiArt dataset. WikiArt is a platform dedicated to showcasing world-famous paintings for art lovers.

[0108] The initial network is configured with an initial Transformer Encoder, an initial Transformer Decoder, and an initial position encoder. The hyperparameters of the initial network can be set according to actual needs. During training, the parameters of the initial Transformer Encoder, the initial Transformer Decoder, and the initial position encoder can be adjusted.

[0109] S41. Input the sample image and the style training image into the initial network to obtain the sample sequence and style sequence with detailed features extracted by the initial Transformer Encoder. The sample sequence contains various detailed features of the sample image, and the style sequence contains the artistic style features of the style training image.

[0110] Specifically, sample images and style training images can be input into the initial network. The initial Transformer Encoder extracts sample sequences and style sequences with detailed features. The style sequence can contain the artistic style features of the style training image, that is, the style sequence can indicate the artistic style of the style training image, and the sample sequence can contain various detailed features of the sample image.

[0111] S42. The sample sequence is position-encoded using the initial position encoder to obtain the position-encoded sample sequence.

[0112] Specifically, the sample sequence can be position-encoded using an initial CPAE to obtain a position-encoded sample sequence.

[0113] S43. Using the initial Transformer Decoder, a stylized sample image is formed by utilizing the position-encoded sample sequence and the style sequence.

[0114] Specifically, the output can be predicted using two MSA layers and one FFN layer of the Transformer Decoder, and the predicted output can be refined by passing it through a three-layer CNN decoder to obtain stylized sample images.

[0115] S44. Based on the image artwork, the style training image, and the stylized sample image, calculate the total loss of the initial network.

[0116] Specifically, the total loss of the initial network and the total loss of the steganalysis network can be determined by comparing the sample images, stylized image works, stylized training images, and the stylized sample images.

[0117] S45. Based on the total loss, adjust the parameters of the initial network until the total loss of the initial network reaches the minimum value that the initial network can achieve. Then, use the final trained initial network as the style transfer network, the final initial Transformer Encoder as the Transformer Encoder, the final initial position encoder as the position encoder, and the final initial Transformer Decoder as the Transformer Decoder.

[0118] Specifically, the parameters of the initial network are adjusted based on the total loss of the initial network. When the loss of the initial network no longer decreases after parameter adjustment, the initial network obtained at this point can be used as the style transfer network. At the same time, the parameters of the steganalysis network can be adjusted based on the total loss of the steganalysis network. When the loss of the steganalysis network no longer decreases after parameter adjustment, the steganalysis network obtained at this point can be used as the trained steganalysis network.

[0119] As can be seen from the above technical solution, this embodiment provides an optional method for training a style transfer network. The Transformer Decoder, Transformer Encoder, and position encoder of the style transfer network can be trained in the above way. After training in the above way, the spatial attention and channel attention of the style transfer network can be improved, and the stylized sample image output by the stylized style transfer network can retain the structural information of the original sample image.

[0120] In some embodiments of this application, the process of calculating the total loss of the initial network based on the image artwork, the style training image, and the stylized sample image is described in detail below:

[0121] S440. Obtain the pre-trained convolutional neural network VGG-19.

[0122] Specifically, a pre-trained VGG-19 network can be obtained. The VGG-19 network is a convolutional neural network that can extract feature maps from input images and compare the similarity of feature maps between input images.

[0123] S441. Using the VGG-19 network, extract the sample feature map of the image work, the style feature map of the style training image, and the transfer feature map of the stylized sample image.

[0124] Specifically, sample images, style training images, and stylized sample images can be input into the VGG-19 network to obtain sample feature maps of the sample images extracted by the ReLU2-2 layer of the VGG-19 network, style feature maps of the style training images, and transfer feature maps of the stylized sample images.

[0125] S442. Compare the differences between the sample feature map and the transfer feature map to obtain the sample loss value corresponding to the initial network. The sample loss value indicates the feature information lost after the sample image is converted into a stylized sample image.

