Visual enhancement robust digital dark watermarking method based on deep learning, storage medium and equipment

Abstract

The invention discloses a vision enhancement robust digital dark watermarking method based on deep learning, a storage medium and equipment, and belongs to the technical field of digital image processing and information security. According to the method, a visual enhancement module comprising a self-adaptive watermark region adjustment module and a frequency enhancement module is constructed, and a multi-region local loss training mechanism and a screen shooting simulation module are combined, so that the visual quality of a watermark image is improved, and meanwhile, the robustness of the watermark image for resisting physical attacks such as screen shooting is enhanced. The specific processing flow comprises the following steps: acquiring an original image and watermark information; embedding the watermark information into the image by using an encoder, wherein an embedded region is optimized by a self-adaptive region guide matrix generated based on the image edge and gray information; in the model training process, a noise layer simulating screen shooting physical distortion is introduced, and the area weight is dynamically adjusted according to residual error distribution in the later stage of training so as to focus and optimize the image quality; watermark information is extracted from an image which may suffer from an attack by using a decoder.

Description

Technical Field

[0001] This invention relates to the fields of digital image processing and information security technology, specifically to a deep learning-based visual enhancement robust digital dark watermarking method, storage medium, and device. Background Technology

[0002] With the widespread dissemination of digital media and the explosive growth of digital content, digital copyright protection technology has become crucial. Digital watermarking technology, as an effective means of imperceptibly embedding identification information into carrier data, plays a key role in copyright authentication, content traceability, and anti-counterfeiting and anti-tampering, and has received widespread attention from academia and industry.

[0003] In recent years, deep learning-based digital watermarking algorithms have demonstrated superior robustness and image quality compared to traditional methods, thanks to their powerful feature extraction and end-to-end optimization capabilities.

[0004] However, in practical applications, especially in scenarios requiring resistance to physical layer attacks such as screen capture, existing mainstream algorithms still face significant challenges in the coordinated optimization of visual quality and robustness, exhibiting significant technical problems, mainly in the following two aspects:

[0005] First, existing solutions, while improving robustness, often severely sacrifice visual quality, resulting in noticeable artifacts and color aberrations. These artifacts significantly impact human visual perception, preventing current watermarking algorithms from being widely applied in practical production despite good experimental performance. For example, the early pioneering end-to-end deep learning watermarking framework HiDDeN introduced an encoder-decoder structure and a noise layer for joint training, but the generated watermarked images were of poor quality and lacked robustness. Subsequent research, StegaStamp, achieved a significant breakthrough in resisting screen capture attacks by introducing a complex differentiable noise layer simulating a physical process. However, to meet extremely high robustness requirements, the model was forced to make substantial modifications to image pixels, resulting in watermarked images generally exhibiting visible "ripple artifacts" and local color distortion, severely affecting visual experience and practical value. Another study optimizing for screen capture attacks, PIMoG, proposed that simply simulating key distortions in the screen capture process (such as moiré patterns) as a noise layer can effectively improve robustness. Although the solution improves image quality metrics, it still cannot completely eliminate artifacts and is prone to color differences in smooth areas, leaving considerable room for improvement in visual fidelity.

[0006] Second, existing methods mostly optimize at the global or pixel level, lacking targeted modeling of human visual characteristics, leading to a mismatch between embedding efficiency and visual perception. Many methods attempt to improve quality by adjusting the loss function, such as using mean squared error (MSE) loss in the pixel domain or introducing adversarial training to enhance concealment. However, these methods typically perform global uniform processing of the image, failing to consider the differences in human eye sensitivity to different regions of the image, and overemphasizing the optimization of detailed pixels while neglecting the global visual effect. For example, in bright, flat-textured areas (such as the sky), small pixel changes are easily noticed; while in complex-textured, low-brightness areas, the human eye's masking effect is stronger, accommodating more modifications without attracting attention. Although some studies have attempted to adaptively adjust the watermark embedding strength based on psychovisual characteristics, their adjustments are still based on the pixel level, failing to guide from the strategic level of region selection or combine with frequency domain characteristics, thus having limited effect on improving overall visual quality.

