Verifiable digital watermark texture generation method based on hierarchical fusion and significance perception

Abstract

The invention discloses a verifiable digital watermark texture generation method based on hierarchical fusion and significance perception, and the method comprises the steps: generating an initial watermark image through forward error correction coding and two-dimensional mapping based on an original binary information stream; based on the target style image, extracting features of the target style image through a style encoder and performing parameterization processing to construct a style vector; based on the initial watermark image, the style vector and the encryption seed, performing step-by-step fusion from macroscopic to microscopic through a fusion generator with a multi-scale layered structure to generate a fusion texture map, and training and optimizing the fusion generator by adopting a composite loss function; and based on the fused texture image, original information is recovered through decoding processing, and authenticity verification is carried out on the fused texture image based on the encryption seed. The invention provides a set of efficient, flexible and safe scheme for solving the contradiction between the concealment and robustness of the traditional watermark, and has wide application prospects in the fields of copyright protection, high-end product anti-counterfeiting and commodity packaging.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer vision and information security, and in particular relates to a verifiable digital watermark texture generation method based on hierarchical fusion and saliency perception. Background Technology

[0002] Digital watermarking technology, as an important branch of information hiding, has wide applications in digital copyright protection, content authentication, and anti-counterfeiting and traceability. The core objective of this technology is to embed specific information into digital media such as images and videos without affecting the visual quality of the medium. After years of development, digital watermarking technology has mainly formed two major technical routes: spatial domain methods and transform domain methods.

[0003] Spatial domain methods, represented by least significant bit substitution, embed information by directly modifying pixel values. They are simple to implement but have weak resistance to image compression and filtering, exhibiting poor robustness. Transform domain methods use tools such as discrete cosine transform and discrete wavelet transform to convert the carrier to the frequency domain, embedding watermarks by modifying frequency domain coefficients. While this improves robustness, significant limitations still exist.

[0004] First, traditional methods struggle to overcome the inherent contradiction between aesthetics and robustness. To ensure watermark robustness, information needs to be embedded in low-to-mid-frequency regions, but this introduces blockiness and artifacts, impairing visual quality. While embedding the watermark in high-frequency regions ensures concealment, it is susceptible to information loss when resisting lossy compression. Second, the embedding patterns of traditional watermarks are fixed and easily analyzed and cracked by attackers, posing security risks. More importantly, in applications with extremely high aesthetic requirements, such as luxury packaging and art reproduction, the regular embedding patterns of traditional watermarks can disrupt the visual effect, preventing the organic integration of watermark information and artistic design.

[0005] With the development of deep learning technology, watermarking methods based on generative adversarial networks (GANs) have achieved information hiding through style transfer, bringing new possibilities to watermarking technology. However, existing deep learning solutions still have three major drawbacks: first, the single-stage generation architecture struggles to balance macroscopic structure and microscopic details, easily leading to watermark structure distortion; second, style control relies on specific reference images, lacking flexible parameterized adjustment capabilities; and third, the indiscriminate protection strategy fails to prioritize and preserve structural features crucial for decoding. These shortcomings limit the application effectiveness of existing technologies in high-end scenarios.

[0006] Therefore, there is an urgent need in this field for a technical solution that can fundamentally solve the problem of balancing aesthetics and robustness, achieve intelligent integration of watermarks and artistic carriers, and at the same time ensure the security and practicality of watermarks. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a verifiable digital watermark texture generation method based on hierarchical fusion and saliency awareness. This method generates aesthetically pleasing watermark information with artistic design elements, ensuring that the key structure of the watermark remains intact during texture generation, thus achieving adaptive protection of the information and achieving an optimal balance between aesthetics and robustness. Furthermore, this invention parametrically controls and blends the artistic style of the generated texture, while assigning a unique and verifiable identity to the generated texture to adapt to diverse commercial design needs and enhance its security in anti-counterfeiting applications.

[0008] This invention provides a verifiable digital watermark texture generation method based on hierarchical fusion and saliency awareness, comprising the following steps:

[0009] Based on the original binary information stream, an initial watermark image is generated through forward error correction coding and two-dimensional mapping, and the initial watermark image is subjected to saliency analysis.

[0010] Based on the target style image, its features are extracted by a style encoder and parameterized to construct a style vector;

[0011] Based on the initial watermark image, the style vector, and the encryption seed, a fusion generator with a multi-scale hierarchical structure is used to perform a step-by-step fusion from macro to micro to generate a fused texture map. The fusion generator is trained and optimized using a composite loss function that includes adversarial loss, style consistency loss, and saliency-aware structure preservation loss.

