A training method and device of a blind watermark embedding and extraction model
By processing satellite imagery and watermarked images using convolutional neural networks, generating watermarked images and controlling the embedding strength, the problem of imperceptibility and robustness of watermarks in image processing is solved, achieving efficient copyright protection and information product production.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-08-31
- Publication Date
- 2026-03-31
AI Technical Summary
How to improve the imperceptibility of digital watermarks to achieve better copyright protection, especially how to balance robustness and imperceptibility in image processing.
By using a convolutional neural network to process satellite image tiles and watermark binary images, a watermark image is generated. The watermark embedding strength is controlled by a loss function. Combined with a Transformer Block and a decoding layer, an end-to-end blind watermark embedding and extraction model training is achieved.
It improves the imperceptibility of watermarks and the robustness of the model, is applicable to any payload remote sensing image in the high-resolution series, enhances the efficiency of information product production, and can resist image attacks.
Smart Images
Figure CN115456852B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital watermarking technology, and in particular to a training method and apparatus for a blind watermark embedding and extraction model. Background Technology
[0002] In recent years, the rapid growth of digital content has led to increasing concerns about copyright. Image watermarking is considered one of the most popular copyright protection methods. With the recent advancements in the application of deep neural networks in image processing, these networks are also being used in image watermarking technology. In the field of digital watermarking technology, robustness and imperceptibility are two challenging characteristics of watermarking methods, and a robust watermark embedding method should satisfy the trade-off between these two properties.
[0003] Improving the imperceptibility of embedded watermarks is a problem that urgently needs to be solved. Summary of the Invention
[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a training method and apparatus for a blind watermark embedding and extraction model.
[0005] The technical solution of this invention is:
[0006] In a first aspect, embodiments of the present invention provide a training method for a blind watermark embedding and extraction model, comprising:
[0007] Satellite image tiles and watermark binary images are input into the blind watermark embedding and extraction model to be trained;
[0008] The satellite image tiles are processed to obtain the satellite image features of the satellite image tiles;
[0009] The watermarked binary image is processed to obtain the watermark features of the watermarked binary image;
[0010] The satellite image features and the watermark features are concatenated to generate a watermarked image;
[0011] Based on the watermarked image, the satellite image tile, and the watermarked binary image, the loss value of the blind watermark embedding and extraction model to be trained is calculated.
[0012] If the loss value is within a preset range, the trained blind watermark embedding and extraction model is used as the final target blind watermark embedding and extraction model.
[0013] Optionally, the blind watermark embedding and extraction model to be trained includes: a mapping layer and an encoding layer.
[0014] The process of processing the satellite image tiles to obtain the satellite image features of the satellite image tiles includes:
[0015] The satellite image tiles are divided into segments to obtain satellite image blocks corresponding to the satellite image tiles;
[0016] The mapping layer is invoked to process the satellite image block, thereby obtaining the satellite image features of the satellite image tile.
[0017] Optionally, the blind watermark embedding and extraction model to be trained further includes: convolutional layers,
[0018] The process of processing the binary watermark image to obtain the watermark features of the binary watermark image includes:
[0019] The watermark binary image is processed by dilated convolution of multiple convolutional layers to obtain the watermark features of the watermark binary image.
[0020] Optionally, the blind watermark embedding and extraction model to be trained further includes: a decoding layer.
[0021] The step of concatenating the satellite image features and the watermark features to generate a watermarked image includes:
[0022] The satellite image features and the watermark features are concatenated to obtain concatenated features;
[0023] The decoding layer is invoked to process the connection features to obtain the watermark image.
[0024] Optionally, after using the trained blind watermark embedding and extraction model as the final target blind watermark embedding and extraction model, the method further includes:
[0025] Acquire test satellite image tiles and test watermark binary images;
[0026] The target blind watermark embedding and extraction model is invoked to perform watermark embedding processing on the test satellite image tiles based on the test watermark binary image to obtain the test watermark image;
[0027] The target blind watermark embedding and extraction model is invoked to perform image attack and watermark extraction operations on the test watermark image to obtain a predicted watermark image.
[0028] Based on the test watermark image and the predicted watermark image, determine the model evaluation index of the target blind watermark embedding and extraction model;
[0029] Based on the model evaluation metrics, the model accuracy of the target blind watermark embedding and extraction model is determined.