[0007] Furthermore, current training strategies typically employ globally consistent optimization objectives. However, in the later stages of training, the residuals caused by watermark embedding often become non-uniformly concentrated in certain local areas of the image. Continuous globally uniform optimization not only fails to efficiently repair these "problem areas" but may also introduce unnecessary perturbations into other converged regions, thus limiting the ultimate upper limit of image quality.

[0008] In summary, there is an urgent need for a digital watermarking technology that can deeply integrate the characteristics of human visual perception, intelligently embed watermarks in both the frequency and spatial domains, and achieve fine optimization through dynamic training strategies. The goal of this method is to generate visually artifact-free, color-difference-free watermarked images that are highly consistent with the original image, while ensuring excellent robustness against strong physical attacks such as screen capture. This will truly meet the practical application needs of high-standard digital rights protection and covert communication. Summary of the Invention

[0009] This invention aims to address the aforementioned problems in the existing technology by providing a deep learning-based visual enhancement robust digital dark watermarking method, storage medium, and device. This method, through the introduction of a visual enhancement module, a multi-region local loss training mechanism, and a noise layer simulating physical attacks, significantly improves the visual quality of the watermarked image while maintaining high robustness. It mainly includes the following core modules and steps:

[0010] Visual Enhancement Module: This module includes an adaptive watermarking region adjustment module and a frequency enhancement module. The adaptive region adjustment module analyzes the edges (using a Canny detector) and grayscale information of the original image to generate a region guidance matrix, guiding the encoder to preferentially embed the watermark information into low-grayscale, high-texture regions that are less sensitive to the human eye, thereby reducing visual interference. The frequency enhancement module introduces an improved focusing frequency domain loss (…). ), optimize the differences in the image in the frequency domain, focusing on reducing the "hard-to-synthesize frequency" components that cause ripple artifacts.

[0011] Multi-region local loss training mechanism: The training process is divided into two stages, an early stage and a late stage, and different loss calculation strategies are adopted. In the early stage of training, the model focuses on establishing a stable embedding and extraction foundation. Therefore, a globally uniform loss calculation method is adopted to calculate uniform frequency domain loss and pixel mean square error (MSE) loss for the entire image to ensure effective embedding of watermark information and initial robustness.

[0012] As training progresses, in the later stages, the residuals caused by watermark embedding are no longer uniformly distributed but significantly concentrated in certain local regions of the image. At this point, uniform global optimization is no longer used. Specifically, the algorithm divides the image into multiple regions based on the image resolution and analyzes the residual distribution within each region in real time. Training weights are dynamically allocated based on the residual size: regions with dense residuals are assigned higher loss weights, allowing the model to focus on optimizing these "problem regions" during backpropagation; regions with smaller residuals or those that have already converged are weighted less to avoid unnecessary interference. This mechanism enables the model's optimization attention to adaptively focus on the local areas that most need improvement, thereby significantly improving the visual fidelity and detail consistency of the watermarked image while maintaining overall image quality.

[0013] Physical robustness enhancement: A sophisticated noise layer is introduced between the encoder and decoder, specifically mimicking the main types of distortion generated during screen capture, including perspective distortion, lighting variations, moiré fringes, and Gaussian noise. By exposing the model to this noise during training, it learns to embed watermarks that resist such physical attacks.

[0014] A deep learning-based visual enhancement robust digital dark watermarking method includes the following steps:

[0015] Step 1: Collect the original image and watermark information;

[0016] Step 2: Input the original image and the watermark information into the encoder to generate a watermarked image. The watermark embedding process of the encoder is guided by the visual enhancement module, which includes an adaptive region adjustment module for determining the priority embedding region based on image edges and grayscale information.

