Multistage video watermark embedding and extracting method and system

Abstract

The invention discloses a multistage video watermark embedding and extracting method and system, and the method comprises the steps: carrying out the perception analysis of an original video through a space-time attention neural network, and dynamically dividing a time domain stable region and a space domain sensitive region; a three-level watermark system is constructed, the three-level watermark system comprises a basic-level watermark, a first-level enhanced watermark and a second-level enhanced watermark, the basic-level watermark adopts an anti-rotation binary matrix generated by a Fourier descriptor, the first-level enhanced watermark is a DCT frequency domain Hash sequence, and the second-level enhanced watermark is a semantic fingerprint generated based on a motion vector. According to the method, through multi-dimensional collaborative optimization of the anti-attack capability, the dynamic security mechanism and the verification efficiency, the core defects that erasing is easy due to insufficient reversibility in a traditional video watermarking technology and counterfeit affiliation chaos is caused by a static key are overcome, and reliable technical guarantee is provided for digital content copyright protection.

Description

Technical Field

[0001] This application relates to the field of watermark embedding technology, and particularly to a method and system for multi-level video watermark embedding and extraction. Background Art

[0002] With the explosive growth of the digital content industry, video, as the core carrier of information dissemination, faces severe challenges in copyright protection.

[0003] Traditional video watermarking technologies mostly adopt a single embedding strategy, making it difficult to balance the contradiction between robustness and invisibility: To resist geometric attacks (such as rotation, scaling) or signal processing attacks (such as compression, noise), it is often necessary to embed high-strength watermarks in the low-frequency components, resulting in visual artifacts; when choosing high-frequency embedding to maintain invisibility, the watermarks are easily destroyed by conventional editing operations. This contradiction directly leads to two major hazards: First, the lack of reversibility of watermarks enables attackers to locate and erase the watermarks, rendering the copyright identifiers ineffective; second, the static key mechanism is vulnerable to cracking or replay attacks, and the injection of forged watermarks further exacerbates the confusion of copyright attribution; pirated videos can strip the watermarks through simple filtering or re-encoding, and the number of illegal rights confirmation events caused by key leakage increases year by year, seriously threatening the rights and interests of creators and the security of the digital content ecosystem. Summary of the Invention

[0004] To solve the above problems, embodiments of the present invention provide a method for multi-level video watermark embedding and extraction, and the method includes:

[0005] Perceptually analyze the original video through a spatio-temporal attention neural network, and dynamically divide the time-domain stable region and the space-domain sensitive region;

[0006] Construct a three-level watermark system, which includes a basic-level watermark, a first-level enhanced watermark, and a second-level enhanced watermark. The basic-level watermark is an anti-rotation binary matrix generated by Fourier descriptors, the first-level enhanced watermark is a DCT frequency-domain hash sequence, and the second-level enhanced watermark is a semantic fingerprint generated based on motion vectors;

[0007] In the time-domain stable region, embed the basic-level watermark into the YUV low-frequency components of I-frames through motion-compensated frame-interframe prediction; in the space-domain sensitive region, utilize a generative adversarial network to simulate the human visual masking effect, and embed the two-level enhanced watermarks into the DCT high-frequency subbands and motion vector residuals of P / B-frames respectively;

[0008] Generate a dynamic encryption key based on video content hashing and blockchain smart contracts, and adopt a double encryption mechanism of elliptic curve cryptography and chaotic scrambling to generate an irreversible watermark identifier;

[0009] When extracting the watermark, correct geometric distortion through a spatio-temporal registration network, first verify the integrity of the basic-level watermark stored on the blockchain, and then gradually decode the enhanced watermarks through a cascaded neural network and perform cross-verification.

[0010] Furthermore, the spatio-temporal attention neural network includes a time-domain branch, a space-domain branch, and a fusion module:

[0011] Time-domain branch: 3D convolution is used to extract the motion consistency feature map to identify the shot boundary and the background stable region;

[0012] Space-domain branch: The Sobel edge detection and HSV color space analysis are used to determine the texture complex region;

[0013] Fusion module: The gating mechanism is used to dynamically generate the probability heat map of the watermark embedding region.