[0030] Secondly, embodiments of the present invention provide a training apparatus for a blind watermark embedding and extraction model, comprising:
[0031] The satellite image input module is used to input satellite image tiles and watermark binary images into the blind watermark embedding and extraction model to be trained;
[0032] The image feature acquisition module is used to process the satellite image tiles to obtain the satellite image features of the satellite image tiles;
[0033] The watermark feature acquisition module is used to process the watermark binary image to obtain the watermark features of the watermark binary image;
[0034] The watermark image generation module is used to connect the satellite image features and the watermark features to generate a watermark image;
[0035] The loss value calculation module is used to calculate the loss value of the blind watermark embedding and extraction model to be trained based on the watermark image, the satellite image tile and the watermark binary image;
[0036] The target watermark embedding model acquisition module is used to take the trained blind watermark embedding and extraction model as the final target blind watermark embedding and extraction model when the loss value is within a preset range.
[0037] Optionally, the blind watermark embedding and extraction model to be trained includes: a mapping layer and an encoding layer.
[0038] The image feature acquisition module includes:
[0039] The image block acquisition unit is used to divide the satellite image tiles to obtain satellite image blocks corresponding to the satellite image tiles;
[0040] The image feature acquisition unit is used to call the mapping layer to process the satellite image block and obtain the satellite image features of the satellite image tile.
[0041] Optionally, the blind watermark embedding and extraction model to be trained further includes: convolutional layers,
[0042] The watermark feature acquisition module includes:
[0043] The watermark feature acquisition unit is used to call multiple convolutional layers to perform dilated convolution processing on the watermark binary image to obtain the watermark features of the watermark binary image.
[0044] Optionally, the blind watermark embedding and extraction model to be trained further includes: a decoding layer.
[0045] The watermark image generation module includes:
[0046] A connection feature acquisition unit is used to perform connection processing on the satellite image features and the watermark features to obtain connection features;
[0047] The watermark image acquisition unit is used to call the decoding layer to process the connection features and obtain the watermark image.
[0048] Optionally, the device further includes:
[0049] The test sample acquisition module is used to acquire test satellite image tiles and test watermark binary images;
[0050] The test watermark image acquisition module is used to call the target blind watermark embedding and extraction model to perform watermark embedding processing on the test satellite image tiles according to the test watermark binary image to obtain the test watermark image.
[0051] The predicted watermark image acquisition module is used to call the target blind watermark embedding and extraction model to perform image attack and watermark extraction operations on the test watermark image to obtain the predicted watermark image.
[0052] The model evaluation index determination module is used to determine the model evaluation index of the target blind watermark embedding and extraction model based on the test watermark image and the predicted watermark image.
[0053] The model accuracy determination module is used to determine the model accuracy of the target blind watermark embedding and extraction model based on the model evaluation index.
[0054] The advantages of this invention compared to existing technologies are as follows: the embodiments of this invention utilize convolutional neural networks to control the embedding strength of digital watermarks, thereby improving the imperceptibility of embedded watermarks; the invention also utilizes neural networks to simulate different image attack methods, thereby improving the robustness of the model and ultimately achieving copyright protection and tracking. Furthermore, it can be applied to any payload remote sensing image in the high-resolution series, greatly improving the efficiency of information product production. Attached Figure Description
[0055] Figure 1 A flowchart illustrating the steps of a training method for a blind watermark embedding and extraction model provided in an embodiment of the present invention;
[0056] Figure 2 This is a schematic diagram of a watermark embedding process provided in an embodiment of the present invention;
[0057] Figure 3 This is a schematic diagram of the structure of a training device for a blind watermark embedding and extraction model provided in an embodiment of the present invention. Detailed Implementation
[0058] Example 1
[0059] Reference Figure 1 The diagram illustrates the steps of a training method for a blind watermark embedding and extraction model provided by an embodiment of the present invention. Figure 1 As shown, the method may include the following steps:
[0060] Step 101: Input the satellite image tiles and the binary watermark image into the blind watermark embedding and extraction model to be trained.
[0061] In this embodiment, when training the blind watermark embedding and extraction model to be trained, satellite image tiles and watermark binary images can be obtained, wherein the satellite image tiles are 256x256 image tiles and the watermark binary image is a 128x128 black and white image.