[0126] Specifically, the sample loss value can be calculated using the following formula:

[0127]

[0128] Wherein, φ(X) s ) is the transfer feature map, φ(X) c ) is the sample feature map, w is the width of the transfer feature map and the sample feature map, Xc is the image artwork, and Xs is the stylized sample image. It is the square of the 2-norm distance between the transferred feature map and the sample feature map.

[0129] S443. Compare the differences between the style feature map and the transfer feature map to obtain the style loss value corresponding to the initial network. The style loss value indicates the ability of the stylized sample image to express the artistic style of the style training image.

[0130] Specifically, the style loss value can be calculated using the following formula:

[0131]

[0132]

[0133] Where h is the length of the transfer feature map and the sample feature map, w is the width of the transfer feature map and the sample feature map, and X... Y It is a style training image, φ(X) s ) is the transfer feature map, Xs is the stylized sample image, and φ(X) is the transfer feature map. Y G(X) is a style feature map. S ) represents X S The corresponding Gram matrix, G(X) Y ) represents X Y The corresponding Gram matrix, j indicates the j-th layer of VGG-19, and c is the number of channels in the transfer feature map and the sample feature map. For G j (X Y ) and G j (X S The square of the 2-norm distance between ).

[0134] We can first treat Xs as X in G(X), and then use the formula corresponding to G(X) and φ(X) as described above. s ) Calculate G(X) S ).

[0135] X can Y As X in G(X), and using the formula corresponding to G(X) above and φ(X) Y ) Calculate G(X) Y ).

[0136] S444. Based on the sample loss value and the style loss value, calculate the total loss of the initial network.

[0137] Specifically, the weights of the sample loss value and the style loss value in the total loss of the initial network can be determined, and the total loss of the initial network can be calculated based on the sample loss value and its corresponding weight, and the style loss value and its corresponding weight.

[0138] As can be seen from the above technical solution, this embodiment provides an optional method for calculating the total loss of the initial network. It considers the total loss of the initial network from two aspects: the difference between the stylized sample image and the original image work, and the difference between the stylized sample image and the style training image. The initial network is trained based on the total loss, which can accelerate the convergence speed of the initial network and provide a better style transfer network for this application.

[0139] In some embodiments of this application, the process of calculating the total loss of the initial network based on the sample loss value and the style loss value in step S444 is described as follows:

[0140] S4440. Using the initial network, the style training image, and the image artwork, a stylized image artwork is formed.

[0141] Specifically, the sequence of image works with detailed features extracted from the initial Transformer Encoder of the initial network and the style sequence of style training images can be used.

[0142] The work sequence is positionally encoded by the initial position encoder to obtain the positionally encoded work sequence.

[0143] Stylized image works are formed by using the initial Transformer Decoder, the position-encoded sequence of works, and the style sequence of style training images.

[0144] S4441. Compare the stylized image artwork and the stylized sample image to determine the encoding loss value corresponding to the initial network. The encoding loss value indicates the similarity between the stylized sample image and the stylized image artwork.

[0145] Specifically, the coding loss value corresponding to the initial network can be calculated using the following formula:

[0146]

[0147] Among them, I style For stylized image works, For the stylized sample images, α and β are weight parameters that can be set according to actual needs. For I styleand LPIPS values ​​between For I style and The Smooth L1 value between.

[0148] LPIPS is a more effective evaluation metric for judging image similarity. Smooth L1 is a loss function, meaning the smoothed L1, which, compared to the L1 norm loss function, can correct the zero-point non-smoothing problem, making the initial network learn faster and converge more easily.

[0149] S4442. Decode the stylized sample image to obtain the decoded copyright information.

[0150] Specifically, considering that the generated stylized sample images also need to contain copyright information, copyright disputes can easily arise if the copyright information cannot be obtained through decoding the stylized sample images. Therefore, when calculating the total loss of the initial network, it is necessary to consider whether the difference between the copyright information after decoding the stylized sample images and the original copyright information is too large.