[0014] Furthermore, the adversarial generative network training includes a generator and a discriminator:

[0015] Generator: Learn to inject invisible watermarks into the DCT high-frequency coefficients;

[0016] Discriminator: Combine the human visual system model and the image quality evaluation index to construct an adversarial loss function.

[0017] Furthermore, the output of the discriminator includes an adversarial loss and a quality evaluation loss:

[0018] Adversarial loss L adv : Based on binary cross-entropy, judge whether the input image is the original unmodified image;

[0019] Quality evaluation loss L IQA : Calculate the visual fidelity of the generated image through the MS-SSIM and the perceptual hash difference.

[0020] Furthermore, the dual encryption mechanism specifically includes:

[0021] Grayscale and size normalization are performed on each frame of the input video, and the discrete cosine transform is used to extract the low-frequency components to generate a 64-bit PHash value;

[0022] The 64-bit PHash value is truncated into a 256-bit binary sequence, and the final private key seed is generated through the HMAC-SHA256 algorithm. Based on the elliptic curve secp256k1 parameters, the final private key seed is mapped to a legal private key;

[0023] The improved Logistic-Tent hyperchaotic system is described using a difference equation, and the difference equation includes:

[0024] where r, μ, and γ are chaotic parameters, and mod1 represents the modulo operation on the numerical value with a modulus of 1;

[0025] Using the chaotic sequence {x n} Control the row cyclic shift amount of the watermark matrix, where the row cyclic shift amount is {x n ·10 4} mod the number of rows; Generate a dynamic mask based on the chaotic sequence {y n} and perform a bitwise exclusive OR operation on each pixel of the watermark matrix;

[0026] Deploy a smart contract on the Ethereum test chain and write the hash value of the encrypted parameter to the blockchain as a verification anchor point.

[0027] Further, the geometric distortion correction includes:

[0028] Construct a spatio-temporal registration network for adversarial sample training that includes rotation, scaling, and shearing;

[0029] Estimate the affine transformation parameters through a deformable convolutional layer and use bilinear interpolation to restore the spatio-temporal alignment of the video.

[0030] Further, a two-level verification mechanism is adopted during verification when extracting the watermark. The two-level verification mechanism includes primary verification and ultimate verification;

[0031] Primary verification: Reconstruct the decryption matrix through the encrypted parameters stored in the blockchain;

[0032] Ultimate verification: Compare the Pearson correlation coefficient between the DCT domain hash sequence and the motion vector semantic fingerprint.

[0033] Further, the construction method of the motion vector semantic fingerprint includes:

[0034] Extract the motion vectors of the P / B frames of the video and statistically analyze their direction distribution and amplitude distribution;

[0035] Map the motion vector features to a fixed-length fingerprint vector through semantic encoding.

[0036] A multi-level video watermark embedding and extraction system, the system includes:

[0037] A region division module, which dynamically divides the time-domain stable region and the space-domain sensitive region by performing perceptual analysis on the original video through a spatio-temporal attention neural network;

[0038] A watermark construction module, which constructs a three-level watermark system. The three-level watermark system includes a basic-level watermark, a first-level enhanced watermark, and a second-level enhanced watermark. The basic-level watermark is an anti-rotation binary matrix generated by Fourier descriptors, the first-level enhanced watermark is a DCT domain hash sequence, and the second-level enhanced watermark is a semantic fingerprint generated based on motion vectors;

[0039] The watermark embedding module embeds the basic-level watermark into the low-frequency component of the YUV of the I-frame through motion compensation inter-frame prediction in the time-domain stable region; in the spatially sensitive region, the adversarial generative network is used to simulate the human eye visual masking effect, and the two-level enhanced watermarks are respectively embedded into the DCT high-frequency subband and the motion vector residual of the P / B-frame;

[0040] The watermark identification module generates a dynamic encryption key based on the video content hash and the blockchain smart contract, and generates an irreversible watermark identification by adopting a dual encryption mechanism of elliptic curve cryptography and chaotic scrambling;

[0041] The watermark verification module corrects geometric distortion through the spatio-temporal registration network during extraction, first verifies the integrity of the basic-level watermark stored on the blockchain, and then gradually decodes the enhanced watermark through the cascaded neural network and performs cross-verification.