[0062] In this example, the number of satellite image tiles and watermarked binary images can be determined according to business needs, and this embodiment does not impose any restrictions on this.
[0063] After obtaining the satellite image tiles and the watermark binary image, the satellite image tiles and the watermark binary image can be input into the blind watermark embedding and extraction model to be trained.
[0064] In this embodiment, the structure of the blind watermark embedding and extraction model to be trained can be shown in Table 1 below:
[0065]
[0066] After inputting the satellite image tiles and the binary watermark image into the blind watermark embedding and extraction model to be trained, step 102 is executed.
[0067] Step 102: Process the satellite image tiles to obtain the satellite image features of the satellite image tiles.
[0068] After inputting satellite image tiles and watermark binary images into the blind watermark embedding and extraction model to be trained, the satellite image tiles can be processed to obtain the satellite image features of the satellite image tiles. Specifically, the following detailed implementation method can be described in detail.
[0069] In one specific implementation of this application, the blind watermark embedding and extraction model to be trained includes a mapping layer and an encoding layer, and step 102 above may include:
[0070] Sub-step A1: Divide the satellite image tiles to obtain the satellite image blocks corresponding to the satellite image tiles.
[0071] In this embodiment, the blind watermark embedding and extraction model to be trained includes a mapping layer and an encoding layer.
[0072] After inputting satellite image tiles into the blind watermark embedding and extraction model to be trained, the satellite image tiles can be first divided to obtain the corresponding satellite image blocks, such as... Figure 2 As shown, the original image is a 256x256 image tile. First, the image is divided into regular, non-overlapping patches.
[0073] After dividing the satellite image tiles into corresponding satellite image blocks, sub-step A2 is executed.
[0074] Sub-step A2: Call the mapping layer to process the satellite image block to obtain the satellite image features of the satellite image tile.
[0075] After dividing the satellite image tiles into corresponding satellite image blocks, a mapping layer can be called to process the satellite image blocks to obtain the satellite image features of the satellite image tiles, such as... Figure 2 As shown, the divided patches are fed into the encoder model. The encoder model first uses Linear Projection encoding, then performs position encoding on the data, and then feeds the encoded data into a series of consecutive Transformer Blocks to obtain the satellite image features of the satellite image tiles.
[0076] Step 103: Process the watermark binary image to obtain the watermark features of the watermark binary image.
[0077] In this embodiment, after inputting the binary watermark image into the blind watermark embedding and extraction model to be trained, the binary watermark image can be processed to obtain the watermark features of the binary watermark image. Specifically, this can be described in detail with reference to the following specific implementation method.
[0078] In one specific implementation of this application, the blind watermark embedding and extraction model to be trained further includes: a convolutional layer, and step 103 above may include:
[0079] Sub-step B1: Call the multiple convolutional layers to perform dilated convolution processing on the watermark binary image to obtain the watermark features of the watermark binary image.
[0080] In this embodiment, after inputting the binary watermark image into the blind watermark embedding and extraction model to be trained, multiple convolutional layers can be called to perform dilated convolution processing on the binary watermark image to obtain the watermark features of the binary watermark image. For example... Figure 2As shown, the watermark data is a 128x128 black and white image. Four dilated convolutional layers are used for feature extraction. Dilated convolution helps the network aggregate multi-scale contextual information without losing resolution. This helps the network have a global view of information and expands the hidden watermark information to a wider range. Pooling layers are not used at this stage because we want to preserve all the information in the input watermark data and maintain the transparency of the reconstructed image.
[0081] After processing the binary watermark image to obtain the watermark features, step 104 is executed.
[0082] Step 104: Connect the satellite image features and the watermark features to generate a watermarked image.
[0083] After obtaining satellite image features and watermark features through the above steps, the satellite image features and watermark features can be concatenated to generate a watermarked image. Specifically, the following detailed implementation method can be described in detail.
[0084] In one specific implementation of this application, the blind watermark embedding and extraction model to be trained further includes a decoding layer, and step 104 above may include:
[0085] Sub-step C1: Perform concatenation processing on the satellite image features and the watermark features to obtain concatenated features.
[0086] In this embodiment, after obtaining the satellite image features and watermark features, a concatenation process can be performed on the satellite image features and watermark features to obtain concatenated features. In this example, by concatenating the extracted image features with the watermark features before sending them to the decoder, the watermark features act as noise to prevent the network from memorizing and force the network to continuously learn.