[0012] Based on the fused texture map, the original information is recovered through decoding, and the authenticity of the fused texture map is verified based on the encrypted seed.

[0013] Optionally, based on the original binary information stream, an initial watermark image is generated through forward error correction coding and two-dimensional mapping, including:

[0014] Perform forward error correction coding on the original binary information stream;

[0015] The encoded data stream is converted into an initial watermark image with structural features using a two-dimensional mapping algorithm.

[0016] Optionally, the process of performing saliency analysis on the initial watermark image includes:

[0017] Key structural features in the initial watermark image are identified using corner detection algorithms or edge detection algorithms.

[0018] A saliency weight map with the same size as the initial watermark image is generated based on the recognition results, wherein the regions corresponding to the key structural features are assigned higher weights.

[0019] Optionally, based on the target style image, its features are extracted by a style encoder and parameterized to construct a style vector, including:

[0020] A style encoder is used to extract multi-scale style feature maps from the target style image;

[0021] Global statistics or pooling operations are performed on the multi-scale style feature maps to obtain low-dimensional style vectors.

[0022] Optionally, based on the initial watermark image, the style vector, and the encryption seed, a fusion generator with a multi-scale hierarchical structure performs a step-by-step fusion from macro to micro to generate a fused texture map, including:

[0023] The fusion generator includes a basic generation layer and several optimization and enhancement layers;

[0024] In the basic generation layer, a low-resolution feature map is generated based on the initial watermark image, the low-frequency texture features of the style vector, and the encryption seed.

[0025] In at least one optimization layer, the feature map output by the previous layer is upsampled, and the initial watermark image features of the corresponding resolution are introduced through skip connections. Then, it is fused with the mid-frequency texture features and high-frequency texture features of the style vector to gradually add texture details and finally output a high-resolution fused texture map.

[0026] Optionally, the fusion operation in the optimization and enhancement layer adopts adaptive instance normalization technology, and the parameters of the adaptive instance normalization are dynamically generated by the style vector corresponding to the current layer.

[0027] Optionally, the saliency-aware structure-preserving loss in the composite loss function is calculated as follows:

[0028] The features of the fused texture map and the initial watermark image are extracted using feature extraction operators, respectively.

[0029] Calculate the feature differences between the fused texture map and the initial watermark image;

[0030] The feature differences are weighted using a significance weighting plot, and the L1 norm of the weighted differences is calculated.

[0031] Optionally, based on the fused texture map, the original information is recovered through decoding processing, including:

[0032] Preprocessing is performed on the fused texture map using geometric correction or illumination compensation;

[0033] Extract watermark structure information from the preprocessed image using a decoder or image processing algorithm;

[0034] The extracted watermark structure information is subjected to two-dimensional inverse mapping and forward error correction decoding to recover the original binary information stream.

[0035] Optionally, the authenticity of the fused texture map is verified based on the encrypted seed, including:

[0036] Using the encrypted seed, original information, and target style, a reference fused texture map is regenerated;

[0037] Calculate the similarity between the fused texture map to be verified and the reference fused texture map;

[0038] If the similarity is higher than a preset threshold, the fused texture map to be verified is determined to be genuine.

[0039] This invention also proposes a verifiable digital watermark texture generation system based on hierarchical fusion and saliency awareness, for performing the method, including:

[0040] The preprocessing and saliency analysis module is used to generate an initial watermark image based on the original binary information stream through forward error correction coding and two-dimensional mapping, and to perform saliency analysis on the initial watermark image.

[0041] The style vector construction module is used to extract features from a target style image using a style encoder and parameterize them to construct a style vector.

[0042] The hierarchical fusion generation module is used to generate a fused texture map by performing a step-by-step fusion from macro to micro based on the initial watermark image, the style vector and the encryption seed through a fusion generator with a multi-scale hierarchical structure. The fusion generator is trained and optimized using a composite loss function including adversarial loss, style consistency loss and saliency-aware structure preservation loss.

[0043] The decoding and verification module is used to recover the original information based on the fused texture map through decoding processing, and to verify the authenticity of the fused texture map based on the encryption seed.