[0042] The technical effects and advantages of the multi-level video watermark embedding and extraction method and system provided by the present invention:

[0043] The present invention solves the core defects of traditional video watermarking technology, such as being easily erased due to insufficient reversibility and the confusion of forged attribution caused by static keys, through multi-dimensional collaborative optimization of anti-attack ability, dynamic security mechanism and verification efficiency, and provides a reliable technical guarantee for digital content copyright protection. The present invention dynamically divides the watermark embedding area based on the spatio-temporal attention mechanism, combines the hierarchical embedding strategy of the basic-level and enhanced-level watermarks, and realizes adaptive defense against different attack types such as geometric attacks, compression noise, and clip tampering; the basic watermark resists geometric deformation through anti-rotation design, and the enhanced-level watermark copes with content tampering through high-frequency embedding and motion vector binding, forming a multi-level protection; by simulating the human eye visual characteristics through the adversarial generative network, the watermark embedding strength is dynamically adjusted in complex texture areas, significantly reducing the visual perceptibility, ensuring the invisibility of the watermark while maintaining the video quality; based on the video content hash and the chaotic encryption system to generate dynamic keys, combined with the blockchain storage technology, the immutability and uniqueness of the watermark parameters are realized, effectively resisting key cracking and replay attacks, and during watermark extraction, through blockchain verification and multi-feature cross-checking (DCT hash and motion vector fingerprint), the dual verification mechanism greatly reduces the false detection rate and ensures that forged content cannot pass the legality verification. Description of the Drawings

[0044] Figure 1 It is the flow chart of the multi-level video watermark embedding and extraction method in Embodiment 1;

[0045] Figure 2 It is the schematic connection diagram of the spatio-temporal attention neural network in Embodiment 1;

[0046] Figure 3 It is the schematic connection diagram of the multi-level video watermark embedding and extraction system in Embodiment 2. Detailed implementation mode

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] Embodiment 1:

[0049] Please refer to Figure 1 As shown, the multi-level video watermark embedding and extraction method in this embodiment includes:

[0050] Perceptually analyze the original video through a spatio-temporal attention neural network, and dynamically divide the time-domain stable region and the space-domain sensitive region;

[0051] Construct a three-level watermark system, which includes a basic-level watermark, a first-level enhanced watermark, and a second-level enhanced watermark. The basic-level watermark is an anti-rotation binary matrix generated by Fourier descriptors, the first-level enhanced watermark is a DCT frequency-domain hash sequence, and the second-level enhanced watermark is a semantic fingerprint generated based on motion vectors;

[0052] In the time-domain stable region, embed the basic-level watermark into the YUV low-frequency component of the I frame through motion-compensated inter-frame prediction; in the space-domain sensitive region, utilize the generative adversarial network to simulate the human eye visual masking effect, and embed the two-level enhanced watermarks into the DCT high-frequency subbands and motion vector residuals of the P / B frames respectively;

[0053] Generate a dynamic encryption key based on video content hashing and blockchain smart contracts, and adopt a double encryption mechanism of elliptic curve cryptography and chaotic scrambling to generate an irreversible watermark identifier;

[0054] During extraction, correct geometric distortion through a spatio-temporal registration network, first verify the integrity of the basic-level watermark stored in the blockchain, and then gradually decode the enhanced watermark through a cascaded neural network and perform cross-verification.

