A robust watermarking system and algorithm against screen-shooting based on deep learning

Abstract

The present invention discloses a deep learning-based robust watermarking system and algorithm for resisting screen shooting, comprising: an encoder, a discriminator and a noise network module connected to the encoder signal, and a decoder connected to the noise network module signal; the encoder is used to construct an information film with secret information to be embedded; the discriminator is used to distinguish whether an image is encoded; the noise network module is used to attack the encoded image; the decoder is used to use the attacked encoded image to recover the embedded secret information. According to the present invention, while ensuring a certain image quality, it has stronger robustness and practicality, and the output information film containing only information can be attached to any image, which has more practical application value, and can also achieve almost 100% extraction accuracy in various real shooting screen scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of multimedia information security, and in particular to a deep learning-based anti-screen-shooting robust watermark algorithm system. Background Art

[0002] In recent years, with the development of the Internet and the popularity of digital mobile devices, we live in an era of information explosion. People have become accustomed to carrying mobile phones as portable terminals for data processing when going out. With the camera function of smartphones, applications closely related to digital image processing technology such as scanning code to pay, scanning code to order, object identification, and logistics information query have been integrated into our lives and become an inseparable part. On the other hand, since the acquisition of digital media has become almost cost-free, the illegal acquisition, theft and tampering of digital media have become more and more serious, which not only brings troubles to the authors of media works, but sometimes even affects the reputation and security of the country. Therefore, copyright protection technology is urgently needed.

[0003] Scanning printed images and photographing printed or displayed images are common image recovery methods in daily life. However, when images on the Internet undergo this recovery process, various complex noises are introduced, and the watermark information in the image is often lost, resulting in the inability to correctly extract information. In this case, how to design a robust watermark that can resist complex noise in the real world has become a major difficulty in the field of digital watermarking technology.

[0004] In recent years, deep convolutional neural networks have made great progress in the fields of image processing and image classification. Some scholars have also proposed the application of deep convolutional neural networks in the field of digital image information hiding, and have achieved excellent results. Compared with traditional methods, it has stronger robustness and can resist various noises while having good image quality. However, they usually encode and embed watermarks based on the carrier image, which is not very practical. This paper will take the anti-screen shooting algorithm based on deep convolutional neural networks as the research object, and explore how to use deep convolutional neural networks: 1. Build an encoding and extraction network for the anti-screen shooting process; 2. The encoding network only encodes secret information, and has nothing to do with the carrier image; 3. Reduce the distortion of the watermark image while ensuring robustness. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a robust watermarking algorithm based on deep learning that is more robust and practical while ensuring a certain image quality. The output is an information film containing only information, which can be attached to any image and has more practical application value. In various real screen shooting scenes, it can also achieve an almost 100% extraction accuracy. In order to achieve the above purpose and other advantages according to the present invention, a robust watermarking system based on deep learning is provided, including:

[0006] An encoder, a discriminator and a noise network module connected to the encoder signal, and a decoder connected to the noise network module signal;

[0007] The encoder is used to construct an information film with secret information to be embedded;

[0008] The discriminator is used to distinguish whether the image is encoded;

[0009] The noise network module is used to attack the encoded image;

[0010] The decoder is used to recover the embedded secret information using the attacked encoded image.

[0011] A deep learning-based robust watermarking algorithm for resisting screen-shooting includes the following steps:

[0012] S1. Input random binary information m0 into the encoder. The m0 is first processed into a 50*50*3 tensor through a linear fully connected layer, and then upsampled to a 400*400*3 tensor. This information preprocessing method helps its convergence.

[0013] S2. Construct a coding network. The function of the coding network is to generate an information film to be attached to the image to be encoded. Apply several convolutional layers and ReLU layers to downsample the tensor generated in step 1 four times to obtain a 50*50*3 tensor, and then upsample it four times to restore it to the original size, and output an information film with a size of 400*400*3;

[0014] S3, adding the image to be encoded and the information film generated in step S2 to obtain a watermark image;

[0015] S4, construct a noise network and use a series of mathematical models to simulate the distortion caused by physical imaging;

[0016] S5, inputting the watermark image generated in step S3 into the noise network constructed in step S4;

[0017] S6. Construct a decoding network, downsample the encoded image into a series of numbers through the decoder, and then divide it into binary;

[0018] S7. The noisy watermark image generated in step S5 is used as the input of the decoding network, and is fed into the decoding network for decoding to obtain the embedded watermark information.