[0044] Compared with the prior art, the present invention has the following advantages and technical effects:

[0045] This invention abandons the traditional "embedded-hidden" watermarking paradigm and proposes a novel "generation-fusion" mechanism. By incorporating watermark information as an intrinsic structural element of the image generation process, the watermark information and artistic textures are organically integrated at the source of generation. The resulting texture image not only carries watermark information but also possesses independent artistic value, which can be directly applied to high-end visual products, achieving a perfect unity of information security and aesthetic expression.

[0046] This invention introduces a saliency-aware structure-preserving loss function during the training phase. By intelligently identifying key structural features in the watermarked image, it applies stronger structure-preserving constraints to these regions. This technique allows the system to maintain artistic freedom while selectively protecting watermark features crucial for decoding, thus maintaining high decoding robustness even in complex application environments, while ensuring the visual appeal of the generated image.

[0047] This invention achieves refined control over artistic styles by parameterizing artistic style features into low-dimensional style vectors. Designers can adjust the values ​​of these style vectors to continuously control texture styles and blend multiple styles. This technological breakthrough transforms artistic style from an abstract, subjective concept into quantifiable and adjustable technical parameters, greatly expanding the dimensions and possibilities of artistic creation.

[0048] This invention introduces a cryptographic seed during the generation process, injecting a unique random feature into each generated texture image. This technique ensures that even if an attacker obtains all the parameters of the generated model, they cannot forge a verifiable, legitimate texture. This mechanism provides the generated texture images with near-cryptographic anti-counterfeiting capabilities, making them suitable for applications with high security requirements such as identity verification and product authentication.

[0049] This invention employs a hierarchical generation architecture comprising a basic generation layer and multiple optimization layers, achieving progressively refined generation from macroscopic structure to microscopic details. This technique effectively overcomes the limitations of traditional single-stage generation methods in handling complex textures, ensuring both the global structural stability of the generated image and the richness of local details, significantly improving the consistency and reliability of output quality. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0051] Figure 1 This is a schematic diagram of the overall process architecture of an embodiment of the present invention;

[0052] Figure 2 This is a flowchart illustrating the training process of the fusion generator model according to an embodiment of the present invention.

[0053] Figure 3 This is a diagram illustrating the watermark generation effect of an embodiment of the present invention. Detailed Implementation

[0054] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0056] Example 1

[0057] This embodiment proposes a verifiable digital watermark texture generation method based on hierarchical fusion and saliency awareness. The method employs a hierarchical fusion structure from macro to micro, ensuring the integrity of the macroscopic watermark structure while maintaining the richness of microscopic details. It also introduces an image saliency analysis strategy to identify and protect watermark structural features that play a crucial role in decoding, thus achieving a balance between decoding performance and visual aesthetics. Furthermore, an encryption seed generation step is introduced during watermark generation, assigning a unique and verifiable identity to each generated watermark texture, aiming to address the technical challenge of insufficient security in existing digital watermarking technologies.

[0058] The method specifically includes the following steps:

[0059] Based on the original binary information stream, an initial watermark image is generated through forward error correction coding and two-dimensional mapping, and the initial watermark image is subjected to saliency analysis.

[0060] Based on the target style image, its features are extracted by a style encoder and parameterized to construct a style vector;

[0061] Based on the initial watermark image, the style vector, and the encryption seed, a fusion generator with a multi-scale hierarchical structure is used to perform a step-by-step fusion from macro to micro to generate a fused texture map. The fusion generator is trained and optimized using a composite loss function that includes adversarial loss, style consistency loss, and saliency-aware structure preservation loss.

[0062] Based on the fused texture map, the original information is recovered through decoding, and the authenticity of the fused texture map is verified based on the encrypted seed.

[0063] A feasible method for generating an initial watermark image based on the original binary information stream through forward error correction coding and two-dimensional mapping includes:

[0064] The original binary information stream is subjected to forward error correction coding; the encoded data stream is then converted into an initial watermark image with structural features using a two-dimensional mapping algorithm.

[0065] An feasible process for performing saliency analysis on the initial watermark image includes:

[0066] Key structural features in the initial watermark image are identified using a corner detection algorithm or an edge detection algorithm; a saliency weight map with the same size as the initial watermark image is generated based on the identification results, wherein the regions corresponding to the key structural features are assigned higher weights.