[0151] Therefore, stylized sample images can be input into a pre-defined STC decoder to obtain the copyright information output by the STC decoder. The STC decoder's decoding process includes downsampling and upsampling. Downsampling is similar to ResNet-18. The upsampling consists of five deconvolutional modules, each composed of a deconvolutional layer, an IN layer, a LeakyReLU activation function layer, and a CBAM layer. Finally, using a Skip Connection method similar to UNet, residual blocks and deconvolutional blocks in the same layer are merged to obtain more comprehensive information.

[0152] Decoding can be performed using the following formula: Ext(Y) = HY, where H ∈ {0, 1} m×n Let m be the public key, and let m be the parity check matrix of the binary linear code C. C(M) is the coset of m, i.e., C(M) = {z∈{0,1}. n |Hz=M}, where M is the decoded copyright information.

[0153] S4443. Compare the decoded copyright information with the copyright information corresponding to the image work to determine the decoding loss value corresponding to the initial network.

[0154] Specifically, the decoding loss value indicates the difference between the copyright information corresponding to the image work and the copyright information in the stylized sample image.

[0155] The decoding loss can be determined using the following formula:

[0156]

[0157] Among them, I c Copyright information for the image works. To decode copyright information, BER stands for bit error rate. For I c and The bit error rate between them. For I c and LPIPS values ​​between For I c and The smooth L1 value between ρ, σ, and τ. ρ, σ, and τ are weighting parameters that can be set according to actual needs.

[0158] S4444. Based on the ability of the encoding loss value, the decoding loss value, the sample loss value, and the style loss value to represent each parameter of the initial network, determine the weights of the encoding loss value, the decoding loss value, the sample loss value, and the style loss value in the total loss of the initial network.

[0159] Specifically, the weights of the encoding loss, decoding loss, sample loss, and style loss in the total loss of the initial network can be determined based on actual needs and / or the importance of the parameters corresponding to each loss value to the initial model. For example, copyright information is a relatively important part of the stylized sample images, so the weights corresponding to the decoding loss and encoding loss values ​​that affect copyright information can be set to be larger.

[0160] S4445. Calculate the total loss of the initial network based on the encoding loss value and its weight in the total loss of the initial network, the decoding loss value and its weight in the total loss of the initial network, the sample loss value and its weight in the total loss of the initial network, and the style loss value and its weight in the total loss of the initial network.

[0161] Specifically, the total loss of the initial network can be calculated using the following formula:

[0162] l total =λl encode +μl decode +ηl content +υl style

[0163] Among them, l encode l is the encoding loss value. decode For decoding loss value, l content Let l be the sample loss value. styleη is the style loss value, λ is the weight of the encoding loss value in the total loss, μ is the weight of the decoding loss value in the total loss, η is the weight of the sample loss value in the total loss, and υ is the weight of the style loss value in the total loss.

[0164] As can be seen from the above technical solution, this embodiment provides an optional method for calculating the total loss. Through the above process, when calculating the total loss of the initial network, this application not only considers the differences between the stylized sample image and the original image work, and the differences between the stylized sample image and the style training image, but also considers the changes in copyright information caused in the process of converting the image work into the stylized sample image. This allows for a more comprehensive adjustment of the various parameters of the initial network, thereby improving the efficiency and reliability of training the style transfer network in this application.

[0165] In some embodiments of this application, to facilitate the processing of copyright information and target images by the steganography network, before step S2, which involves extracting multi-dimensional features of the target image through the feature extraction layer to obtain the target sequence of the target image, the following steps are also included:

[0166] S20. Align the target image and the copyright information by size to obtain aligned copyright information.

[0167] Specifically, before inputting the copyright information and target image into the steganography network, the copyright information can be converted into a 256*256 copyright information matrix, and the target image can be aligned to a size of 256*256.