[0019] Preferably, step S4 includes five types of noise, namely perspective transformation, motion blur, random discrete noise, color distortion, and JPEG compression noise, for simulating the noise generated in the actual screen shooting process.

[0020] Preferably, the perspective transformation is a process in which the camera is not aligned with the image during the screen shooting process, which may cause the image to be deformed. The present invention uses a random homography perspective transformation to map the original image and the deformed image, and projects an image to a new viewing plane. The general transformation formula is:

[0021]

[0022] Where (u, v) is the original image pixel coordinate, is the pixel coordinate of the image after transformation. During the training process, the perspective transformation degree of the four corners is controlled within 10% of the length and width of the image, and bilinear interpolation is used during the transformation to complete the area that needs to be filled. The image after perspective transformation is recorded as I a .

[0023] Preferably, the motion blur is caused by camera movement or inaccurate focus during the screen capture process, which may cause the captured image to be blurred. A value is set to control the variance of the blur kernel to generate a linear Gaussian blur kernel G with a width of 7 pixels, and the rotation angle is controlled between 0 and 2π. Blurred image I b It can be expressed by the following formula:

[0024]

[0025] Where (i, j) represents the image I after adding blur noise b The coordinates of , (m, n) represent the coordinates of the rotated Gaussian convolution kernel.

[0026] Preferably, the random discrete noise is various noises introduced when the camera system is imaging, such as photon noise, dark noise and shot noise, etc., which are generated by Gaussian noise (N g ) model, image I after adding noise c It can be expressed as:

[0027] I c =I b +N g .

[0028] Preferably, the color distortion is due to the limited color gamut of printers and displays compared to the entire RGB color space, and these distortions are approximated by a series of random color transformations, such as hue adjustment: color compensation is multiplied by I randomly set between -0.1 and 0.1 c The RGB components are generated, denoted as I d , Brightness adjustment: Use linear transformation to generate brightness adjustment value, the formula is as follows: I br .=I d .m+b, limit the value of b to between -0.3 and 0.3, and the value of m to between 0.5 and 1.5. Saturation adjustment: We first convert the image from the RGB color domain to the YUV color domain, where the Y component represents the brightness information of the image, denoted by I y , and then linearly combine the original image with the Y component.

[0029] The final image can be expressed as follows:

[0030] I e =(1+t).(I br +I d )+tI y .

[0031] Preferably, the JPEG compression noise is quantized by calculating a discrete cosine transform (DCT) on each 8×8 block of the image and rounding to the nearest integer, as shown in the following formula:

[0032]

[0033] Where q(x) is the result of JPEG compression, and x represents the pixel value of the point.

[0034] In step S7, the input of the decoding network is the noisy watermark image, and the watermark information can be correctly extracted with the corresponding BCH error correction code.

[0035] Compared with the prior art, the present invention has the following beneficial effects: using a deep learning algorithm, an end-to-end neural network architecture is constructed, which is driven by a large data image set and has better universality. The encoding network is used to generate an information film containing only watermark information, which can be attached to any image. Unlike other methods, the generation of the information film does not require the participation of the carrier image, so it has more practical application value. Then the information film is added to the image to be encoded to generate a watermark image. In order to improve the robustness of the watermark image in real situations, a noise network needs to be added during the training of the deep convolutional network so that the subsequent watermark extraction network can be robust to noise. We use a series of mathematical models to simulate the noise generated during the screen shooting process, and experiments have shown that watermark information can be transmitted more reliably. The main function of the decoding network is to extract the original information from the watermark image with noise added. In order to convert the two-dimensional watermark image into a one-dimensional string of 01 numbers, we use several layers of convolutional layers to reduce the image size, and then map it to one dimension through a fully connected layer. In order to resist the deformation that will occur in the real environment, a spatial transformation network (STN) is used before feeding the coded noise image into the decoder network. STN will perform an affine transformation on the image to correct the deformation. In addition, the introduction of residual map supervision based on image JND during the training process can further reduce the distortion of the watermark image after covering the information film. Therefore, compared with traditional watermarking algorithms and other watermarking algorithms based on deep learning, the present invention has stronger robustness and practicality while ensuring a certain image quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a coding network diagram of the deep learning-based anti-screen-shooting robust watermarking algorithm according to the present invention;