[0067] Specifically, the process of watermark information preprocessing and saliency analysis includes:

[0068] After the raw binary information stream is input into the system, it first uses forward error correction coding (e.g., Reed-Solomon codes) to add redundant information, thereby effectively improving the information's noise resistance. Next, the encoded data stream is mapped into an initial watermark image (IWI) with structural features using a two-dimensional mapping algorithm (e.g., Hadamard transform or fractal construction). This embodiment introduces watermark saliency analysis in this step, generating a saliency weight map (SWM) of the same size as the IWI. The pixel values ​​in the SWM range from 0 to 1, representing the importance of each pixel position during watermark decoding. For example, Harris corner detection, Sobel edge detection, or a lightweight CNN model can be used to identify key feature points in the watermark and assign them higher weights in the SWM.

[0069] Implementable methods for constructing style vectors from target style images by extracting and parameterizing their features using a style encoder, including:

[0070] A style encoder is used to extract multi-scale style feature maps from the target style image; global statistics or pooling operations are performed on the multi-scale style feature maps to obtain low-dimensional style vectors.

[0071] Specifically, the process of style feature parameter extraction includes:

[0072] In this embodiment, one or more images can be selected as target style images. Style encoder networks, such as partially convolutional layers of a VGG network or a specially trained encoder, are used to extract style features from these images. Unlike existing technologies, this embodiment performs subsequent processing on the extracted feature maps, transforming them into one or a set of low-dimensional style vectors (SVs) for more refined style control in later stages. Using this mechanism, style information is specifically parameterized, and each component of the style vector (SV) corresponds to different style attributes, such as texture coarseness, directionality, or color distribution. By weighted combination of multiple style vectors, a smoothly transitioning hybrid style with new features can be formed within the style space.

[0073] Implementable, based on the initial watermark image, the style vector, and the encryption seed, a fusion generator with a multi-scale hierarchical structure performs a step-by-step fusion from macro to micro to generate a fused texture map, including:

[0074] The fusion generator includes a basic generation layer and several optimization and enhancement layers. In the basic generation layer, a low-resolution feature map is generated based on the initial watermark image, the low-frequency texture features of the style vector, and the encryption seed. In at least one optimization and enhancement layer, the feature map output by the previous layer is upsampled, and the initial watermark image features of the corresponding resolution are introduced through skip connections. Then, it is fused with the mid-frequency texture features and high-frequency texture features of the style vector to gradually add texture details and finally output a high-resolution fused texture map.

[0075] Furthermore, the fusion operation in the optimization and enhancement layer adopts adaptive instance normalization technology, and the parameters of the adaptive instance normalization are dynamically generated by the style vector corresponding to the current layer.

[0076] Specifically, the process of generating hierarchical textures based on encrypted seeds includes:

[0077] This embodiment uses a hierarchical fusion generator (HFG) to generate the final fused texture map (FTI). The HFG receives the initial watermark image IWI, style vector SV, and encryption seed CS. Internally, it contains multiple scale layers, including a basic generation layer and several optimization layers. The basic generation layer is located in the bottom layer of the HFG; therefore, the input to the HFG is the input to this layer, but the style vector (SV) within it is a subset reflecting global features. The task of the basic generation layer is to generate a low-resolution feature map, providing a basis for the global topology of the fused texture map (FTI). The encryption seed CS is generated by deterministically controlling the convolution kernel or random noise to achieve a unique mapping of the output, thus ensuring the verifiability of the watermark. The optimization layers are responsible for upsampling the feature map from the previous layer, obtaining the IWI features of any layer through an efficient skip connection mechanism, and fusing them with other subsets of SV (representing mid- and high-frequency texture features). The fusion operation can be implemented using techniques such as adaptive instance normalization (AdaIN), with its normalization parameters generated in real-time by SV. By progressively refining the image, the system continuously adds textures and details to the skeleton of the low-resolution base image, eventually obtaining a full-resolution FTI.

[0078] The feasible, saliency-aware structure-preserving loss in the composite loss function is calculated as follows:

[0079] The features of the fused texture image and the initial watermark image are extracted using feature extraction operators respectively; the feature differences between the fused texture image and the initial watermark image are calculated; the feature differences are weighted using a saliency weight map, and the L1 norm of the weighted differences is calculated.

[0080] Specifically, the adversarial training process based on a composite loss function includes:

[0081] To train the HFG network, this embodiment designs a refined composite loss function to achieve a balance among multiple optimization objectives. This loss function includes at least the following components:

[0082] Adversarial loss: One or more discriminators with multi-scale structures are used to distinguish the fused texture map (FTI) generated by HFG from the real image. Adversarial loss makes the FTI output by the generator more visually realistic and has a natural artistic style.