[0055] The time-domain stable region is for basic-level watermark embedding, and the space-domain sensitive region is for enhanced-level watermark embedding. The enhanced-level watermark includes a first-level enhanced watermark and a second-level enhanced watermark.

[0056] As Figure 2 shown, the spatio-temporal attention neural network includes a time-domain branch, a space-domain branch, and a fusion module:

[0057] Time-domain branch: Use 3D convolution to extract the motion consistency feature map and identify the shot boundary and the background stable region;

[0058] The 3D convolutional architecture uses a three-dimensional convolutional kernel (5×5×3) that is continuous across frames to extract temporal motion features in the RGB channel dimension. After the convolutional layer, a ReLU activation function and a max-pooling layer are connected to form a spatio-temporal feature pyramid. The modeling of the motion consistency feature map can be achieved by calculating the cosine similarity matrix of adjacent frame feature maps, constructing a motion vector field, and using the non-maximum suppression algorithm to detect the boundaries of abrupt camera cuts. The method for identifying the background stable region includes performing temporal median filtering on the feature map, extracting the region where the proportion of low-frequency energy exceeds a set threshold as the candidate background region, and eliminating holes through morphological closing operations.

[0059] Spatial domain branch: Determine the texture complex region through Sobel edge detection and HSV color space analysis, thereby generating the texture complexity map (S_map) of the spatial domain branch.

[0060] For the detection of multi-scale edges, the Sobel operator is used to calculate the gradients in the horizontal, vertical, and diagonal directions respectively, and the anisotropic edge response maps are fused to generate the initial texture distribution. The analysis of the HSV color space includes converting the RGB frame to the HSV color space and constructing texture complexity metrics in the following dimensions:

[0061] Hue channel: Calculate the local color variance to reflect the intensity of hue change.

[0062] Saturation channel: Use a sliding window to count the aggregation degree of high-saturation pixels.

[0063] Value channel: Apply the Laplacian operator to detect regions of abrupt brightness change.

[0064] The determination of texture composition is achieved by weighted fusion of the three elements of edge density, color variance, and brightness gradient to generate a 0-1 normalized spatial domain texture complexity map to determine the texture complex region.

[0065] Fusion module: Dynamically generate a probability heat map of the watermark embedding region using a gating mechanism.

[0066] The method for generating the gating weight matrix in the gating mechanism includes: inputting the motion consistency feature map (T_map) of the spatial domain branch and the texture complexity map (S_map) of the spatial domain branch into a two-channel gating network, and generating the gating weight matrix through the sigmoid function.

[0067] The feature interaction mechanism in the gating mechanism includes:

[0068] Spatial attention: Use a channel-wise spatial transformer to perform an affine transformation on T_map for alignment.

[0069] Cross-domain feature concatenation: Concatenate the aligned T_map and S_map along the channel dimension.

[0070] Adaptive filtering: A 1×1 convolutional kernel is applied to achieve dimensionality reduction and information screening of feature channels.

[0071] The method for generating the probability heat map includes: bilinearly interpolating the low-resolution fused T_map and S_map, upsampling to the original video resolution, and using the softmax function to generate the pixel-level watermark embedding probability distribution. Dynamically divide the watermark embedding area based on the spatio-temporal attention mechanism, and combine the hierarchical embedding strategy of the basic-level and enhanced-level watermarks to achieve adaptive defense against different attack types such as geometric attacks, compression noise, and clip tampering; the basic watermark resists geometric deformation through anti-rotation design, and the enhanced-level watermark copes with content tampering through high-frequency embedding and motion vector binding, forming a multi-level protection.

[0072] The training of the generative adversarial network includes a generator and a discriminator:

[0073] Generator: Learn to inject invisible watermarks into the DCT high-frequency coefficients;

[0074] The learning method includes:

[0075] The generator receives the original image (the original video is composed of the original images, and one frame of the original video is the original image) and the watermark information to be embedded. The original image is divided into blocks (e.g., 8×8 pixel blocks) and transformed to the frequency domain through the discrete cosine transform, separating the DCT low-frequency coefficients and the DCT high-frequency coefficients.

