Digital image watermark embedding and extracting method and device and storage medium

Abstract

The invention discloses a digital image watermark embedding and extracting method and device, and a storage medium, and the method comprises the steps: combining histogram displacement with neural network embedding, enabling an embedded data flow to generate the minimum noise under JPEG compression through the histogram displacement after DCT transformation, and enabling the embedded data flow to generate the minimum noise under JPEG compression; and the influence of the data stream on the host image is smaller through a neural network compensation algorithm. By adopting the technical scheme of the invention, the problem that the existing neural network embedded digital watermark technology is poor in performance when coping with the JPEG compression attack is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image information hiding, and in particular relates to a digital image watermark embedding and extraction method and device, and a storage medium. Background Art

[0002] Neural network digital watermarking technology has been widely used due to its good concealment and small impact on image quality. This technology achieves low-distortion watermark embedding by learning the characteristics of the host image. However, neural networks have certain limitations in fitting deterministic mathematical functions, which makes their learning ability for simple and clear paradigms relatively weak. For example, neural networks are generally difficult to accurately learn and fit precise mathematical function models.

[0003] In the process of Internet image transmission, websites usually use JPEG compression format to reduce the demand for upstream and downstream bandwidth. Therefore, images are mainly exposed to attacks based on JPEG compression during transmission, and may also be affected by other types of attacks such as Gaussian noise and salt and pepper noise. The performance of neural network digital watermarking technology for JPEG compression is still obviously insufficient. The main reason is that the discrete cosine transform (DCT) involved in JPEG compression and the quantization table under different quality coefficients are both expressed as complex mathematical function relationships, and these function characteristics are difficult for neural networks to learn and accurately fit. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a digital image watermark embedding and extraction method and device, and a storage medium to solve the problem that the existing neural network embedding digital watermark technology performs poorly when dealing with JPEG compression attacks.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A digital image watermark embedding and extraction method, comprising the following steps:

[0007] Step S1: Information embedding

[0008] Step 11: Entropy decode the host image X to obtain all its discrete cosine transform DCT blocks, and reorganize the (1,1) block in the block, that is, the DC coefficient block, to form a DC vector

[0009] a (1,1) ={a1[1,1],a2[1,1]......} and 63 AC vectors a (1,2) ,...,a (8,8) ;

[0010] Step 12: Use DC prediction neural network based on AC vector a (1,2) ,...,a(8,8) Predicted DC vector The prediction error e is calculated by subtracting the predicted DC vector from the actual DC vector:

[0011] Step 13: By shifting the histogram, the embedded data stream can produce the minimum noise under JPEG compression, and the watermark data stream m is embedded into the prediction error e to generate the prediction error e with the watermark. w ;

[0012] Step 14: Use the formula Reconstruct the DC vector b with watermark (1,1) , b (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0013] Step 15, entropy encoding is performed on all modified DCT blocks to generate a watermark image Y embedded with a watermark;

[0014] Step S2: Extracting secret information and reconstructing the image

[0015] Step 21: Secret image Perform entropy decoding to obtain all its discrete cosine transform DCT blocks, and reorganize the (1,1) block in the block, that is, the DC coefficient block, to form a DC vector and 63 AC vectors (high frequency vectors)

[0016] Step 22: Use DC prediction neural network based on AC vector Predicted DC vector The prediction error is calculated by subtracting the predicted DC vector from the actual DC vector

[0017] Step 23: From the prediction error Extract the watermark m' and restore the prediction error e';

[0018] Step 24: Use the formula Reconstruct the DC vector a' (1,1) , a' (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0019] Step 25: Perform entropy coding on all modified DCT blocks to generate a restored image X'.

[0020] Preferably, in step 13,

[0021]

[0022] Among them, q is the quantization coefficient corresponding to the (1,1) block position in the quantization table in the JPEG algorithm, and the quantization coefficient of the quantization table with quality factor QF=50 has the best effect; ι, β are hyperparameters, ι is used to control the size of the histogram shift, and β is used to control the amount of watermark data embedded and the size of the histogram shift.