[0087] Sub-step C2: Call the decoding layer to process the connection features to obtain the watermark image.
[0088] After obtaining the connection features, the decoding layer can be called to process the connection features to obtain the watermark image.
[0089] After obtaining the watermarked image, proceed to step 105.
[0090] Step 105: Calculate the loss value of the blind watermark embedding and extraction model to be trained based on the watermark image, the satellite image tile, and the watermark binary image.
[0091] After obtaining the watermark image, the loss value of the blind watermark embedding and extraction model to be trained can be calculated based on the watermark image, satellite image tiles, and watermark binary image.
[0092] In this embodiment, optimization functions, learning rate decay functions, loss functions, etc., can be selected, and parameters such as epoch, number of iterations, and parameter ratio can be set to train the model. The model loss function is shown in formula (1), where λ1 and λ2 are coefficients that control the trade-off between the two contrast features of the watermark and the image, respectively, and L1 is the loss function between the original image and the image with embedded watermark, representing the difference between the two images. In this invention, mean squared error (MSE) is used as the comparison index between the input image and the output image, as shown in formula (2). L2 is the loss function for the difference between the input watermark and the extracted watermark. In this invention, binary cross-entropy is used as the comparison index between the input and output watermarks, as shown in formula (3).
[0093] Loss=λ1L1+λ2L2 (1)
[0094]
[0095]
[0096] After calculating the loss value of the blind watermark embedding and extraction model to be trained based on the watermark image, satellite image tiles and watermark binary image, step 106 is executed.
[0097] Step 106: If the loss value is within a preset range, the trained blind watermark embedding and extraction model is used as the final target blind watermark embedding and extraction model.
[0098] After calculating the loss value of the blind watermark embedding and extraction model to be trained based on the watermark image and satellite image tiles, the convergence of the blind watermark embedding and extraction model to be trained can be determined based on the loss value.
[0099] If the loss value is within the preset range, it is determined that the trained blind watermark embedding and extraction model has converged. At this time, the trained blind watermark embedding and extraction model can be used as the final target blind watermark embedding and extraction model.
[0100] In this embodiment, after the initial training of the model to obtain the target blind watermark embedding and extraction model, test samples can be used for testing. Specifically, the following detailed implementation method can be described in detail.
[0101] In one specific implementation of this application, after step 106 above, the following may also be included:
[0102] Step D1: Obtain test satellite image tiles and test watermark binary images.
[0103] In this embodiment, the model test samples and model training samples can be obtained simultaneously. In this example, the ratio of the training set to the validation set can be preset to 7:3, etc.
[0104] After training and obtaining the target blind watermark embedding and extraction model, test satellite image tiles and test watermark binary images can be acquired.
[0105] After obtaining the test satellite image tiles and the test watermark binary image, proceed to step D2.
[0106] Step D2: Call the target blind watermark embedding and extraction model to perform watermark embedding processing on the test satellite image tiles based on the test watermark binary image to obtain the test watermark image.
[0107] After obtaining the test satellite image tiles and the test watermark binary image, the target blind watermark embedding and extraction model can be called to perform watermark embedding processing on the test satellite image tiles based on the test watermark binary image to obtain the test watermark image.
[0108] Step D3: Call the target blind watermark embedding and extraction model to perform image attack and watermark extraction operations on the test watermark image to obtain the predicted watermark image.
[0109] After obtaining the test watermark image, the target blind watermark embedding and extraction model can be called to perform image attack and watermark extraction operations on the test watermark image to obtain the predicted watermark image.
[0110] Step D4: Determine the model evaluation index of the target blind watermark embedding and extraction model based on the test watermark image and the predicted watermark image.
[0111] After obtaining the predicted watermark image, the model evaluation index of the target watermark embedding and extraction model can be determined based on the test watermark image and the predicted watermark image.
[0112] Step D5: Determine the model accuracy of the target blind watermark embedding and extraction model based on the model evaluation index.
[0113] After obtaining the model evaluation index of the target watermark embedding and extraction model, the model accuracy of the target blind watermark embedding and extraction model can be determined based on the model evaluation index.