[0017] Step 3: Set a noise layer and an adversarial discriminator between the encoder and the decoder. The noise layer is used to apply noise perturbation to the watermarked image to simulate the physical distortion of screen shooting, so as to obtain a noisy image. The adversarial discriminator and the encoder form an adversarial joint training framework.

[0018] Step 4: Input the noisy image into the decoder to extract the watermark information;

[0019] Step 5: During the training process, a multi-region local loss function is used for optimization. In the later stage of training (i.e. when the watermark residual tends to concentrate in the local area of ​​the image), the multi-region local loss function dynamically allocates training weights according to the residual size of each region of the image.

[0020] During training, neural networks continuously adjust their parameters through backpropagation of errors to minimize the loss function. Building upon this, this invention introduces an adaptive region adjustment module to generate a region guidance matrix. This allows for precise spatial-aware control of the network training process. Region guidance matrix. The size is consistent with the resolution of the original image. Each element in the matrix directly corresponds to a pixel position in the original image. The matrix is ​​combined with pixel-wise loss through matrix dot product to achieve regional weighted control of the overall loss value.

[0021] Specifically, the regional guidance matrix The design follows the characteristics of human vision: for low grayscale and high texture areas in an image, Assigning lower weights to these regions reduces their contribution to the total loss, effectively reducing the loss penalty the model incurs when embedding watermarks in these areas, thus encouraging the encoder to embed more information in these regions; conversely, for smooth regions with high grayscale values ​​and low texture, These regions are then assigned higher weights to emphasize visual fidelity, thereby suppressing the embedding intensity of the watermark in easily perceptible areas. Through this dynamic weight allocation mechanism, the training process is effectively guided, causing the watermark information to adaptively and preferentially embed into visually insensitive low-grayscale, high-texture regions. The specific steps are as follows:

[0022] Canny edge detection is performed on the original image to obtain an edge information matrix. The boundary region matrix value is set to 0, and the non-boundary region matrix value is set to 1; the original image is then converted to grayscale to obtain a grayscale information matrix. And perform normalization; according to the formula Calculate the region guidance matrix , where α and β are weighting coefficients.

[0023] The visual enhancement module also includes a frequency enhancement module for calculating the focusing frequency domain loss. The expression is: ;

[0024] in, This is a frequency optimization matrix used to weight different frequency components; This is a distance metric between the original image and the watermarked image in the frequency domain.

[0025] The encoder is built on an improved U-Net network architecture, whose core comprises a downsampling path, a global feature extraction layer, and an upsampling path. These three components work together to achieve high-fidelity fusion of watermark information and deep image features. The downsampling path progressively extracts multi-scale features and compresses spatial dimensions through operations including convolution and max pooling, constructing a feature pyramid from low-level texture to high-level semantics, providing robust feature representation for watermark embedding. The global feature extraction layer, based on downsampling, aggregates the overall statistical information and contextual semantics of the image through global pooling, providing global visual guidance for watermark embedding and ensuring that the embedding process is consistent with the overall image structure. The upsampling path progressively restores the feature map resolution through transposed convolution or interpolation operations, fusing shallow detail features at the same scale through skip connections at each step, while integrating the processed watermark information into the feature stream. This mechanism enables the network to comprehensively utilize local details, high-level semantics, and global context during reconstruction, ultimately mapping the fused features back to the image space through 1×1 convolution, generating a watermarked image with the same size and visual perception as the original image.

[0026] The adversarial discriminator and encoder in step 3 form an adversarial joint training framework, which aims to continuously improve the imperceptibility and visual quality of watermarked images through a game-theoretic mechanism. The parameters are... Adversarial discriminator It is trained as a binary classification network to distinguish between the original image and the watermark image; the parameters are... encoder Used to store the original watermark information Embedded into the original image Watermark image generated later The specific training process is as follows: In each iteration, the adversarial discriminator... Receive raw image and watermarked images As input, and by maximizing the discriminant loss To update parameters To enhance its ability to distinguish between genuine and counterfeit goods, the loss is defined as... ,in This indicates that the discriminator considers the input image to be the original image. The confidence level output.