[0037] Figure 2 A decoding network diagram of the deep learning-based anti-screen-shooting robust watermarking algorithm according to the present invention;

[0038] Figure 3 It is an end-to-end network framework diagram used by the deep learning-based anti-screen-shooting robust watermarking algorithm according to the present invention;

[0039] Figure 4 A portion of the image to be embedded in the anti-screen-shooting robust watermarking algorithm based on deep learning according to the present invention;

[0040] Figure 5 It is an information film image containing embedded information of the anti-screen-shooting robust watermarking algorithm based on deep learning according to the present invention;

[0041] Figure 6 A watermark image of a part of the image additional information film to be embedded in the anti-screen-shooting robust watermark algorithm based on deep learning according to the present invention;

[0042] Figure 7 These are experimental pictures of screen shooting at different angles according to the deep learning-based anti-screen-shooting robust watermarking algorithm of the present invention. DETAILED DESCRIPTION

[0043] 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.

[0044] Reference Figure 1-7 , a deep learning-based anti-screen-shooting robust watermarking system, comprising: an encoder, a discriminator and a noise network module connected to the encoder signal, and a decoder connected to the noise network module signal;

[0045] The encoder is used to construct an information film with secret information to be embedded;

[0046] The discriminator is used to distinguish whether the image is encoded;

[0047] The noise network module is used to attack the encoded image;

[0048] The decoder is used to recover the embedded secret information using the attacked encoded image.

[0049] Example

[0050] Step 1: The random binary information m0 is used as the input of the encoder. It first passes through a linear fully connected layer and is processed into a 50*50*3 tensor, and then upsampled to a 400*400*3 tensor.

[0051] Step 2: Build the encoding network. Figure 1 As shown in the figure, several convolutional layers and ReLU layers are applied to downsample the tensor generated in step 1 four times to obtain a 50*50*3 tensor, and then upsample it four times to restore it to its original size. The output is an information film with a size of 400*400*3. Suppose the image to be encoded is I o , the information film is attached to I o Generate watermark image I en , convert both images from RGB color domain to YUV color domain, and use L2 loss function to evaluate the similarity of the two images:

[0052] L2(I o , I en )=||I o -I en || 2 / (C·H·W)

[0053] Where C·H·W is the size of the image. On this basis, the image perception loss function is introduced to further supervise the image quality, which can be expressed as LPIPS (I o , I en ). In order to minimize the distortion of the watermark image, a residual graph supervision function L based on image JND is added jnd , in summary, the multi-task loss function of the encoding network can be expressed as follows:

[0054] L en =k1·L2+k2·L jnd +k3·LPIPS(I o , I en )

[0055] Step 3: Build a noise network and use a series of mathematical models to simulate the printing and shooting imaging processes in the real physical world, including perspective transformation, motion blur, random discrete noise, color distortion, and JPEG compression.

[0056] Step 4: Build a decoding network, such as Figure 2 As shown in Figure 2. The role of the decoding network is to extract the information embedded in the watermark image. Let the original information be M0, and the watermark information extracted from the encoded watermark image after passing through the decoder be M de , use the cross entropy formula to calculate the difference between the two sets of information:

[0057] Where M0(n) represents the nth bit of the original watermark information, M de (n) represents the nth bit of the extracted information, and s represents the length of the information. The greater the information entropy, the greater the difference between the two pieces of information. Therefore, the goal of the network is to make the output value of the loss function as small as possible.

[0058] Step 5: The end-to-end network model is built, such as Figure 3 As shown. Set the number of information bits that can be embedded in the model to 32 bits, including 16 information bits (two characters) and 16 error correction bits. Use the COCO dataset as the image to be encoded during training, and adjust the image size to 400*400 for training;

[0059] Step 6: Use the trained model to embed watermarks. First, set the embedding information to xm and generate the corresponding information film, such as Figure 2 As shown, the watermark image covering the information film is obtained, such as Figure 3 As shown;

[0060] Step 7: Use the trained model to decode the watermark image. With the BCH (32, 16) error correction code, 3 bits of error can be corrected. Experiments have shown that the watermark information can be correctly restored. These 32 bits of information can form 2 32 Different kinds of information are sufficient for general applications.