[0083] Style consistency loss: This loss ensures that the generated image maintains a high degree of consistency with the target image at the style level by calculating the difference between the statistical indices of the generated FTI and the target style image in several intermediate layer features of the style encoder. This includes consistency in texture structure and color distribution.

[0084] Saliency-aware structure-preserving loss: Its design aims to suppress the destruction of the initial watermark image (IWI) structural information by HFG during training. Its calculation form is as follows:

[0085] ;

[0086] in, This represents a feature extraction operator, such as the Sobel gradient operator or a specific convolutional layer; This indicates an element-wise multiplication operation. Through the saliency weighting mechanism of SWM, the model is more constrained in salient regions (such as corners and edges), while retaining more freedom of style transformation in low saliency regions, thus achieving a balance between structural fidelity and artistic style.

[0087] Implementable methods include recovering the original information through decoding based on the fused texture map, including:

[0088] The fused texture map is preprocessed with geometric correction or illumination compensation; watermark structure information is extracted from the preprocessed image using a decoder or image processing algorithm; the extracted watermark structure information is then subjected to two-dimensional inverse mapping and forward error correction decoding to recover the original binary information stream.

[0089] Furthermore, the authenticity of the fused texture map is verified based on the encrypted seed, including:

[0090] Using the encrypted seed, original information, and target style, a reference fused texture map is regenerated; the similarity between the fused texture map to be verified and the reference fused texture map is calculated; if the similarity is higher than a preset threshold, the fused texture map to be verified is determined to be genuine.

[0091] Specifically, the decoding and verification process includes:

[0092] Decoding process: After obtaining the FTI from a physical medium (such as a printed carrier), the image is first corrected and preprocessed to compensate for possible geometric distortions or lighting deviations. Next, the decoder network or traditional image processing algorithm performs the reverse steps of the encoding stage to extract the watermark structure information from the FTI. Through forward error correction decoding, the complete original data content can be recovered.

[0093] Verification process: When using an encrypted seed, the system verifies the source of the FTI. It obtains the FTI to be tested and the corresponding seed (or public key), and determines whether the graph was generated by a legitimate HFG using that seed by replaying the generation steps or calling a trained verification network. Finally, it returns "true" or "false".

[0094] On the other hand, this embodiment also proposes a verifiable digital watermark texture generation system based on hierarchical fusion and saliency awareness, for performing the method, including:

[0095] The preprocessing and saliency analysis module is used to generate an initial watermark image based on the original binary information stream through forward error correction coding and two-dimensional mapping, and to perform saliency analysis on the initial watermark image.

[0096] The style vector construction module is used to extract features from a target style image using a style encoder and parameterize them to construct a style vector.

[0097] The hierarchical fusion generation module is used to generate a fused texture map by performing a step-by-step fusion from macro to micro based on the initial watermark image, the style vector and the encryption seed through a fusion generator with a multi-scale hierarchical structure. The fusion generator is trained and optimized using a composite loss function including adversarial loss, style consistency loss and saliency-aware structure preservation loss.

[0098] The decoding and verification module is used to recover the original information based on the fused texture map through decoding processing, and to verify the authenticity of the fused texture map based on the encryption seed.

[0099] Compared with the prior art, the method proposed in this embodiment has the following beneficial effects:

[0100] This embodiment achieves a deep integration of art and technology: It abandons the traditional "embedded-hidden" digital watermarking paradigm and proposes a novel "generation-fusion" mechanism, making the watermark information an intrinsic structure within the image generation process, rather than a later-added steganographic element. The resulting texture map itself possesses artistic design value and can be directly applied to high-end visual products, achieving an organic unity of information security and aesthetic expression.

[0101] Synergistic optimization of decoding robustness and visual aesthetics: This embodiment introduces a saliency-aware structure-preserving loss function during the training phase, enabling the system to automatically identify and enhance key areas of watermark information for targeted protection. This design ensures high robustness in complex environments while retaining the creative flexibility of style generation networks, thereby generating textured images that are both practical and visually appealing.

[0102] Controllability and Creative Breakthroughs in Style: With the help of parametric style technology, artistic style is no longer an abstract, subjective concept, but a quantifiable and adjustable parameter system. Designers can control the intensity of texture styles like adjusting volume, and can also mix multiple styles to create unprecedented visual effects, achieving true artistic innovation.