[0168] Based on this, the process of embedding the copyright information into the Y channel of the target image using the target sequence through the embedding layer of the steganography network to obtain the content image is described in detail below:

[0169] S30. Through the embedding layer of the steganography network, the aligned copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image.

[0170] Specifically, the embedding layer can embed the aligned copyright information into the Y channel of the aligned target image to obtain the content image.

[0171] As can be seen from the above technical solution, this embodiment can process the format and size of the target image and copyright information before inputting them into the steganography network, making it easier for the steganography network to generate content images and making this application more efficient.

[0172] In some embodiments of this application, in order to increase the embedding capacity and make the final stylized content image more visually appealing, after forming the stylized content image in step S7 using the Transformer Decoder with the position-encoded content sequence and style sequence, a processing procedure for the stylized content image can be added. This process will be described in detail below, with the steps as follows:

[0173] S8. The target sequence is position-encoded using the position encoder to obtain the position-encoded target sequence.

[0174] Specifically, a position editor can be used to positionally encode the target sequence in order to maintain the structural features of the target image.

[0175] S9. Using the Transformer Decoder, a stylized target image is formed by utilizing the position-encoded target sequence and the style sequence.

[0176] Specifically, a style transfer network can be used to transfer the artistic style of a style image to a target image, resulting in a stylized target image.

[0177] S10. Obtain the Y-channel information of the stylized content image and the Y-channel information of the stylized target image.

[0178] Specifically, it can acquire the Y-channel information of the stylized content image and the Y-channel information of the stylized target image.

[0179] S11. Subtract the Y-channel information of the stylized target image from the Y-channel information of the stylized content image to obtain a difference image.

[0180] Specifically, the Y-channel information of the stylized content image and the Y-channel information of the stylized target image can be subtracted to obtain a difference image.

[0181] S12. Add the Y channel information of the difference image and the Y channel information of the stylized content image to obtain a summed image, and use the summed image as the final stylized content image.

[0182] Specifically, the Y-channel information of the difference image and the Y-channel information of the stylized content image can be added together to obtain the final stylized content image.

[0183] As can be seen from the above technical solution, this embodiment provides a way that is more visually appealing and better able to represent the works created by the creator. The above method can not only increase the upper limit of the copyright information that can be embedded in this application, but also better showcase the creator's works.

[0184] The copyright information steganography apparatus based on arbitrary image style transfer provided in the embodiments of this application is described below. The copyright information steganography apparatus based on arbitrary image style transfer described below can be referred to in correspondence with the copyright information steganography method based on arbitrary image style transfer described above.

[0185] See Figure 2 , Figure 2 This is a schematic diagram of a copyright information steganography device based on arbitrary image style transfer disclosed in an embodiment of this application.

[0186] like Figure 2 As shown, the copyright information steganography device based on arbitrary image style transfer may include:

[0187] The steganalysis network acquisition unit 1 is used to acquire a preset steganalysis network, a target image, and copyright information corresponding to the target image. The steganalysis network includes a feature extraction layer and an embedding layer.

[0188] The target sequence extraction unit 2 is used to extract multi-dimensional features of the target image through the feature extraction layer to obtain the target sequence of the target image;

[0189] Copyright information utilization unit 3 is used to embed the copyright information into the Y channel of the target image through the embedding layer of the steganography network and the target sequence to obtain a content image;

[0190] Style transfer network acquisition unit 4 is used to acquire style transfer network, style image and content image, wherein the style transfer network includes Transformer Encoder, Transformer Decoder and position encoder;

[0191] The content sequence extraction unit 5 is used to extract the style sequence of the style image and the content sequence of the content image through the Transformer Encoder, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image;

[0192] The position encoding unit 6 is used to perform position encoding on the content sequence through the position encoder to obtain the position-encoded content sequence;

[0193] Style forming unit 7 is used to form a stylized content image by using the position-encoded content sequence and style sequence through the Transformer Decoder.