[0023] Preferably, in step 23,

[0024]

[0025] Preferably, the DC prediction neural network includes: a predictor, a generator and a discriminator based on PatchGAN; wherein the predictor is based on the U-Net architecture and consists of 5 upsampling modules and 5 downsampling modules; the total loss of the generator is a balanced combination of prediction loss, image loss and adversarial loss.

[0026] The present invention also provides a digital image watermark embedding and extraction device, comprising: an information embedding module and a secret information extraction and image reconstruction module; wherein,

[0027] The information embedding module includes:

[0028] The first processing unit is used to perform entropy decoding on X to obtain all discrete cosine transform DCT blocks thereof, and reorganize the (1,1) block, i.e., the DC coefficient block, in the block to form a DC vector

[0029] a (1,1) ={a1[1,1],a2[1,1]......} and 63 AC vectors a (1,2) ,...,a (8,8) ;

[0030] The second processing unit is used to use a DC prediction neural network based on the AC vector a (1,2) ,...,a (8,8) Predicted DC vector The prediction error e is calculated by subtracting the predicted DC vector from the actual DC vector:

[0031] The third processing unit is used to embed the watermark data stream m into the prediction error e by shifting the histogram so that the embedded data stream can generate the minimum noise under JPEG compression, and generate the prediction error e with the watermark w ;

[0032] The fourth processing unit is used to pass the formula Reconstruct the DC vector b with watermark (1,1) , b (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0033] a fifth processing unit, configured to perform entropy coding on all modified DCT blocks to generate an image Y embedded with a watermark;

[0034] The modules for extracting secret information and reconstructing images include:

[0035] The sixth processing unit is used to process the secret image Perform entropy decoding to obtain all its discrete cosine transform DCT blocks, and reorganize the (1,1) block in the block, that is, the DC coefficient block, to form a DC vector and 63 AC vectors (high frequency vectors)

[0036] The seventh processing unit is used to predict the AC vector using a DC neural network. Predicted DC vector The prediction error is calculated by subtracting the predicted DC vector from the actual DC vector

[0037] The eighth processing unit is used to obtain the prediction error Extract the watermark m' and restore the prediction error e';

[0038] The ninth processing unit is used to pass the formula Reconstruct the DC vector a' (1,1) , a' (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0039] The tenth processing unit is configured to perform entropy coding on all modified DCT blocks to generate a restored image X'.

[0040] Preferably, in the third processing unit,

[0041]

[0042] Among them, q is the quantization coefficient corresponding to the (1,1) block position in the quantization table in the JPEG algorithm, and the quantization coefficient of the quantization table with quality factor QF=50 has the best effect; ι, β are hyperparameters, ι is used to control the size of the histogram shift, and β is used to control the amount of watermark data embedded and the size of the histogram shift.

[0043] Preferably, in the eighth processing unit,

[0044]

[0045] Preferably, the DC prediction neural network includes: a predictor, a generator and a discriminator based on PatchGAN; wherein the predictor is based on the U-Net architecture and consists of 5 upsampling modules and 5 downsampling modules; the total loss of the generator is a balanced combination of prediction loss, image loss and adversarial loss.

[0046] The present invention also provides a storage medium, on which a computer program is stored, and the computer program executes the digital image watermark embedding and extraction method when running.

[0047] In order to reduce the complexity of neural network learning, the present invention introduces a histogram shift paradigm to solve the solution to minimize the noise impact of JPEG compression on digital watermarks under different conditions, thereby effectively improving the robustness and reversibility of the watermark while reducing the learning burden of the neural network. Specifically, the location of the watermark embedding is selected in the low-frequency coefficient (DC) area of ​​the image after discrete cosine transform (DCT), which makes the embedded watermark more robust. At the same time, in order to further ensure the concealment of the digital watermark, the present invention adopts a neural network compensation mechanism to minimize the impact of watermark embedding on image quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0049] Figure 1 This is a flow chart of a digital image watermark embedding and extraction method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] Embodiment 1:

[0053] like Figure 1As shown, the embodiment of the present invention provides a digital image watermark embedding and extraction method, which combines histogram shift with neural network embedding. Different from the existing neural network embedded watermark, it is designed specifically for JPEG compression attacks. After DCT transformation, the histogram shift is first used to make the embedded data stream produce the minimum noise under JPEG compression, and then the neural network compensation algorithm is used to make the data stream have less impact on the host image. Specifically, it includes the following steps:

[0054] Step S1: Information embedding

[0055] Step 11: Entropy decode the host image X to obtain all its discrete cosine transform (DCT) blocks, and reorganize the (1,1) blocks in these blocks, namely the DC coefficient (low-frequency coefficient) blocks, to form a DC vector a (1,1) ={a1[1,1],a2[1,1]......} and 63 AC vectors (high frequency vectors) a (1,2) ,...,a (8,8) ;

[0056] Step 12: Use DC prediction neural network based on AC vector a (1,2) ,...,a (8,8) Predicted DC vector The prediction error e is calculated by subtracting the predicted DC vector from the actual DC vector:

[0057] Step 13: Using formula (1), the embedded data stream m is embedded into the prediction error e by histogram shifting to generate the minimum noise under JPEG compression, and the prediction error e with watermark is generated. w ; The choice of histogram shift depends on how to make the watermark produce less noise after JPEG compression;

[0058] Step 14: Use the formula Reconstruct the DC vector b with watermark (1,1) , b (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0059] Step 15: Perform entropy coding on all modified DCT blocks to generate a watermarked image Y embedded with a watermark.

[0060] Formula (1) is as follows:

[0061]

[0062] Among them, q is the quantization coefficient corresponding to the (1,1) block position in the quantization table in the JPEG algorithm. The quantization coefficient of the quantization table with quality factor QF=50 has the best effect; ι, β are hyperparameters, ι is used to control the size of the histogram shift, and β is used to control the amount of watermark data embedded and the size of the histogram shift. The coefficients carrying the watermark bit 0 are adjacent to the coefficients that are not used to embed the watermark. External interference may easily move these coefficients to the wrong side, so they are more susceptible to extraction errors than the coefficients carrying the watermark bit 1. However, when restoring the main coefficients, these coefficients do not need to be modified, ensuring that even if errors occur when extracting the watermark bit 0, it will not affect the quality of the restored image. In addition, the number of bit 0 in the watermark sequence can be predetermined, denoted as N0, which helps to extract the watermark bit 0. In the extraction stage, select the coefficients located in the interval The N0 coefficients in the watermark sequence that are at the largest distance from the origin are assumed to be originally used to embed the watermark bit 0. Note that N0 does not need to be transmitted to the receiver, as it can be pre-set to half the length of the watermark sequence, considering that the number of bits 0 and 1 in the encrypted message sequence is roughly the same. In addition, the message sequence can be truncated or supplemented to ensure that the number of bits 0 matches N0.

[0063] Further, in step 12, the AC coefficient of the image is used as the input of the prediction neural network, and the task is to predict the corresponding DC coefficient. The DC prediction neural network includes: a predictor, a generator, and a discriminator; wherein the predictor is based on the U-Net architecture, consisting of 5 upsampling modules and 5 downsampling modules, and the discriminator is a discriminator based on PatchGAN to provide adversarial supervision. The input image X is transformed by 8×8 block DCT to separate the AC coefficient and DC coefficient.

[0064] To enhance the robustness of the network under noisy communication channels, data augmentation was performed on the AC coefficients during training, including the addition of Gaussian noise (μ = 0, σ = 0.1) and JPEG distortion (the quality factor QF varies randomly between 50 and 100). This augmentation strategy enables the predictor to accurately predict the DC coefficient in noisy and perturbed environments.