[0061] According to the above steps, the experiment was conducted. For the quality of the watermarked image, the peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and image perceptual similarity (LPIPS) were used as measurement indicators. The larger the PSNR and SSIM indicators, the smaller the distortion of the image compared to the original carrier image. The LPIPS indicator is the perceptual similarity of the image. The smaller the value, the better the visual effect, which can better reflect the visual perception of the human eye. During the experiment, we used 60 original images, such as Figure 4 As shown, the information film with embedded information "xm" is covered and generated, such as Figure 5 As shown, the corresponding watermark image is obtained, such as Figure 6 As shown, their PSNR, SSIM, and LPIPS are calculated respectively, and the average is taken. The results are shown in Table 1:

[0062] Table 1 Watermark image quality

[0063] index value PSNR 31.825145 SSIM 0.952779 LPIPS 0.014151

[0064] For the robustness test in real environment and digital environment, two indicators are designed to measure, one is the bit error rate after direct decoding:

[0065]

[0066] Among them, N bc Represents the number of correct bits in each decoded bit string, L bs is the length of each bit string

[0067] The other is the average accuracy ε2 after error correction coding restoration:

[0068]

[0069] Where N represents the number of test images, N c Indicates the number of images from which watermark information is correctly extracted.

[0070] The robustness experiment in the digital environment can reflect the resistance of the deep learning model to noise to a certain extent. In this experiment, we selected five types of noise that may appear in the process of shooting the screen, including: Gaussian noise, scaling noise, edge coverage, center cropping, and JPEG compression. For Gaussian noise attacks, the size of the Gaussian kernel directly determines the degree of damage to the image, so in the experiment, Gaussian kernels with lengths and widths of 3, 5, 7, 9, 11, and 23 were selected as convolution kernels when adding noise operations. For image scaling attacks, the image size determines the strength of the attack. Since the network output is a 400*400 three-channel image, in order to prevent decimals in the length and width after scaling, in the experiment, 4 / 5, 3 / 4, 3 / 5, 1 / 2, 2 / 5, 1 / 4, and 1 / 5 were selected as the scaling size (that is, the image size after scaling is 320*320, 300*300, 240*240, 200*200, 160*160, 100×100, and 80×80). For edge coverage attacks, in the experiment, we use a border around the image to add edge occlusion. The pixel value RGB used for occlusion is (0, 0, 0), and the thickness of the frame is 10, 20, 30, 40, and 50 pixels respectively. For center cropping attacks, we use black square areas of different sizes to cover the center of the watermark image. In the experiment, we selected black pixel blocks of 20×20, 40×40, 60×60, 80×80, 100×100 and 120×120 to cover the center of the image embedded with the watermark. For JPEG compression, the parameter that affects the quality of the compressed image is the quality factor. In the experiment, we selected 40, 50, 60, 70, 80 and 90 as the JPEG quality factor values. We conducted robustness tests in a digital environment under the above five attacks on 59 watermark images covered with embedded information "xm". The results of the average bit error rate ε1 of direct decoding and the average accuracy rate ε2 after error correction coding restoration are shown in Tables 2, 3, 4, 5 and 6:

[0071] Table 2 Average bit error rate ε1 and average accuracy ε2 of the model under Gaussian blur kernel attacks of different sizes

[0072] Gaussian Kernel Average bit error rate ε1 Average accuracy ε2 3 0.003939 100％ 5 0.007878 99.1597％ 7 0.016817 95.7983％ 9 0.031513 89.916％ 11 0.056197 76.4706％

[0073] Table 3 Average bit error rate ε1 and average accuracy ε2 of the model under different size scaling attacks

[0074] Zoom Average bit error rate ε1 Average accuracy ε2 320*320 0.002889 100％ 280*280 0.003676 100％ 240*240 0.003151 100％ 200*200 0.003676 100％ 160*160 0.004464 100％ 120*120 0.006828 100％ 80*80 0.012605 99.1597％

[0075] Table 4 Average bit error rate ε1 and average accuracy ε2 of the model under edge coverage attacks of different sizes

[0076]

[0077]

[0078] Table 5 Average bit error rate ε1 and average accuracy ε2 of the model under different size center cropping attacks

[0079] Center Crop Average bit error rate ε1 Average accuracy ε2 20 0.002626 100％ 40 0.002889 100％ 60 0.004989 100％ 80 0.005252 100％ 100 0.007090 98.3193％ 120 0.008403 99.1597％