[0194] Optionally, the style transfer network acquisition unit mentioned above may include:

[0195] The sample image acquisition subunit is used to acquire sample images, style training images, and an initial network. The initial network includes an initial Transformer Encoder, an initial Transformer Decoder, and an initial position encoder. The sample images are image works with corresponding copyright information implicitly written into them.

[0196] The sample sequence acquisition subunit is used to input the sample image and the style training image into the initial network to obtain the sample sequence and style sequence with detailed features extracted by the initial Transformer Encoder. The sample sequence contains various detailed features of the sample image, and the style sequence contains the artistic style features of the style training image.

[0197] The sample sequence encoding subunit is used to perform position encoding on the sample sequence using the initial position encoder to obtain a position-encoded sample sequence.

[0198] The stylized sample image acquisition subunit is used to form a stylized sample image by using the position-encoded sample sequence and the style sequence through the initial Transformer Decoder;

[0199] The total loss calculation subunit is used to calculate the total loss of the initial network based on the image artwork, the style training image, and the stylized sample image;

[0200] The parameter tuning subunit is used to adjust the parameters of the initial network based on the total loss until the total loss of the initial network reaches the minimum value that the initial network can achieve. The final trained initial network is used as the style transfer network, the final initial Transformer Encoder is used as the Transformer Encoder, the final initial position encoder is used as the position encoder, and the final initial Transformer Decoder is used as the Transformer Decoder.

[0201] Optionally, the total loss calculation subunit of this application may include:

[0202] VGG-19 network acquisition sub-unit, used to acquire pre-trained convolutional neural network VGG-19 network;

[0203] The feature map extraction subunit is used to extract sample feature maps of the image work, style feature maps of the style training images, and transfer feature maps of the stylized sample images using the VGG-19 network.

[0204] The sample feature map comparison subunit is used to compare the differences between the sample feature map and the transfer feature map to obtain the sample loss value corresponding to the initial network. The sample loss value indicates the feature information lost after the sample image is converted into a stylized sample image.

[0205] The style feature map comparison subunit is used to compare the differences between the style feature map and the transfer feature map to obtain the style loss value corresponding to the initial network. The style loss value indicates the ability of the stylized sample image to express the artistic style of the style training image.

[0206] The style loss value is utilized by a sub-unit to calculate the total loss of the initial network based on the sample loss value and the style loss value.

[0207] Optionally, the style loss value mentioned above can utilize sub-units that include:

[0208] The initial network utilizes sub-units to form stylized image works using the initial network, the style training images, and the image works;

[0209] The encoding loss value determination subunit is used to compare the stylized image artwork and the stylized sample image to determine the encoding loss value corresponding to the initial network. The encoding loss value indicates the similarity between the stylized sample image and the stylized image artwork.

[0210] An image decoding subunit is used to decode the stylized sample image to obtain decoded copyright information;

[0211] The decoding loss value determination subunit is used to compare the decoded copyright information with the copyright information corresponding to the image work to determine the decoding loss value corresponding to the initial network.

[0212] The weight determination subunit is used to determine the weights of the encoding loss value, decoding loss value, sample loss value, and style loss value in the total loss of the initial network based on the ability of the encoding loss value, decoding loss value, sample loss value, and style loss value to represent each parameter of the initial network.

[0213] The decoding loss value utilization subunit is used to calculate the total loss of the initial network based on the encoding loss value and its weight in the total loss of the initial network, the decoding loss value and its weight in the total loss of the initial network, the sample loss value and its weight in the total loss of the initial network, and the style loss value and its weight in the total loss of the initial network.

[0214] Optionally, the aforementioned copyright information steganography device based on arbitrary image style transfer may further include:

[0215] The position encoder utilization unit is used to perform position encoding on the target sequence through the position encoder to obtain the position-encoded target sequence;

[0216] A stylized target image acquisition unit is used to form a stylized target image by using the Transformer Decoder with the position-encoded target sequence and the style sequence;

[0217] Y-channel information acquisition unit, used to acquire the Y-channel information of the stylized content image and the Y-channel information of the stylized target image;

[0218] The difference image acquisition unit is used to subtract the Y channel information of the stylized target image from the Y channel information of the stylized content image to obtain the difference image;

[0219] The image acquisition unit is used to add the Y channel information of the difference image and the Y channel information of the stylized content image to obtain an image, and to use the image as the final stylized content image.