[0114] In this embodiment, PSNR and SSIM parameters can be used as model evaluation indicators to evaluate model accuracy from multiple angles and in all aspects, as shown in the following formulas (4) and (5):
[0115]
[0116] SSIM = [l(x,y)]α [c(x,y)] β [s(x,y)] γ (5)
[0117] The method of this invention is applicable to adding digital watermarks to image tiles. Compared with the traditional method of embedding information in the frequency domain, the method of this invention uses the MAE model to hide information in the feature map space, which is more robust than the traditional method and can resist alteration or unauthorized access.
[0118] The network model of the method of this invention has high robustness, which helps to integrate watermark information into the image without being easily detected.
[0119] The method of this invention is an end-to-end network model that attempts to improve both robustness and imperceptibility simultaneously, and uses a loss function to control the trade-off between these two properties. At the same time, this method adopts an asymmetric model design, which helps to shorten the pre-training time of the network.
[0120] The method of this invention uses a transformer network, which is not only efficient and convenient, but also overcomes the instability and concealment of traditional watermark embedding methods. At the same time, the similarity between the watermarked image and the original image is as high as 0.99, which will not affect the user's viewing experience of the image.
[0121] The method of this invention has good transferability and is applicable to any payload remote sensing image of the high-resolution series, which greatly improves the efficiency of information product production.
[0122] Example 2
[0123] Reference Figure 3 The diagram shows a schematic representation of the structure of a training device for a blind watermark embedding and extraction model provided in an embodiment of the present invention. Figure 3 As shown, the device may include the following modules:
[0124] The satellite image input module 310 is used to input satellite image tiles and watermark binary images into the blind watermark embedding and extraction model to be trained;
[0125] The image feature acquisition module 320 is used to process the satellite image tiles to obtain the satellite image features of the satellite image tiles;
[0126] The watermark feature acquisition module 330 is used to process the watermark binary image to obtain the watermark features of the watermark binary image;
[0127] The watermark image generation module 340 is used to connect the satellite image features and the watermark features to generate a watermark image;
[0128] The loss value calculation module 350 is used to calculate the loss value of the blind watermark embedding and extraction model to be trained based on the watermark image, the satellite image tile and the watermark binary image.
[0129] The target watermark embedding model acquisition module 360 is used to take the trained blind watermark embedding and extraction model as the final target blind watermark embedding and extraction model when the loss value is within a preset range.
[0130] Optionally, the blind watermark embedding and extraction model to be trained includes: a mapping layer and an encoding layer.
[0131] The image feature acquisition module includes:
[0132] The image block acquisition unit is used to divide the satellite image tiles to obtain satellite image blocks corresponding to the satellite image tiles;
[0133] The image feature acquisition unit is used to call the mapping layer to process the satellite image block and obtain the satellite image features of the satellite image tile.
[0134] Optionally, the blind watermark embedding and extraction model to be trained further includes: convolutional layers,
[0135] The watermark feature acquisition module includes:
[0136] The watermark feature acquisition unit is used to call multiple convolutional layers to perform dilated convolution processing on the watermark binary image to obtain the watermark features of the watermark binary image.
[0137] Optionally, the blind watermark embedding and extraction model to be trained further includes: a decoding layer.
[0138] The watermark image generation module includes:
[0139] A connection feature acquisition unit is used to perform connection processing on the satellite image features and the watermark features to obtain connection features;
[0140] The watermark image acquisition unit is used to call the decoding layer to process the connection features and obtain the watermark image.
[0141] Optionally, the device further includes:
[0142] The test sample acquisition module is used to acquire test satellite image tiles and test watermark binary images;
[0143] The test watermark image acquisition module is used to call the target blind watermark embedding and extraction model to perform watermark embedding processing on the test satellite image tiles according to the test watermark binary image to obtain the test watermark image.
[0144] The predicted watermark image acquisition module is used to call the target blind watermark embedding and extraction model to perform image attack and watermark extraction operations on the test watermark image to obtain the predicted watermark image.
[0145] The model evaluation index determination module is used to determine the model evaluation index of the target blind watermark embedding and extraction model based on the test watermark image and the predicted watermark image.
[0146] The model accuracy determination module is used to determine the model accuracy of the target blind watermark embedding and extraction model based on the model evaluation index.