[0080] Table 6 Average bit error rate ε1 and average accuracy ε2 of the model under different JPEG compression factor attacks

[0081] JPEG compression Average bit error rate ε1 Average accuracy ε2 40 0.004202 100％ 50 0.004464 100％ 60 0.003676 100％ 70 0.004727 100％ 80 0.002889 99.1597％ 90 0.002889 100％

[0082] Robustness experiment in real environment. The watermark image used in the experiment is obtained by shooting the real screen. The real screen shooting process will produce more complex noise. The experimental results can better prove the robustness of the model. In order to make the experimental results more convincing, we divided the experiment into four groups, namely: shooting at different distances, shooting at different angles, shooting under different lighting conditions, and shooting with different devices. For the screen shooting experiment at different distances, we used iPhone xs under 150lux lighting conditions to shoot three groups of watermark images, 20 images in each group, respectively: 18cm, 35cm and 55cm. For the screen shooting experiment at different angles, such as Figure 7 As shown in the figure, we fixed the shooting distance at 35cm each time under the lighting condition of 150lux, and used iPhonexs to shoot six groups of watermark images, 20 images in each group, namely: 15 degrees to the left, 30 degrees to the left, 45 degrees to the left, 15 degrees to the right, 30 degrees to the right, and 45 degrees to the right. For the screen shooting experiment under different lighting conditions, we fixed the shooting distance at 35cm each time, and used iPhonexs to shoot three groups of watermark images, namely 50lux, 150lux and 500lux. For the screen shooting experiment under different shooting devices, we fixed the shooting distance at 35cm each time under the lighting condition of 15lux, and used iPhonexs, HUAWEIP40 Pro+, and Xiaomi 11 to shoot three groups of watermark images, 20 images in each group. We conducted robustness tests in real environments under the above four conditions for 20 watermark images covered with embedded information "xm". The results of the average bit error rate ε1 of direct decoding and the average accuracy ε2 after error correction coding restoration are shown in Tables 7, 8, 9, and 10:

[0083] Table 7 Average bit error rate ε1 and average accuracy ε2 of the model after actually shooting the screen at different distances

[0084] Shooting distance Average bit error rate ε1 Average accuracy ε2 18cm 0.0516 85％ 35cm 0.0125 100％ 55cm 0.0312 95％

[0085] Table 8 Average bit error rate ε1 and average accuracy ε2 of the model after actually shooting the screen at different angles

[0086] Shooting Angle Average bit error rate ε1 Average accuracy ε2 15 degrees left 0.0203 100％ 30 degrees left 0.0219 100％ 45 degrees left 0.0203 100％ 15 degrees right 0.0016 100％ 30 degrees right 0.0016 100％ 45 degrees to the right 0.0078 100％

[0087] Table 9 Average bit error rate ε1 and average accuracy ε2 of the model after actually shooting the screen under different lighting conditions

[0088] Lighting conditions Average bit error rate ε1 Average accuracy ε2 50lux 0.0187 100％ 150lux 0.0125 100％ 500lux 0.0187 100％

[0089] Table 10 Average bit error rate ε1 and average accuracy ε2 of the model after actually shooting the screen under different shooting devices

[0090] Filming equipment Average bit error rate ε1 Average accuracy ε2 iPhone xs 0.0125 100％ Huawei P40 Pro 0.0109 100％ Xiaomi Mi 11 0.0250 100％

[0091] Functions and Effects of the Embodiments

[0092] According to the deep learning-based anti-screen shooting robust watermarking algorithm involved in this embodiment, the encoding network is used to generate an information film containing only information, the information film is superimposed on any image to generate a watermark image, the noise layer simulates the noise that may be generated during the screen shooting process, and the decoding network is used to decode the watermark image after the noise layer. This embodiment conducts robustness experiments on the model in a digital environment and a robustness experiment in a real environment. The experimental results show that compared with the traditional anti-screen shooting method based on image processing, the anti-screen shooting watermarking algorithm based on deep neural network training has better performance in indicators such as information robustness and watermark image visual quality, and can also achieve nearly 100% decoding accuracy in various complex real shooting screen situations. In addition, the present invention is different from other anti-screen shooting methods based on deep learning in that the generation of the information film only requires the participation of embedded information, which means that it can be superimposed on any original image, so the method proposed in the present invention has more practical application value.