[0220] Optionally, the aforementioned copyright information steganography device based on arbitrary image style transfer may further include:

[0221] The size alignment unit is used to align the target image and the copyright information to obtain the aligned copyright information.

[0222] Based on this, the aforementioned content image acquisition unit may include:

[0223] The copyright information embedding unit is used to embed aligned copyright information into the Y channel of the target image through the embedding layer of the steganography network, thereby obtaining a content image.

[0224] The copyright information steganography device based on arbitrary image style transfer provided in this application embodiment can be applied to copyright information steganography devices based on arbitrary image style transfer, such as PC terminals, Raspberry Pi, cloud platforms, servers, and server clusters. Optionally, Figure 3 The hardware structure block diagram of a copyright information steganography device based on arbitrary image style transfer is shown. (Refer to...) Figure 3 The hardware structure of a copyright information steganography device based on arbitrary image style transfer may include: at least one processor 1, at least one communication interface 2, at least one memory 3 and at least one communication bus 4;

[0225] In this embodiment of the application, the number of processor 1, communication interface 2, memory 3, and communication bus 4 is at least one, and processor 1, communication interface 2, and memory 3 communicate with each other through communication bus 4;

[0226] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0227] Memory 3 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0228] The memory stores a program, which the processor can call. The program is used for:

[0229] Obtain a pre-set steganography network, a target image, and copyright information corresponding to the target image. The steganography network includes a feature extraction layer and an embedding layer.

[0230] Through the feature extraction layers, multi-dimensional features of the target image are extracted to obtain the target sequence of the target image;

[0231] By using the embedding layers of the steganography network, the copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image;

[0232] Obtain a style transfer network and a style image, wherein the style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder;

[0233] The Transformer Encoder extracts the style sequence of the style image and the content sequence of the content image, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image;

[0234] The content sequence is position-encoded using the position encoder to obtain a position-encoded content sequence.

[0235] The Transformer Decoder uses the position-encoded content sequence and style sequence to form a stylized content image.

[0236] Optionally, the refined and extended functions of the program can be referred to the above description.

[0237] This application embodiment also provides a readable storage medium that can store a program suitable for execution by a processor, the program being used for:

[0238] Obtain a pre-set steganography network, a target image, and copyright information corresponding to the target image. The steganography network includes a feature extraction layer and an embedding layer.

[0239] Through the feature extraction layers, multi-dimensional features of the target image are extracted to obtain the target sequence of the target image;

[0240] By using the embedding layers of the steganography network, the copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image;

[0241] Obtain a style transfer network and a style image, wherein the style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder;

[0242] The Transformer Encoder extracts the style sequence of the style image and the content sequence of the content image, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image;

[0243] The content sequence is position-encoded using the position encoder to obtain a position-encoded content sequence.

[0244] The Transformer Decoder uses the position-encoded content sequence and style sequence to form a stylized content image.

[0245] Optionally, the refined and extended functions of the program can be referred to the above description.