[0076] The generator learns to embed watermarks in the DCT high-frequency coefficients through a neural network, specifically including:

[0077] Predict the modification amount of the DCT high-frequency coefficients through the convolutional layer, so that the watermark is integrated into the image texture features;

[0078] Introduce noise interference or compression perturbation during training to enhance the anti-attack ability of the watermark in the DCT high-frequency coefficients;

[0079] The modified DCT high-frequency coefficients and the DCT low-frequency coefficients are combined and reconstructed through the inverse discrete cosine transform into a generated image with an invisible watermark.

[0080] Discriminator: Construct an adversarial loss function by combining the human visual system model and the image quality evaluation index.

[0081] The modeling of the human visual system model includes contrast sensitivity function filtering, brightness masking effect constraint, and frequency domain attention mechanism;

[0082] Contrast sensitivity function filtering: Simulate the insensitivity of the human eye to the DCT high-frequency coefficients, and reduce the response weight to the DCT high-frequency region in the discriminator;

[0083] Brightness masking effect constraint: Dynamically adjust the tolerance threshold for the modification amplitude of the DCT high-frequency coefficients according to the average brightness of the image block;

[0084] Frequency domain attention mechanism: In the discriminator feature extraction layer, assign higher attention weights to the DCT low-frequency coefficients to enhance the detection of structural distortions in the image.

[0085] The fusion of image quality assessment metrics includes multi-scale structural similarity and perceptual hash alignment;

[0086] Multi-scale structural similarity: Calculate the structural similarity loss between the generated image and the original image on the DCT low-frequency coefficients and DCT high-frequency coefficients;

[0087] Perceptual hash alignment: Constrain the statistical distribution of the DCT high-frequency coefficients of the generated image to be consistent with the DCT high-frequency coefficient distribution of the natural image library.

[0088] Simulate the human eye visual characteristics through the adversarial generation network, dynamically adjust the watermark embedding strength in the complex texture area, significantly reduce the visual perceptibility, and ensure the invisibility of the watermark while maintaining the video quality.

[0089] The output of the discriminator includes adversarial loss and quality assessment loss:

[0090] Adversarial loss L adv : Based on binary cross-entropy, judge whether the input image is an original unmodified image;

[0091] Quality assessment loss L IQA (such as the SSIM value): Calculate the visual fidelity of the generated image through the MS-SSIM and perceptual hash differences;

[0092] Therefore, the total loss function L D is: L D = L adv + λ · L IQA , where λ is the balance coefficient.

[0093] The double encryption mechanism specifically includes:

[0094] Grayscale and size normalize each frame of the input video (for example, 32×32 pixels), and use the discrete cosine transform (DCT) to extract the low-frequency components to generate a 64-bit PHash value;

[0095] Truncate the 64-bit PHash value to a 256-bit binary sequence, generate the final private key seed through the HMAC-SHA256 algorithm, and map the final private key seed to a legal private key based on the elliptic curve secp256k1 parameters.

[0096] The improved Logistic-Tent hyperchaotic system is used for cyclic shift and exclusive-or confusion of the watermark matrix;

[0097] Specifically, it includes:

[0098] The improved Logistic-Tent hyperchaotic system is described using a difference equation, and the difference equation includes:

[0099] where r, μ, and γ are chaotic parameters, mod1 represents the modulo operation on the value, and the modulus is 1;

[0100] The row cyclic shift amount of the watermark matrix is controlled using the chaotic sequence {x n}, and the row cyclic shift amount is {x n ·10 4} mod the number of rows; a dynamic mask is generated based on the chaotic sequence {y n}, and a bitwise exclusive-or operation is performed pixel by pixel on the watermark matrix;

[0101] Every time a preset number of iterations is reached, r, μ, and γ are dynamically fine-tuned through the low-order bytes of the PHash value to prevent the chaotic system from falling into a periodic state.