[0093] The number of devices and processing scales described here are used to simplify the description of the present invention, and the application, modification and variation of the present invention will be obvious to those skilled in the art.

[0094] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A deep learning-based robust watermarking algorithm for resisting screen snapping, including a robust watermarking system for resisting screen snapping, the system including an encoder, a discriminator and a noise network module connected to the encoder signal, and a decoder connected to the noise network module signal; The encoder is used to construct an information film with secret information to be embedded; The discriminator is used to distinguish whether the image is encoded; The noise network module is used to attack the encoded image; The decoder is used to recover embedded secret information using the attacked coded image; characterized in that, The algorithm specifically includes the following steps: S1, random binary information Input encoder, the First, it passes through a linear fully connected layer and is processed into a 50*50*3 tensor, and then upsampled to a 400*400*3 tensor; S2, constructing a coding network, the role of the coding network is to generate an information film for attaching to the image to be encoded; S3, adding the image to be encoded and the information film generated in step S2 to obtain a watermark image; S4, construct a noise network and use a series of mathematical models to simulate the distortion caused by physical imaging; S5, inputting the watermark image generated in step S3 into the noise network constructed in step S4; S6. Construct a decoding network, downsample the encoded image into a series of numbers through the decoder, and then divide it into binary; S7. The noisy watermark image generated in step S5 is used as the input of the decoding network, and is fed into the decoding network for decoding to obtain the embedded watermark information.

2. The deep learning-based robust watermarking algorithm for screen-snap protection as claimed in claim 1, characterized in that: The step S4 includes five types of noise, namely perspective transformation, motion blur, random discrete noise, color distortion, and JPEG compression noise, which are used to simulate the noise generated in the process of actually shooting the screen.

3. The deep learning-based robust watermarking algorithm for screen-snap protection as claimed in claim 2, characterized in that: The perspective transformation is the process of projecting an image onto a new viewing plane when the camera is not aligned with the image during the screen capture process. The general transformation formula is: ； in is the original image pixel coordinate, is the pixel coordinate of the image after transformation. During the training process, the perspective transformation degree of the four corners is controlled within 10% of the length and width of the image, and bilinear interpolation is used during the transformation to complete the area that needs to be filled. The image after perspective transformation is recorded as .

4. The deep learning-based robust watermarking algorithm for resisting screen-shooting as claimed in claim 3, characterized in that: The motion blur is caused by camera movement or inaccurate focus during the screen capture process, which may lead to blurry images. A value is set to control the variance of the blur kernel to generate a linear Gaussian blur kernel G with a width of 7 pixels, and the rotation angle is controlled between 0 and 2π; Blurred Image It can be expressed by the following formula: ； in Represents the image after adding blur noise The coordinates of Represents the coordinates of the rotated Gaussian convolution kernel.

5. The deep learning-based robust watermarking algorithm for screen-snap protection as claimed in claim 4, characterized in that: The random discrete noise is various noises introduced when the camera system is imaging, such as photon noise, dark noise and shot noise, etc. Model, image after adding noise It can be expressed as: 。 6. The deep learning-based robust watermarking algorithm for resisting screen-shooting as claimed in claim 5, characterized in that: The color distortion is due to the limited color gamut of printers and monitors compared to the entire RGB color space. These distortions are approximated by a series of random color transformations, such as hue adjustment: color compensation is multiplied by a parameter randomly set between -0.1 and 0.1 The RGB components are generated, denoted as ,Brightness adjustment: Use linear transformation to generate the brightness adjustment value, the formula is as follows: ,Will The value of is limited to between -0.3 and 0.

3. The value of is limited to between 0.5 and 1.

5. Saturation adjustment: We first convert the image from the RGB color domain to the YUV color domain, where the Y component represents the brightness information of the image, recorded as , and then linearly combine the original image with the Y component; The final image can be expressed as follows: 。 7. The deep learning-based robust watermarking algorithm for resisting screen-shooting as claimed in claim 6, characterized in that: The JPEG compression noise is quantized by computing the discrete cosine transform (DCT) on each 8×8 block of the image and rounding to the nearest integer, as shown in the following formula: ； in It is the result of JPEG compression. Indicates the pixel value of this point.

Images