[0027] At the same time, encoder The goal is to deceive the adversarial discriminator by minimizing the adversarial loss. To update parameters This forces the discriminator to output a confidence level close to 1 for the watermarked image, thereby driving the watermarked image to approximate the distribution of the original image in the feature space. The adversarial loss is defined as: Through this dynamic adversarial optimization process, the encoder is driven to learn how to embed the watermark with as little visual distortion as possible, thereby significantly enhancing the concealment and overall image quality of the watermarked image while maintaining the extractability of the watermark information.

[0028] The calculation method for the multi-region local loss function is as follows:

[0029] The focusing frequency domain loss for each region was calculated both in the early and late stages of training. and pixel mean square error loss ;

[0030] In the early stages of training, the weighted sum of the calculated losses is used directly as the loss function: ,in The weighting coefficient is set to 2.0 in this invention to emphasize the importance of the frequency domain loss function;

[0031] Decoder D extracts information from the watermark image to obtain the decoding result. By calculating the decoding result Compared with the original watermark information The mean square error between them yields the information decoding loss. ,in ;

[0032] Loss With information decoding loss visual enhancement loss and combating losses The total loss in the early stage is obtained by performing a weighted summation.

[0033] In the later stages of training, the image is divided into multiple regions based on different image resolutions;

[0034] Assign weights to each region based on the residual size. The larger the residual, the higher the weight.

[0035] Total local loss ,in and They represent the first Focusing frequency domain loss of each image region and the The MSE loss of each image region is used to simultaneously guide macro-level optimization and pixel-by-pixel optimization of digital watermarking, thereby improving the quality of the watermarked image.

[0036] Total local losses With information decoding loss visual enhancement loss and combating losses The total loss in the later stages is obtained by performing a weighted summation.

[0037] A storage medium storing at least one instruction, at least one program, a code set, or an instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-described deep learning-based visual enhancement robust digital dark watermarking method.

[0038] A server includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the above-described deep learning-based visual enhancement robust digital dark watermarking method.

[0039] The beneficial effects of this invention are:

[0040] (1) Significantly improved visual effect: By adaptively selecting the embedding region and optimizing the frequency domain, the ripple artifacts and color differences of the watermarked image are effectively reduced, and the quality of the watermarked image is greatly improved. This makes the watermarked image visually closer to the original image.

[0041] (2) Enhanced resistance to high-intensity physical attacks: The noise layer makes the algorithm robust against real-world attacks such as screen capture and screenshots, and the decoding accuracy is high.

[0042] (3) Superior overall performance: This invention significantly reduces ripple artifacts and color differences in watermarked images. Experiments on public datasets show that the method of this invention significantly outperforms existing mainstream algorithms such as HiDDeN, StegaStamp, and PIMog in terms of image quality metrics such as peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), while maintaining excellent decoding robustness and high decoding accuracy for operations such as screen capture and screenshots, making it suitable for practical application scenarios such as digital rights protection.

[0043] (4) The present invention actually provides a storage medium and an algorithm program stored on a server and run on a processor, which can more intuitively feel the advantages of the algorithm invention and the effectiveness of solving the problems of the current watermarking algorithm. Attached Figure Description

[0044] Figure 1This is a schematic diagram of the overall framework of the digital watermarking method provided in an embodiment of the present invention.

[0045] Figure 2 This is an example of adaptive watermark embedding region selection in this invention, which shows the difference in the impact of embedding watermarks in different regions of an image on the visual effect on the human eye.

[0046] Figure 3 This is a schematic diagram of the ripple artifact phenomenon commonly seen in existing watermarking algorithms. (a) is the original image, and (b) is the watermarked image with obvious artifacts after embedding the watermark using other open-source algorithms.