[0102] Deploy a smart contract on the Ethereum test chain (such as Rinkeby or Goerli), and write the hash value of the encryption parameter into the blockchain as a verification anchor point. During watermark extraction, through blockchain verification and multi-feature cross-verification (DCT hash and motion vector fingerprint), the double verification mechanism significantly reduces the false detection rate and ensures that forged content cannot pass the legality verification.

[0103] Geometric distortion correction includes:

[0104] Construct a spatio-temporal registration network for adversarial sample training that includes rotation, scaling, and shearing;

[0105] Apply a geometric transformation with affine transformation parameters to the input video frame to generate diverse distorted samples. Through the adversarial game between the generator (spatio-temporal registration network) and the discriminator, optimize the affine transformation parameter estimation ability of the generator. The output of the generator is the corrected frame, and the discriminator judges the distribution difference between it and the true aligned frame.

[0106] Estimate the affine transformation parameters through a deformable convolutional layer, and use bilinear interpolation to restore the spatio-temporal alignment of the video;

[0107] The deformable convolutional layer learns local geometric deformation features by dynamically offsetting sampling points and outputs a feature map. It estimates the affine transformation parameters using the deformable convolutional layer, obtains an affine matrix composed of the affine transformation parameters, and maps each pixel point of the corrected image (a frame of the video) back to the coordinate system of the original distorted image. The weighted average of the four integer pixel points around the non-integer coordinates is calculated, and the weight is determined by the fractional part of the coordinates. Among them, if the affine matrix is irreversible, a regularization term (such as determinant constraint) needs to be added to ensure numerical stability. The deformable convolution and the adversarial training network are used to automatically correct complex geometric distortions such as rotation and scaling, improving the stability and accuracy of watermark extraction.

[0108] During the verification of watermark extraction, a two-level verification mechanism is adopted, which includes primary verification and ultimate verification.

[0109] Primary verification: Reconstruct the decryption matrix through the encrypted parameters stored in the blockchain.

[0110] When embedding a watermark in a video, a set of encrypted parameters (such as symmetric keys, initialization vectors, and matrix generation seeds) are generated, and their hash values and timestamps are written into the blockchain (such as Ethereum and Hyperledger) to ensure the data cannot be tampered with. When extracting the watermark, the original encrypted parameters can be queried through the blockchain transaction ID or smart contract interface to verify their integrity and authenticity (such as comparing the hash value on the chain with the local parameter hash). The preliminary decryption of the watermark uses the decryption matrix to perform inverse transformation on the encrypted area of the watermarked video (such as DCT low-frequency coefficients or motion vector differences).

[0111] Ultimate verification: Compare the Pearson correlation coefficient between the DCT domain hash sequence and the motion vector semantic fingerprint.

[0112] The generation method of the DCT domain hash sequence includes:

[0113] The key frames (I-frames) of the video are divided into blocks, the DCT DC components and low-frequency AC components of each block are extracted, the DCT DC components and low-frequency AC components are quantized, and at the same time, binarization is performed to generate a binary hash sequence.

[0114] The construction method of the motion vector semantic fingerprint includes:

[0115] Extract the motion vectors of the P / B frames of the video and statistically analyze their direction distribution (such as horizontal and vertical component histograms) and amplitude distribution.

[0116] Map the motion vector features to a fixed-length fingerprint vector through semantic encoding (such as Huffman coding or Word2Vec).

[0117] The multi - stage cascade decoding and multi - stage verification mechanism optimize the extraction process, significantly reducing the computational complexity while ensuring robustness, and supporting real - time or near - real - time watermark extraction; the lightweight blockchain verification design reduces the consumption of storage and computing resources, supporting efficient deployment in large - scale copyright management scenarios.