[0065] The total loss of the generator is a balanced combination of prediction loss, image loss, and adversarial loss, and its formula is: L = λ1L pred +λ2L img +λ3L adv . Among them, λ1, λ2, λ3 are hyperparameters that adjust the importance of each loss.

[0066] Prediction loss L pred Defined as the predicted DC coefficient and the true value a (1,1) The L2 distance between:

[0067] Adversarial loss L adv Aims to improve generation performance through a competitive learning environment between the generator and the discriminator.

[0068] In addition, the image loss L is introduced img , to maintain the consistency between the predicted DC coefficient and the corresponding AC coefficient. This loss measures the difference between the original image X and the predicted DC coefficient. With the original AC coefficient {a (1,2) ,...,a (8,8) The L2 distance between the reconstructed images generated by the combination is defined as: This loss helps to restore a high-fidelity image by using consistent DC and AC coefficients.

[0069] Step S2: Extracting secret information and reconstructing the image

[0070] The receiver will receive a secret image The image may be interfered by the channel, making it not equal to the image sent by the sender. Extracting secret information and reconstructing the image specifically include:

[0071] Step 21: Perform entropy decoding to obtain all its discrete cosine transform (DCT) blocks, and reorganize the (1,1) blocks in these blocks, that is, the DC coefficient (low-frequency coefficient) blocks, to form a DC vector and 63 AC vectors (high frequency vectors)

[0072] Step 22: Use DC prediction neural network based on AC vector Predicted DC vector The prediction error is calculated by subtracting the predicted DC vector from the actual DC vector

[0073] Step 23: Use formula (2) to calculate the prediction error Extract the watermark m' and restore the prediction error e';

[0074] Step 24: Use the formula Reconstruct the DC vector a' (1,1) , a' (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0075] Step 25: Perform entropy coding on all modified DCT blocks to generate a restored image X'.

[0076] Formula (2) is as follows:

[0077]

[0078] The embodiment of the present invention optimizes the watermark embedding position under JPEG compression by introducing histogram shift, thereby minimizing the noise generated by the embedded data after compression, and at the same time combines the neural network compensation mechanism to improve the concealment of watermark embedding and image quality.

[0079] In the embodiment of the present invention, the watermark embedding position selects the low-frequency coefficients (DC block) in the DCT domain, which significantly improves the robustness of the watermark under JPEG compression and other channel attacks, while reducing the impact on the high-frequency details of the image.

[0080] The embodiment of the present invention uses a DC prediction neural network based on the U-Net architecture, combined with a PatchGAN discriminator, and uses Gaussian noise and JPEG distortion data enhancement strategies to improve the robustness to compression and noise attacks and ensure high accuracy of watermark extraction and image restoration.

[0081] The embodiment of the present invention adopts the method of efficiently extracting watermark data from the prediction error and using the prediction neural network to restore the low-frequency coefficients of the image, thereby ensuring the dual protection of the watermark and the original image quality during the image reconstruction process.

[0082] Embodiment 2:

[0083] The embodiment of the present invention also provides a digital image watermark embedding and extraction device, including: an information embedding module and a secret information extraction and image reconstruction module; wherein,

[0084] The information embedding module includes:

[0085] The first processing unit is used to perform entropy decoding on X to obtain all discrete cosine transform DCT blocks thereof, and reorganize the (1,1) block, i.e., the DC coefficient block, in the block to form a DC vector

[0086] a (1,1) ={a1[1,1],a2[1,1]......} and 63 AC vectors a (1,2) ,...,a (8,8) ;

[0087] The second processing unit is used to use a DC prediction neural network based on the AC vector a (1,2) ,...,a (8,8) Predicted DC vector The prediction error e is calculated by subtracting the predicted DC vector from the actual DC vector:

[0088] The third processing unit is used to embed the watermark data stream m into the prediction error e by shifting the histogram so that the embedded data stream can generate the minimum noise under JPEG compression, and generate the prediction error e with the watermark w ;