[0047] Figure 4 This is a schematic diagram comparing the distribution of residual plots in the early and late stages of the training process of this invention. The same row shows the residual plot itself and the effect of magnifying the residuals tenfold, so as to intuitively understand the distribution pattern of residuals. Among them, (a) is the distribution of residual plots in the early stage of training, and (b) is the distribution of residual plots in the late stage of training. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0049] like Figure 1 As shown, the digital watermarking system based on visual enhancement and deep learning provided in this embodiment mainly includes an encoder, an adversary, a noise layer, a decoder, and a visual enhancement module.

[0050] Dataset and Preprocessing: This embodiment uses the COCO dataset for training and testing. 10,000 images are randomly selected as the training set, and 2,000 images are selected as the test set. All images are uniformly scaled to 128×128 pixels.

[0051] Encoder: The encoder employs a network structure based on an improved U-Net. The input is the original image. and watermark information The network extracts multi-level features through convolution and downsampling, and deeply fuses watermark information with image features through upsampling and skip connections, ultimately outputting a watermarked image. .

[0052] Visual enhancement module:

[0053] Adaptive region adjustment: such as Figure 2 As shown, for the original image Canny edge detection was performed to obtain And perform grayscale conversion and normalization to obtain According to the formula Generate guiding matrix This matrix encourages the network to embed watermarks in dark, textured areas (such as bird feathers) and avoids making obvious modifications in bright, flat areas (such as the sky or the bottle). It also performs an element-wise summation operation based on the matrix element values ​​to obtain the region-guided loss. .in Defined as a region guiding matrix The sum of all elements in the formula. , For the regional guiding matrix Located in row i, the The column element values, where L is the size of the matrix.

[0054] Frequency enhancement: Perform Fourier transform on the original image and the watermarked image to calculate their frequency domain representation. ,calculate Loss, where the frequency optimization matrix This is used to enhance the optimization of mid-to-high frequency components that are key to visual impact, thereby suppressing... Figure 3 The wavy, ripple-like artifacts shown.

[0055] Noise Layer: To simulate screen capture, the noise layer integrates four differentiable perturbations: random perspective transformation, brightness and contrast adjustment, pattern superposition simulating moiré fringes, and Gaussian noise addition. This layer is applied to the watermarked image during training. Generate noisy images This is to improve the robustness of the model under actual physical attacks.

[0056] Decoder: The decoder is a convolutional neural network that receives noisy images. As input, the data undergoes a series of convolution, pooling, and fully connected operations, ultimately outputting the extracted watermark information. The goal is to minimize With the original Errors between them (such as MSE loss).

[0057] Adversarial Discriminant: Introduce an adversarial discriminant. It engages in a game with the encoder. Trying to distinguish the original image and watermarked images The encoder attempts to generate... Indistinguishable This countermeasure helps improve the concealment of the watermark.

[0058] Training strategy and loss function (the loss function directly guides the training direction of the model and plays a crucial role in directly guiding the training approach):

[0059] The training is divided into two phases: the early stage and the later stage.

[0060] Early stage: Focus on establishing basic embedding-extraction capabilities and robustness. The total loss function is shown in the following formula: ,in (Including frequency domain and pixel MSE loss) is a globally fundamental loss calculated uniformly, without segmentation weighting of the image. For information decoding loss, Visual enhancement loss (focusing frequency domain loss) +Regional guided loss ), To mitigate losses. (Default) =1, =2, =1, =0.001.

[0061] Later stage: After the decoding accuracy stabilizes, multi-region local loss is introduced. .like Figure 4 As shown, the residuals are no longer uniformly distributed in the later stages of training. At this point, the image is divided into multiple regions, the loss for each region is calculated, and weights are dynamically assigned based on the residual amount within each region. Regions with large residuals have high weights, and the summation results in multi-regional local losses. The total loss in the later stages is shown in the following formula: default =1, =1, =1, =0.001. This mechanism enables the model to focus on optimizing local regions where residuals are clustered, which greatly promotes the improvement of image quality in later stages.