[0103] Significantly enhanced anti-counterfeiting security: By injecting a unique random feature into each generated texture image using an encrypted seed, it possesses an uncopyable "digital gene." Even if an attacker obtains the generation parameters and model weights, they cannot forge or clone a verifiable legitimate texture. Therefore, this scheme possesses near-cryptographic level anti-counterfeiting capabilities, making it suitable for fields requiring high fidelity and high security, such as identity verification and product authentication.

[0104] Significant improvements in generation quality and stability: Employing a hierarchical, multi-scale generation architecture allows for precise control over both the overall image and its details at both macro and micro levels. Compared to traditional single-stage generation methods, it effectively ensures stable output quality and rich watermark details.

[0105] This embodiment provides an efficient, flexible, and secure solution to resolve the contradiction between the concealment and robustness of traditional watermarks, and has broad application prospects in the fields of copyright protection, anti-counterfeiting of high-end products, and commodity packaging.

[0106] Example 2

[0107] This embodiment provides a method for generating verifiable digital watermark texture maps based on hierarchical fusion and saliency awareness. For example... Figure 1 The diagram shown illustrates the overall process architecture of the method described in this embodiment, which mainly includes the following steps:

[0108] (1) Watermark information preprocessing and saliency analysis:

[0109] This step aims to transform the raw binary information stream into a structured, error-correcting initial watermark image (IWI) and generate a saliency weight map (SWM) to guide the subsequent generation process.

[0110] First, the system receives the raw binary information to be embedded, such as a copyright text or product ID. To enhance the information's resistance to interference during subsequent printing, scanning, and other processes, forward error correction coding is used to encode the binary information stream, and redundant check bits are added.

[0111] Next, the FEC-encoded binary data stream is converted into a two-dimensional initial watermark image using a deterministic two-dimensional mapping algorithm. In one embodiment, the Hadamard transform can be used. For example, using the rows or columns of the Hadamard matrix as a basis, the binary sequence is mapped into a grayscale image with obvious block structure and orthogonality. Such a structured image is easier to extract stably during decoding compared to random noise points.

[0112] Subsequently, this embodiment introduces a core saliency analysis module to analyze the generated IWI to identify structural features that play a crucial role in decoding. The output of this module is a saliency weight map (SWM) of the same size as the IWI, with pixel values ​​ranging from 0 to 1. In one embodiment, classic image processing operators can be used, such as the Harris corner detection algorithm to identify corners in the image, and the Sobel or Canny edge detection algorithm to identify edges. Detected corners and edge regions are assigned higher weights (e.g., close to 1.0) in the SWM, while flat, informationally redundant regions are assigned lower weights (e.g., close to 0). In another embodiment, a pre-trained lightweight convolutional neural network can also be used to predict the importance of each pixel in the IWI end-to-end to obtain a more refined SWM.

[0113] (2) Parametric extraction of style features:

[0114] This step aims to extract the artistic style from one or more target style images and represent it as a controllable low-dimensional parameter.

[0115] The user provides one or more images in the desired artistic style as input. The system employs a pre-trained deep convolutional neural network as a style encoder, such as the VGG-19 network. The style images are input into this network, and activation feature maps from its multiple intermediate convolutional layers are extracted. These feature maps at different depths capture information such as texture, brushstrokes, and color distribution of the style images at different scales.

[0116] To achieve quantitative control and flexible blending of styles, this embodiment parametrically processes the extracted high-dimensional feature maps. Specifically, by applying global average pooling or calculating their statistical properties (such as mean and variance) to the feature maps of each selected layer, they are reduced to one or a set of low-dimensional style vectors (SVs). For example, the SV components corresponding to shallow network features may control color and basic brushstrokes, while the SV components corresponding to deep network features control more complex texture structures. By weighted interpolation of SVs from images of different styles, entirely new hybrid art styles that fall between the two can be created.

[0117] (3) Hierarchical texture generation based on encrypted seeds:

[0118] This is the core step in generating the final blended texture map (FTI), which is accomplished by a hierarchical blending generator (HFG). The HFG takes IWI, SV, and a cryptographic seed (CS) as input.

[0119] like Figure 2 As shown, HFG has a multi-scale hierarchical structure, including a basic generation layer and multiple optimization and enhancement layers.

[0120] Basic Generation Layer: This layer receives a downsampled IWI, a subset of SV representing global style features, and a cryptographic seed CS. Its task is to generate a low-resolution feature map that lays the foundation for the macroscopic structure and overall tone of the final texture map FTI. The cryptographic seed CS here deterministically influences the initialization of the convolutional kernels or generates a random noise map concatenated with the input, ensuring that even if all other inputs are identical, different CSs will generate unique texture maps, providing a basis for subsequent validation.