[0147] The specific embodiments described in this application are intended to enable those skilled in the art to gain a more comprehensive understanding of this application, but do not limit this application in any way. Therefore, those skilled in the art should understand that modifications or equivalent substitutions can still be made to this application; and all technical solutions and improvements that do not depart from the spirit and technical essence of this application should be covered within the scope of protection of this patent application.
[0148] The contents not described in detail in this specification are common knowledge to those skilled in the art.
Claims
1. A method for training a blind watermark embedding and extraction model, characterized in that, The method comprises the steps of: inputting a satellite image tile and a watermark binary image into a blind watermark embedding and extraction model to be trained; the blind watermark embedding and extraction model is based on an MAE architecture, and comprises a mapping layer, an encoding layer, a convolution layer and a decoding layer; performing division processing on the satellite image tile to obtain a satellite image block corresponding to the satellite image tile; performing linear mapping by calling the mapping layer, performing position encoding by calling the encoding layer, and then inputting into a continuous Transformer Block to extract satellite image features of the satellite image tile; performing hole convolution processing on the watermark binary image by calling multiple layers of the convolution layer to obtain watermark features of the watermark binary image; performing connection processing on the satellite image features and the watermark features to obtain connection features; performing processing on the connection features by calling the decoding layer to obtain a watermark image; calculating a loss value of the blind watermark embedding and extraction model to be trained according to the watermark image, the satellite image tile and the watermark binary image; in a case where the loss value is within a preset range, taking the trained blind watermark embedding and extraction model to be trained as a final target blind watermark embedding and extraction model.
2. The method of claim 1, wherein, After the step of taking the trained blind watermark embedding and extraction model to be trained as the final target blind watermark embedding and extraction model, the method further comprises the steps of: obtaining a test satellite image tile and a test watermark binary image; performing watermark embedding processing on the test satellite image tile according to the test watermark binary image by calling the target blind watermark embedding and extraction model to obtain a test watermark image; performing image attack and watermark extraction operations on the test watermark image by calling the target blind watermark embedding and extraction model to obtain a predicted watermark image; determining a model evaluation index of the target blind watermark embedding and extraction model according to the test watermark image and the predicted watermark image; determining a model precision of the target blind watermark embedding and extraction model according to the model evaluation index.
3. A device for training a blind watermark embedding and extraction model, characterized in that, The method comprises the steps of: a satellite image input module is configured to input a satellite image tile and a watermark binary image into a blind watermark embedding and extraction model to be trained; the blind watermark embedding and extraction model is based on an MAE architecture, and comprises a mapping layer, an encoding layer, a convolution layer and a decoding layer; an image feature acquisition module is configured to perform division processing on the satellite image tile to obtain a satellite image block corresponding to the satellite image tile; perform linear mapping by calling the mapping layer, perform position encoding by calling the encoding layer, and then input into a continuous Transformer Block to extract satellite image features of the satellite image tile; a watermark feature acquisition module is configured to perform hole convolution processing on the watermark binary image by calling multiple layers of the convolution layer to obtain watermark features of the watermark binary image; a watermark image generation module is configured to perform connection processing on the satellite image features and the watermark features to obtain connection features; perform processing on the connection features by calling the decoding layer to obtain a watermark image; a loss value calculation module is configured to calculate a loss value of the blind watermark embedding and extraction model to be trained according to the watermark image, the satellite image tile and the watermark binary image. The target watermark embedding model acquisition module is configured to, in a case where the loss value is within a preset range, take the trained blind watermark embedding and extraction model to be trained as a final target blind watermark embedding and extraction model.
4. The apparatus of claim 3, wherein, The device further includes: A test sample acquisition module configured to acquire a test satellite image tile and a test watermark binary image; A test watermark image acquisition module configured to call the target blind watermark embedding and extraction model to perform watermark embedding processing on the test satellite image tile according to the test watermark binary image, to obtain a test watermark image; A predicted watermark image acquisition module configured to call the target blind watermark embedding and extraction model to perform image attack and watermark extraction operations on the test watermark image, to obtain a predicted watermark image; A model evaluation index determination module configured to determine a model evaluation index of the target blind watermark embedding and extraction model according to the test watermark image and the predicted watermark image; A model precision determination module configured to determine a model precision of the target blind watermark embedding and extraction model according to the model evaluation index.
Citation Information
Patent Citations
Self-coding color image robust watermark processing method based on visual perception
CN114841846A