[0062] Example effect:

[0063] On the COCO test set, the average metrics achieved by this method are: PSNR = 41.57 dB, SSIM = 0.992. After simulating a screen capture attack, the bit accuracy (BER) of watermark extraction reaches 98.32%. As shown in Table 1, this method outperforms the comparison algorithms in both visual quality and robustness against screen capture attacks.

[0064] Table 1. Performance comparison of each algorithm on the COCO dataset.

[0065]

[0066] To fully verify the performance of the proposed watermarking algorithm, it was tested on different datasets. The algorithm randomly selected 1000 images from each of the following datasets for experimentation. The results in Table 2 show that the watermarking algorithm not only performs well on the COCO dataset, but also outperforms it on the COCO dataset in terms of PSNR metrics on other datasets. Furthermore, the SSIM metric results on these two datasets are similar to those observed on the COCO dataset.

[0067] Table 2 Performance comparison of the present invention on different datasets

[0068]

[0069] In the experimental phase, ablation studies were conducted on the watermark region adjustment module, frequency enhancement module, and multi-region local loss proposed in this invention. Furthermore, the watermark region adjustment module was further divided into a grayscale module and an edge module for experimental testing. The model data used in the experiments were randomly selected from 1000 images in the COCO dataset as the original images for testing.

[0070] Table 3 Ablation Experiment Results

[0071]

[0072] The methods described above are presented and provided as website service functions. The core functions of the system are divided into two main modules: single image processing and batch tasks. Users can embed invisible digital watermark information into images through a simple upload operation, utilizing the method model described in this invention deployed in the backend. This achieves the covert embedding of copyright information while maintaining the high visual effect and image quality of the watermarked image. Similarly, the system also supports decoding and detection of watermarked images, accurately extracting hidden information and calculating matching accuracy to verify the effectiveness of the watermark. To meet the needs of different scenarios, the webpage is specially designed with a batch processing function, supporting asynchronous background tasks. Users can submit multiple images at once for watermark embedding or extraction and view task progress and statistical results in real time. In addition, the interface also integrates a system guide panel and a sample image library, making it convenient for new users to quickly get started and test, and intuitively experience the technical advantages of invisible watermarks in terms of anti-attack and concealment.

Claims

1. A deep learning-based visual enhancement robust digital dark watermarking method, characterized in that, Includes the following steps: Step 1: Collect the original image and watermark information; Step 2: Input the original image and the watermark information into the encoder to generate a watermarked image. The watermark embedding process of the encoder is guided by the visual enhancement module, which includes an adaptive region adjustment module for determining the priority embedding region based on image edges and grayscale information. Step 3: Set a noise layer and an adversarial discriminator between the encoder and the decoder. The noise layer is used to apply noise perturbation to the watermarked image to simulate the physical distortion of screen shooting, so as to obtain a noisy image. The adversarial discriminator and the encoder form an adversarial joint training framework. Step 4: Input the noisy image into the decoder to extract the watermark information; Step 5: During the training process, a multi-region local loss function is used for optimization. In the later stage of training, when the watermark residual tends to concentrate in the local area of ​​the image, the training weights are dynamically allocated according to the residual size of each region of the image.

2. The deep learning-based visual enhancement robust digital dark watermarking method according to claim 1, characterized in that, The adaptive region adjustment module generates a region guidance matrix. This allows for precise spatial awareness control of the network training process; the region guidance matrix... The size is consistent with the resolution of the original image. Each element in the matrix directly corresponds to a pixel position in the original image. A combination of matrix dot product and pixel-wise loss is used to achieve regional weighted control of the overall loss value. This allows the watermark information to be adaptively and preferentially embedded into visually insensitive low-grayscale, high-texture regions. The specific steps are as follows: Canny edge detection is performed on the original image to obtain an edge information matrix. The boundary region matrix value is set to 0, and the non-boundary region matrix value is set to 1; the original image is then converted to grayscale to obtain a grayscale information matrix. And perform normalization; according to the formula Calculate the region guidance matrix , where α and β are weighting coefficients.