[0089] The fourth processing unit is used to pass the formula Reconstruct the DC vector b with watermark (1,1) , b (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0090] a fifth processing unit, configured to perform entropy coding on all modified DCT blocks to generate an image Y embedded with a watermark;

[0091] The modules for extracting secret information and reconstructing images include:

[0092] The sixth processing unit is used to process the secret image Perform entropy decoding to obtain all its discrete cosine transform DCT blocks, and reorganize the (1,1) block in the block, that is, the DC coefficient block, to form a DC vector and 63 AC vectors (high frequency vectors)

[0093] The seventh processing unit is used to predict the AC vector using a DC neural network. Predicted DC vector The prediction error is calculated by subtracting the predicted DC vector from the actual DC vector

[0094] The eighth processing unit is used to obtain the prediction error Extract the watermark m' and restore the prediction error e';

[0095] The ninth processing unit is used to pass the formula Reconstruct the DC vector a' (1,1) , a' (1,1) Each coefficient in is placed back to its original position in the corresponding DCT block;

[0096] The tenth processing unit is configured to perform entropy coding on all modified DCT blocks to generate a restored image X'.

[0097] As an implementation of the embodiment of the present invention, in the third processing unit,

[0098]

[0099] Among them, q is the quantization coefficient corresponding to the (1,1) block position in the quantization table in the JPEG algorithm, and the quantization coefficient of the quantization table with quality factor QF=50 has the best effect; ι, β are hyperparameters, ι is used to control the size of the histogram shift, and β is used to control the amount of watermark data embedded and the size of the histogram shift.

[0100] As an implementation of the embodiment of the present invention, in the eighth processing unit,

[0101]

[0102] As an implementation method of an embodiment of the present invention, a DC prediction neural network includes: a predictor, a generator and a discriminator based on PatchGAN; wherein the predictor is based on a U-Net architecture and consists of 5 upsampling modules and 5 downsampling modules; the total loss of the generator is a balanced combination of prediction loss, image loss and adversarial loss.

[0103] Embodiment 3:

[0104] An embodiment of the present invention further provides a storage medium, on which a computer program is stored, and the computer program executes the digital image watermark embedding and extraction method when running.

[0105] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A digital image watermark embedding and extraction method, characterized in that: The following steps are involved: Step S1: Information embedding Step 11: Entropy decode the host image X to obtain all its discrete cosine transform DCT blocks, and reorganize the (1, 1) block in the block, that is, the DC coefficient block, to form a DC vector a (1，1) ={a1[1,1], a2[1,1]...} and 63 AC vectors a (1，2) , ..., a (8，8) ; Step 12: Use DC prediction neural network based on AC vector a (1，2) , ..., a (8，8) Predicted DC vector The prediction error e is calculated by subtracting the predicted DC vector from the actual DC vector: Step 13: By shifting the histogram, the embedded data stream can produce the minimum noise under JPEG compression, and the watermark data stream m is embedded into the prediction error e to generate the prediction error e with the watermark. w ; Step 14: Use the formula Reconstruct the DC vector b with watermark (1，1) , b (1，1) Each coefficient in is placed back to its original position in the corresponding DCT block; Step 15, entropy encoding is performed on all modified DCT blocks to generate a watermark image Y embedded with a watermark; Step S2: Extracting secret information and reconstructing the image Step 21: Secret image Entropy decoding is performed to obtain all discrete cosine transform DCT blocks, and the (1, 1) block in the block, that is, the DC coefficient block, is reorganized to form a DC vector and 63 AC vectors (high frequency vectors) Step 22: Use DC prediction neural network based on AC vector Predicted DC vector The prediction error is calculated by subtracting the predicted DC vector from the actual DC vector Step 23: From the prediction error Extract the watermark m' and restore the prediction error e'; Step 24: Use the formula Reconstruct the DC vector a' (1，1) , a' (1，1) Each coefficient in is placed back to its original position in the corresponding DCT block; Step 25: Perform entropy coding on all modified DCT blocks to generate a restored image X'.