[0121] Optimization and Enhancement Layers: Each optimization and enhancement layer receives the feature map output from the previous layer and performs upsampling operations (such as transposed convolution or pixel reassembly). Simultaneously, through skip connections, this layer acquires IWI features that match the current resolution. Then, the upsampled feature map, the IWI feature map, and the SV subset corresponding to the current level of detail are fused. In a preferred embodiment, the fusion operation employs adaptive instance normalization. The scaling and translation parameters of AdaIN are dynamically generated from the SV subset of the current layer through a small fully connected network, thereby accurately injecting style information into the content structure. This process proceeds layer by layer, from coarse to fine, continuously superimposing and optimizing micro-texture details on the macroscopic structure, ultimately outputting a full-resolution fused texture map (FTI).

[0122] (4) Adversarial training based on composite loss function:

[0123] To train the HFG network, this embodiment designs a composite loss function to collaboratively optimize the artistic appeal, style consistency, watermark robustness, and verifiability of the generated images. For example... Figure 2 As shown, the training process is a competitive game.

[0124] Total loss function It can be represented as:

[0125] ;

[0126] in, These are the weighting coefficients for each type of loss.

[0127] Adversarial loss: One or more discriminator networks with multi-scale receptive fields are used to distinguish between FTI generated by HFG and real artistic texture images. This loss term drives HFG to generate natural textures that are visually indistinguishable from real artworks.

[0128] Style Consistency Loss: Intermediate feature maps of the FTI and target style images are extracted separately using a style encoder (such as a VGG network). The differences (such as L1 or L2 distance) between the Gram matrices or channel statistics (mean, variance) of these feature maps are calculated. This loss term ensures that the FTI is highly consistent with the target style in terms of texture, color, etc.

[0129] Saliency-aware structure preservation loss: It is used to ensure that key information in IWI is not destroyed during stylization. Its calculation formula is:

[0130] ;

[0131] in, It is a feature extraction operator (such as the Sobel gradient operator or a fixed convolutional layer). This indicates element-wise multiplication. This represents the L1 norm. The core of this loss function lies in the weighting effect of SWM: in regions with high SWM weights (such as the edges and corners of the watermark), the structural differences between FTI and IWI are severely penalized, forcing the network to accurately preserve these features; while in regions with low SWM weights (flat regions), the penalty is smaller, giving the network greater freedom for artistic creation. This achieves a delicate balance between decoding robustness and visual aesthetics.

[0132] (5) Decoding and verification:

[0133] Decoding Process: When recovering original information from an FTI (e.g., an image scanned from printed material), image preprocessing is first performed, including geometric correction (such as affine transformation alignment), illumination compensation, and noise reduction. Then, the processed image is input into a decoder network (whose structure can be viewed as the inverse of a generator) or traditional image processing algorithms are used to extract the embedded IWI structure. Finally, the extracted IWI is subjected to a two-dimensional inverse mapping to obtain a binary sequence, and any errors are corrected by an FEC decoder to recover the original binary information, such as... Figure 3 As shown.

[0134] Verification Process: To verify the authenticity of an FTI, the system utilizes its unique "digital gene"—the encrypted seed CS. In one embodiment, the verification module receives the FTI to be tested, the claimed CS, and the original information and style used to generate the FTI (for reconstructing IWI and SV). The verification module uses these inputs to regenerate a reference FTI using a pre-trained HFG. By comparing the similarity between the FTI to be tested and the reference FTI (e.g., calculating pixel-level differences or structural similarity SSIM), if the similarity is higher than a preset threshold, the FTI to be tested is determined to be "genuine"; otherwise, it is "fake." This method makes it difficult for forgers to counterfeit even if they possess the network model, as they cannot obtain the correct CS.

[0135] In summary, this embodiment ensures image quality through a hierarchical generation architecture, achieves synergistic optimization of robustness and aesthetics through saliency analysis, and provides cryptographic-level security and anti-counterfeiting features through an encrypted seed mechanism, providing a complete and innovative solution for the deployment of digital watermarking technology in high-end application fields.