3. The deep learning-based visual enhancement robust digital dark watermarking method according to claim 1, characterized in that, The visual enhancement module also includes a frequency enhancement module for calculating the focusing frequency domain loss. The expression is: ;in, This is a frequency optimization matrix used to weight different frequency components; This is a distance metric between the original image and the watermarked image in the frequency domain.

4. The deep learning-based visual enhancement robust digital dark watermarking method according to claim 1, characterized in that, The encoder is built on an improved U-Net network architecture, which includes a downsampling path, a global feature extraction layer, and an upsampling path. The three work together to achieve high-fidelity fusion of watermark information and deep image features. The downsampling path extracts multi-scale features and compresses spatial dimensions step by step through operation modules including convolution and max pooling, and constructs a feature pyramid from low-level texture to high-level semantics, providing robust feature representation for watermark embedding. The global feature extraction layer aggregates the overall statistical information and contextual semantics of the image through global pooling based on downsampling, providing global visual guidance for watermark embedding and ensuring that the embedding process is consistent with the overall structure of the image. The upsampling path gradually restores the feature map resolution through transposed convolution or interpolation operations, and fuses shallow detail features at the same scale through skip connections in each step, while integrating the processed watermark information into the feature stream. Finally, the fused features are mapped back to the image space through 1×1 convolution to generate a watermarked image with the same size and visual perception as the original image.

5. The deep learning-based visual enhancement robust digital dark watermarking method according to claim 1, characterized in that, The adversarial discriminator and encoder in step 3 form an adversarial joint training framework that continuously improves the imperceptibility and visual quality of the watermarked image through a game-theoretic mechanism; where the parameters are... Adversarial discriminator It is trained as a binary classification network to distinguish between the original image and the watermark image; the parameters are... encoder Used to store the original watermark information Embedded into the original image Watermark image generated later The specific training process is as follows: In each iteration, the adversarial discriminator... Receive raw image and watermarked images As input, and by maximizing the discriminant loss To update parameters To enhance its ability to distinguish between genuine and counterfeit goods, the loss is defined as... ,in This indicates that the discriminator considers the input image to be the original image. The confidence level output; At the same time, encoder The goal is to deceive the adversarial discriminator by minimizing the adversarial loss. To update parameters This forces the discriminator to output a confidence level close to 1 for the watermarked image, thereby driving the watermarked image to approximate the distribution of the original image in the feature space. The adversarial loss is defined as: .

6. The deep learning-based visual enhancement robust digital dark watermarking method according to claim 1, characterized in that, The calculation method for the multi-region local loss function is as follows: The focusing frequency domain loss for each region was calculated both in the early and late stages of training. and pixel mean square error loss ; In the early stages of training, the weighted sum of the calculated losses is used directly as the loss function: ,in These are the weighting coefficients; Decoder D extracts information from the watermark image to obtain the decoding result. By calculating the decoding result Compared with the original watermark information The mean square error between them yields the information decoding loss. , ; Loss With information decoding loss visual enhancement loss and combating losses The total prior period loss is obtained by performing a weighted summation. In the later stages of training, the image is divided into multiple regions based on different image resolutions; Weights are assigned to each region based on the residual size. The larger the residual, the higher the weight of the region; Total local loss ,in and They represent the first Focusing frequency domain loss of each image region and the MSE loss for each image region; Total local losses With information decoding loss visual enhancement loss and combating losses The total loss in the later stages is obtained by performing a weighted summation. in, To focus on frequency domain loss With regional guidance loss sum; Defined as a region guiding matrix The sum of all elements in the formula. , For the regional guiding matrix Located in row i, the The column element values, where L is the size of the matrix.

7. A storage medium, characterized in that, The storage medium stores at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or instruction set is loaded and executed by a processor to implement the deep learning-based visual enhancement robust digital dark watermarking method according to any one of claims 1-6.

8. A server, characterized in that, The server includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the deep learning-based visual enhancement robust digital dark watermarking method according to any one of claims 1-6.

Images