2. The digital image watermark embedding and extraction method according to claim 1, characterized in that: In step 13, Among them, q is the quantization coefficient corresponding to the (1,1) block position in the quantization table in the JPEG algorithm, and the quantization coefficient of the quantization table with quality factor QF=50 has the best effect; ι, β are hyperparameters, ι is used to control the size of the histogram shift, and β is used to control the amount of watermark data embedded and the size of the histogram shift.

3. The digital image watermark embedding and extraction method as claimed in claim 2, characterized in that: In step 23, 4. The digital image watermark embedding and extraction method as claimed in claim 3, characterized in that: The DC prediction neural network includes: a predictor, a generator and a discriminator based on PatchGAN; the predictor is based on the U-Net architecture and consists of 5 upsampling modules and 5 downsampling modules; the total loss of the generator is a balanced combination of prediction loss, image loss and adversarial loss.

5. A digital image watermark embedding and extraction device, characterized in that: include: Information embedding module and secret information extraction and image reconstruction module; wherein, The information embedding module includes: The first processing unit is used to perform entropy decoding on X to obtain all discrete cosine transform DCT blocks thereof, and reorganize the (1, 1) block in the block, i.e., the DC coefficient block, to form a DC vector a (1，1) ={a1[1,1], a2[1,1]...} and 63 AC vectors a (1，2) , ..., a (8，8) ; The second processing unit is used to use a DC prediction neural network based on the AC vector a (1，2) , ..., a (8，8) Predicted DC vector The prediction error e is calculated by subtracting the predicted DC vector from the actual DC vector: The third processing unit is used to embed the watermark data stream m into the prediction error e by shifting the histogram so that the embedded data stream can generate the minimum noise under JPEG compression, and generate the prediction error e with the watermark w ; The fourth processing unit is used to pass the formula Reconstruct the DC vector b with watermark (1，1) , b (1，1) Each coefficient in is placed back to its original position in the corresponding DCT block; a fifth processing unit, configured to perform entropy coding on all modified DCT blocks to generate an image Y embedded with a watermark; The modules for extracting secret information and reconstructing images include: The sixth processing unit is used to process the secret image Entropy decoding is performed to obtain all discrete cosine transform DCT blocks, and the (1, 1) block in the block, that is, the DC coefficient block, is reorganized to form a DC vector and 63 AC vectors (high frequency vectors) The seventh processing unit is used to predict the AC vector using a DC neural network. Predicted DC vector The prediction error is calculated by subtracting the predicted DC vector from the actual DC vector The eighth processing unit is used to obtain the prediction error Extract the watermark m' and restore the prediction error e'; The ninth processing unit is used to pass the formula Reconstruct the DC vector a' (1，1) , a' (1，1) Each coefficient in is placed back to its original position in the corresponding DCT block; The tenth processing unit is configured to perform entropy coding on all modified DCT blocks to generate a restored image X'.

6. The digital image watermark embedding and extraction device as claimed in claim 5, characterized in that: In the third processing unit, Among them, q is the quantization coefficient corresponding to the (1, 1) block position in the quantization table in the JPEG algorithm, and the quantization coefficient of the quantization table with quality factor QF=50 has the best effect; ι, β are hyperparameters, ι is used to control the size of the histogram displacement, and β is used to control the amount of watermark data embedded and the size of the histogram displacement.

7. The digital image watermark embedding and extraction device as claimed in claim 6, characterized in that: In the eighth processing unit, 8. The digital image watermark embedding and extraction device as claimed in claim 7, characterized in that: The DC prediction neural network includes: a predictor, a generator and a discriminator based on PatchGAN; the predictor is based on the U-Net architecture and consists of 5 upsampling modules and 5 downsampling modules; the total loss of the generator is a balanced combination of prediction loss, image loss and adversarial loss.

9. A storage medium, characterized in that: The storage medium stores a computer program, which executes the digital image watermark embedding and extraction method as claimed in any one of claims 1 to 4 when running.

Images