[0118] Embodiment 2:

[0119] As Figure 3 shown, based on the same inventive concept as the multi - stage video watermark embedding and extraction method in the foregoing embodiment, this application provides a multi - stage video watermark embedding and extraction system. The system in the embodiment of this application and the method embodiment are based on the same inventive concept. Among them, the system includes:

[0120] A region division module that performs perceptual analysis on the original video through a spatio - temporal attention neural network to dynamically divide the time - domain stable region and the space - domain sensitive region;

[0121] A watermark construction module that constructs a three - level watermark system. The three - level watermark system includes a basic - level watermark, a first - level enhanced watermark, and a second - level enhanced watermark. The basic - level watermark is an anti - rotation binary matrix generated by Fourier descriptors, the first - level enhanced watermark is a DCT frequency - domain hash sequence, and the second - level enhanced watermark is a semantic fingerprint generated based on motion vectors;

[0122] A watermark embedding module that, in the time - domain stable region, embeds the basic - level watermark into the YUV low - frequency component of the I - frame through motion - compensated inter - frame prediction; in the space - domain sensitive region, uses a generative adversarial network to simulate the human visual masking effect and embeds the two - level enhanced watermarks into the DCT high - frequency sub - band and the motion vector residual of the P / B - frame respectively;

[0123] A watermark identification module that generates a dynamic encryption key based on video content hashing and blockchain smart contracts, and generates an irreversible watermark identifier using a dual - encryption mechanism of elliptic curve cryptography and chaotic scrambling;

[0124] A watermark verification module that corrects geometric distortion through a spatio - temporal registration network during extraction, first verifies the integrity of the basic - level watermark stored on the blockchain, and then gradually decodes the enhanced watermarks through a cascade neural network and performs cross - verification.

[0125] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention also intends to include these changes and modifications.

[0126] The above are only the preferred specific embodiments of the present application. However, the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application, according to the technical solution and its concept of the present application, making equivalent substitutions or changes should be covered within the protection scope of the present application.

Claims

1. A multi-level video watermark embedding and extraction method, characterized in that Including: Perceptually analyze the original video through a spatio-temporal attention neural network, and dynamically divide the temporally stable region and the spatially sensitive region; Construct a three-level watermark system, which includes a basic-level watermark, a first-level enhanced watermark, and a second-level enhanced watermark. The basic-level watermark uses an anti-rotation binary matrix generated by Fourier descriptors. The first-level enhanced watermark is a DCT frequency-domain hash sequence, and the second-level enhanced watermark is a semantic fingerprint based on motion vectors; In the temporally stable region, embed the basic-level watermark into the YUV low-frequency component of the I-frame through motion-compensated inter-frame prediction; in the spatially sensitive region, utilize a generative adversarial network to simulate the human visual masking effect, and embed the two-level enhanced watermarks into the DCT high-frequency subbands and motion vector residuals of the P / B frames respectively; Generate a dynamic encryption key based on video content hashing and blockchain smart contracts, and adopt a dual encryption mechanism of elliptic curve cryptography and chaotic scrambling to generate an irreversible watermark identifier; When extracting the watermark, correct geometric distortion through a spatio-temporal registration network. First, verify the integrity of the basic-level watermark stored on the blockchain, and then gradually decode the enhanced watermark through a cascaded neural network and perform cross-verification.

2. The multi-level video watermark embedding and extraction system according to claim 1, wherein The spatio-temporal attention neural network includes a temporal branch, a spatial branch, and a fusion module: Temporal branch: Use 3D convolution to extract motion consistency feature maps and identify shot boundaries and background stable regions; Spatial branch: Determine texture complex regions through Sobel edge detection and HSV color space analysis; Fusion module: Dynamically generate a probability heat map of the watermark embedding region using a gating mechanism.

3. The multi-level video watermark embedding and extraction system according to claim 1, wherein The training of the generative adversarial network includes a generator and a discriminator: Generator: Learn to inject invisible watermarks into DCT high-frequency coefficients; Discriminator: Construct an adversarial loss function by combining the human visual system model and image quality evaluation metrics.