[0246] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0247] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0248] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. The various embodiments of this application can be combined with each other. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A copyright information steganography method based on arbitrary image style transfer, characterized by, include: Obtain a pre-set steganography network, a target image, and copyright information corresponding to the target image. The steganography network includes a feature extraction layer and an embedding layer. Through the feature extraction layers, multi-dimensional features of the target image are extracted to obtain the target sequence of the target image; By using the embedding layers of the steganography network, the copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image; Obtain a style transfer network and a style image, wherein the style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder; The Transformer Encoder extracts the style sequence of the style image and the content sequence of the content image, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image; The content sequence is position-encoded using the position encoder to obtain a position-encoded content sequence. The Transformer Decoder uses the position-encoded content sequence and style sequence to form a stylized content image. The target sequence is position-encoded using the position encoder to obtain a position-encoded target sequence. The Transformer Decoder uses the position-encoded target sequence and the style sequence to form a stylized target image. Obtain the Y-channel information of the stylized content image and the Y-channel information of the stylized target image; Subtract the Y-channel information of the stylized target image from the Y-channel information of the stylized content image to obtain a difference image; The Y-channel information of the difference image and the Y-channel information of the stylized content image are added together to obtain a summed image, which is then used as the final stylized content image.

2. The copyright information steganography method based on arbitrary image style transfer according to claim 1, characterized in that, The method for obtaining the style transfer network includes: Acquire sample images, style training images, and an initial network. The initial network includes an initial TransformerEncoder, an initial Transformer Decoder, and an initial position encoder. The sample images are image works with corresponding copyright information implicitly written into them. The sample image and the style training image are input into the initial network to obtain the sample sequence and style sequence with detailed features extracted by the initial TransformerEncoder. The sample sequence contains various detailed features of the sample image, and the style sequence contains the artistic style features of the style training image. The sample sequence is position-encoded using the initial position encoder to obtain a position-encoded sample sequence. Using the initial Transformer Decoder, a stylized sample image is formed by utilizing the position-encoded sample sequence and the style sequence; Based on the image artwork, the style training image, and the stylized sample image, calculate the total loss of the initial network; Based on the total loss, the parameters of the initial network are adjusted until the total loss of the initial network reaches the minimum value that the initial network can achieve. The initial network obtained by the final training is used as the style transfer network, the initial Transformer Encoder obtained by the final training is used as the Transformer Encoder, the initial position encoder obtained by the final training is used as the position encoder, and the initial Transformer Decoder obtained by the final training is used as the Transformer Decoder.

3. The copyright information steganography method based on arbitrary image style transfer according to claim 2, characterized in that, Based on the image artwork, the style training images, and the stylized sample images, the total loss of the initial network is calculated, including: Obtain the pre-trained convolutional neural network VGG-19; The VGG-19 network is used to extract sample feature maps of the image works, style feature maps of the style training images, and transfer feature maps of the stylized sample images; By comparing the differences between the sample feature map and the transfer feature map, the sample loss value corresponding to the initial network is obtained. The sample loss value indicates the feature information lost after the sample image is converted into a stylized sample image. By comparing the differences between the style feature map and the transfer feature map, the style loss value corresponding to the initial network is obtained. The style loss value indicates the ability of the stylized sample image to express the artistic style of the style training image. Based on the sample loss value and the style loss value, the total loss of the initial network is calculated.

4. The copyright information steganography method based on arbitrary image style transfer according to claim 3, characterized in that, The step of calculating the total loss of the initial network based on the sample loss value and the style loss value includes: Stylized image works are formed by using the initial network, the style training images, and the image works; By comparing the stylized image artwork with the stylized sample image, the encoding loss value corresponding to the initial network is determined. The encoding loss value indicates the similarity between the stylized sample image and the stylized image artwork. The stylized sample image is decoded to obtain the decoded copyright information; By comparing the decoded copyright information with the copyright information corresponding to the image work, the decoding loss value corresponding to the initial network is determined; Based on the ability of the encoding loss value, the decoding loss value, the sample loss value, and the style loss value to represent each parameter of the initial network, determine the weights of the encoding loss value, the decoding loss value, the sample loss value, and the style loss value in the total loss of the initial network; The total loss of the initial network is calculated based on the encoding loss value and its weight in the total loss of the initial network, the decoding loss value and its weight in the total loss of the initial network, the sample loss value and its weight in the total loss of the initial network, and the style loss value and its weight in the total loss of the initial network.