[0136] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A verifiable digital watermark texture generation method based on hierarchical fusion and saliency awareness, characterized in that, Includes the following steps: Based on the original binary information stream, an initial watermark image is generated through forward error correction coding and two-dimensional mapping, and the initial watermark image is subjected to saliency analysis. Based on the target style image, its features are extracted by a style encoder and parameterized to construct a style vector; Based on the initial watermark image, the style vector, and the encryption seed, a fusion generator with a multi-scale hierarchical structure is used to perform a step-by-step fusion from macro to micro to generate a fused texture map. The fusion generator is trained and optimized using a composite loss function that includes adversarial loss, style consistency loss, and saliency-aware structure preservation loss. Based on the fused texture map, the original information is recovered through decoding, and the authenticity of the fused texture map is verified based on the encrypted seed.

2. The method according to claim 1, characterized in that, Based on the original binary information stream, an initial watermark image is generated through forward error correction coding and two-dimensional mapping, including: Perform forward error correction coding on the original binary information stream; The encoded data stream is converted into an initial watermark image with structural features using a two-dimensional mapping algorithm.

3. The method according to claim 1, characterized in that, The process of performing saliency analysis on the initial watermark image includes: Key structural features in the initial watermark image are identified using corner detection algorithms or edge detection algorithms. A saliency weight map with the same size as the initial watermark image is generated based on the recognition results, wherein the regions corresponding to the key structural features are assigned higher weights.

4. The method according to claim 1, characterized in that, Based on the target style image, features are extracted using a style encoder and parameterized to construct a style vector, including: A style encoder is used to extract multi-scale style feature maps from the target style image; Global statistics or pooling operations are performed on the multi-scale style feature maps to obtain low-dimensional style vectors.

5. The method according to claim 1, characterized in that, Based on the initial watermark image, the style vector, and the encryption seed, a fusion generator with a multi-scale hierarchical structure performs a step-by-step fusion from macro to micro to generate a fused texture map, including: The fusion generator includes a basic generation layer and several optimization and enhancement layers; In the basic generation layer, a low-resolution feature map is generated based on the initial watermark image, the low-frequency texture features of the style vector, and the encryption seed. In at least one optimization layer, the feature map output by the previous layer is upsampled, and the initial watermark image features of the corresponding resolution are introduced through skip connections. Then, it is fused with the mid-frequency texture features and high-frequency texture features of the style vector to gradually add texture details and finally output a high-resolution fused texture map.

6. The method according to claim 5, characterized in that, The fusion operation in the optimization and enhancement layer adopts adaptive instance normalization technology, and the parameters of the adaptive instance normalization are dynamically generated by the style vector corresponding to the current layer.

7. The method according to claim 3, characterized in that, The saliency-aware structure-preserving loss in the composite loss function is calculated as follows: The features of the fused texture map and the initial watermark image are extracted using feature extraction operators, respectively. Calculate the feature differences between the fused texture map and the initial watermark image; The feature differences are weighted using a significance weighting plot, and the L1 norm of the weighted differences is calculated.

8. The method according to claim 1, characterized in that, Based on the fused texture map, the original information is recovered through decoding processing, including: Preprocessing is performed on the fused texture map using geometric correction or illumination compensation; Extract watermark structure information from the preprocessed image using a decoder or image processing algorithm; The extracted watermark structure information is subjected to two-dimensional inverse mapping and forward error correction decoding to recover the original binary information stream.

9. The method according to claim 1, characterized in that, The authenticity of the fused texture map is verified based on the encrypted seed, including: Using the encrypted seed, original information, and target style, a reference fused texture map is regenerated; Calculate the similarity between the fused texture map to be verified and the reference fused texture map; If the similarity is higher than a preset threshold, the fused texture map to be verified is determined to be genuine.

10. A verifiable digital watermark texture generation system based on hierarchical fusion and saliency awareness, characterized in that, For performing the method as described in any one of claims 1 to 9, comprising: The preprocessing and saliency analysis module is used to generate an initial watermark image based on the original binary information stream through forward error correction coding and two-dimensional mapping, and to perform saliency analysis on the initial watermark image. The style vector construction module is used to extract features from a target style image using a style encoder and parameterize them to construct a style vector. The hierarchical fusion generation module is used to generate a fused texture map by performing a step-by-step fusion from macro to micro based on the initial watermark image, the style vector and the encryption seed through a fusion generator with a multi-scale hierarchical structure. The fusion generator is trained and optimized using a composite loss function including adversarial loss, style consistency loss and saliency-aware structure preservation loss. The decoding and verification module is used to recover the original information based on the fused texture map through decoding processing, and to verify the authenticity of the fused texture map based on the encryption seed.

Images