4. The multi-level video watermark embedding and extraction system according to claim 3, wherein, The output of the discriminator includes an adversarial loss and a quality evaluation loss: Adversarial loss L adv : Based on binary cross-entropy, determine whether the input image is the original unmodified image; Quality assessment loss L IQA : Calculate the visual fidelity of the generated image through MS - SSIM and perceptual hash difference.

5. The multi-level video watermark embedding and extraction system according to claim 1, characterized in that The specific dual encryption mechanism includes: Grayscale and size-normalize each frame of the input video, and use discrete cosine transform to extract low-frequency components to generate a 64-bit PHash value; Truncate the 64-bit PHash value into a 256-bit binary sequence, generate the final private key seed through the HMAC-SHA256 algorithm, and map the final private key seed to a legal private key based on the elliptic curve secp256k1 parameters; Use a difference equation to describe an improved Logistic-Tent hyperchaotic system. The difference equation includes: Among them, r, μ, and γ are chaotic parameters, and mod1 represents the modulo operation on the value, with the modulus being 1; Use the chaotic sequence {x n} to control the row cyclic shift amount of the watermark matrix, and the row cyclic shift amount is {x n ·10 4} mod the number of rows; generate a dynamic mask based on the chaotic sequence {y n}, and perform a bitwise XOR operation on each pixel of the watermark matrix; Deploy a smart contract on the Ethereum test chain, and write the hash value of the encryption parameters into the blockchain as a verification anchor point.

6. The multi-level video watermark embedding and extraction system according to claim 1, characterized in that Geometric distortion correction includes: Construct an adversarial sample training spatio-temporal registration network that includes rotation, scaling, and shearing; Estimate affine transformation parameters through a deformable convolutional layer, and use bilinear interpolation to restore video spatio-temporal alignment.

7. The multi-level video watermark embedding and extraction system according to claim 1, characterized in that, A two-level verification mechanism is adopted during verification when extracting the watermark. The two-level verification mechanism includes primary verification and ultimate verification; Primary verification: Reconstruct the decryption matrix through the encryption parameters stored on the blockchain; Ultimate verification: Compare the Pearson correlation coefficient between the DCT-domain hash sequence and the motion vector semantic fingerprint.

8. The multi-level video watermark embedding and extraction system according to claim 7, wherein The construction method of the motion vector semantic fingerprint includes: Extract the motion vectors of the P / B frames of the video and statistically analyze their direction distribution and amplitude distribution; Map the motion vector features to a fixed-length fingerprint vector through semantic encoding.

9. Multi - level video watermark embedding and extraction system, characterized in that, The system includes: A region division module that perceptually analyzes the original video through a spatio-temporal attention neural network to dynamically divide the temporally stable region and the spatially sensitive region; A watermark construction module that constructs a three-level watermark system. The three-level watermark system includes a basic-level watermark, a first-level enhanced watermark, and a second-level enhanced watermark. The basic-level watermark is an anti-rotation binary matrix generated by Fourier descriptors. The first-level enhanced watermark is a DCT frequency-domain hash sequence, and the second-level enhanced watermark is a semantic fingerprint generated based on motion vectors; A watermark embedding module that, in the temporally stable region, embeds the basic-level watermark into the low-frequency component of the YUV of the I-frame through motion-compensated inter-frame prediction; in the spatially sensitive region, uses a generative adversarial network to simulate the human visual masking effect and embeds the two-level enhanced watermarks into the DCT high-frequency sub-bands and motion vector residuals of the P / B frames respectively; A watermark identification module that generates a dynamic encryption key based on video content hashing and blockchain smart contracts and generates an irreversible watermark identifier using a dual encryption mechanism of elliptic curve cryptography and chaotic scrambling; A watermark verification module that corrects geometric distortion through a spatio-temporal registration network during extraction, first verifies the integrity of the basic-level watermark stored on the blockchain, and then gradually decodes the enhanced watermark through a cascaded neural network and performs cross-verification.

Images