5. The copyright information steganography method based on arbitrary image style transfer according to claim 1, characterized in that, Before extracting multi-dimensional features of the target image through the feature extraction layer to obtain the target sequence of the target image, the process further includes: The target image and the copyright information are sized and aligned to obtain the aligned copyright information. By embedding the copyright information into the Y channel of the target image using the target sequence through the embedding layer of the steganography network, a content image is obtained, including: By using the embedding layers of the steganography network, aligned copyright information is embedded into the Y channel of the target image using the target sequence to obtain the content image.

6. A copyright information steganography device based on arbitrary image style transfer, characterized in that, include: A steganalysis network acquisition unit is used to acquire a preset steganalysis network, a target image, and copyright information corresponding to the target image. The steganalysis network includes a feature extraction layer and an embedding layer. The target sequence extraction unit is used to extract multi-dimensional features of the target image through the feature extraction layer to obtain the target sequence of the target image; The copyright information utilization unit is used to embed the copyright information into the Y channel of the target image through the embedding layer of the steganography network and the target sequence to obtain a content image; A style transfer network acquisition unit is used to acquire a style transfer network, a style image, and a content image. The style transfer network includes a Transformer Encoder, a Transformer Decoder, and a position encoder. The content sequence extraction unit is used to extract the style sequence of the style image and the content sequence of the content image through the Transformer Encoder, wherein the style sequence indicates the artistic style of the style image and the content sequence contains various detailed features of the content image; A position encoding unit is used to perform position encoding on the content sequence through the position encoder to obtain a position-encoded content sequence; The style forming unit is used to form a stylized content image by using the position-encoded content sequence and style sequence through the Transformer Decoder; The position encoder utilization unit is used to perform position encoding on the target sequence through the position encoder to obtain the position-encoded target sequence; A stylized target image acquisition unit is used to form a stylized target image by using the Transformer Decoder with the position-encoded target sequence and the style sequence; Y-channel information acquisition unit, used to acquire the Y-channel information of the stylized content image and the Y-channel information of the stylized target image; The difference image acquisition unit is used to subtract the Y channel information of the stylized target image from the Y channel information of the stylized content image to obtain the difference image; The image acquisition unit is used to add the Y channel information of the difference image and the Y channel information of the stylized content image to obtain an image, and to use the image as the final stylized content image.

7. The copyright information steganography device based on arbitrary image style transfer according to claim 6, characterized in that, The style transfer network acquisition unit includes: The sample image acquisition unit is used to acquire sample images, style training images, and an initial network. The initial network includes an initial Transformer Encoder, an initial Transformer Decoder, and an initial position encoder. The sample images are image works with corresponding copyright information implicitly written into them. The sequence acquisition unit is used to input the sample image and the style training image into the initial network to obtain the sample sequence and style sequence with detailed features extracted by the initial Transformer Encoder. The sample sequence contains various detailed features of the sample image, and the style sequence contains the artistic style features of the style training image. A position encoding unit is used to perform position encoding on the sample sequence using the initial position encoder to obtain a position-encoded sample sequence; A stylized sample image unit is used to form a stylized sample image by means of the initial Transformer Decoder, the position-encoded sample sequence, and the style sequence; The total loss calculation unit is used to calculate the total loss of the initial network based on the image artwork, the style training image, and the stylized sample image; The parameter adjustment unit is used to adjust the parameters of the initial network based on the total loss until the total loss of the initial network reaches the minimum value that the initial network can achieve. The final trained initial network is used as the style transfer network, the final initial Transformer Encoder is used as the Transformer Encoder, the final initial position encoder is used as the position encoder, and the final initial Transformer Decoder is used as the Transformer Decoder.

8. A copyright information steganography device based on arbitrary image style transfer, characterized in that, Including memory and processor; The memory is used to store programs; The processor is configured to execute the program to implement each step of the copyright information steganography method based on arbitrary image style transfer as described in any one of claims 1-5.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the copyright information steganography method based on arbitrary image style transfer as described in any one of claims 1-5.

Images