Screen shooting video watermark resisting method based on fast Zernike moment and SuperPoint network

Abstract

The invention discloses an anti-screen shooting video watermarking method based on a fast Zernike moment and a SuperPoint network, which mainly comprises two steps of watermark embedding and watermark extraction, and the core is that geometric invariance of the Zernike moment and high-precision feature positioning of the SuperPoint network are fused, and video watermarking processing for resisting screen capture attack is realized in combination with improved quantization index modulation (MQIM). In the aspect of anti-attack capability, the method is outstanding in performance when facing specific composite interference of screen shooting; after screen shooting attack including moire and geometric distortion, the average normalized correlation coefficient (NC) of watermark extraction is greater than 0.98, and the extraction success rate is stabilized at 95% or above; in the aspect of calculation efficiency, the watermark embedding and extracting speed is increased to 35 FPS, and the requirement for real-time processing is completely met.

Description

Technical Field

[0001] This invention belongs to the field of digital watermarking and information security technology, specifically relating to a robust video watermarking method for video content copyright protection, which is particularly suitable for resisting screen capture attacks (such as geometric distortion and moiré interference caused by mobile phone screen capture), and can be applied to digital media copyright authentication, leakage tracking and content integrity verification. Background Technology

[0002] Existing video watermarking technologies mainly include hybrid transform domain watermarking algorithms, quantization index modulation methods, and deep learning-based watermarking models. These technologies all exhibit significant limitations when dealing with the specific attack mode of screen capture (screen recording). The screen capture attack process involves digital-to-analog-to-digital conversion, introducing complex distortions including moiré dynamic interference, lens geometric distortion (such as barrel / pincushion distortion), and CMOS sensor noise, posing a severe challenge to the robustness of watermarks.

[0003] Limitations of traditional transform domain and quantization methods: Methods based on transform domain methods (such as DCT and DWT) or quantization index modulation suffer from a core deficiency: reliance on spectral decomposition of fixed basis functions or preset quantization step sizes. These methods struggle to maintain stable feature extraction when faced with dynamic geometric deformations caused by screen capture (such as perspective, rotation, and scaling), easily losing global geometric information and leading to a significant drop in watermark extraction accuracy. Furthermore, they are poorly adaptable to complex noise unique to screen capture (such as moiré patterns and lighting variations). Increasing watermark embedding strength to enhance anti-attack capabilities often directly results in unacceptable degradation of video quality, making it difficult to balance imperceptibility and robustness.

[0004] Challenges of Deep Learning Models: While deep learning-based watermarking models can adaptively optimize watermark embedding and extraction strategies through end-to-end training and simulate screen capture distortion to some extent, they generally suffer from high computational complexity and require large-scale labeled data for training, making it difficult to meet the low-latency requirements of real-time scenarios such as surveillance videos and live streams. Furthermore, these models have limited generalization ability to unknown attacks not covered by the training data or variations caused by different device combinations (such as differences in resolution and sensors between displays and shooting devices), resulting in poor cross-device compatibility. For example, a combination of a low-resolution screen and a high-resolution camera can lead to a watermark extraction rate drop of over 10%.

[0005] In summary, existing video watermarking technologies have significant drawbacks when dealing with screen capture attacks, including weak resistance to geometric attacks, difficulty in balancing imperceptibility and robustness, poor cross-device compatibility, and low computational efficiency. Summary of the Invention

[0006] This invention provides a watermarking method for anti-screen capture video based on fast Zernike moments and SuperPoint networks, aiming to systematically solve the technical challenges of watermarking under screen capture attacks. The primary objective is to resist moiré interference, lens geometric distortion, and CMOS noise in screen capture attacks, ensuring accurate extraction of watermark information even in high-distortion environments. While maintaining high video quality (PSNR ≥ 45dB), it improves the robustness of the watermark against geometric attacks (such as rotation and scaling), compression attacks (such as H.264 / HEVC), and frame loss attacks, achieving a synergistic optimization of imperceptibility and attack resistance. Algorithm optimization reduces computational complexity, enabling real-time watermark embedding and extraction (inference speed ≥ 30FPS) to meet the needs of real-time applications such as surveillance videos and live streams. Improved cross-device compatibility ensures a stable watermark extraction rate of over 95% across different monitor and shooting device combinations, overcoming performance fluctuations caused by device differences. We optimize the watermark embedding position and strength strategy to avoid the impact of local area loss caused by screen capture (such as border cropping and texture occlusion) on the integrity of the watermark. We enhance local robustness through dynamic feature anchoring and adaptive quantization step size adjustment.

[0007] A method for resisting screen capture video watermarking based on fast Zernike moments and SuperPoint networks includes the following steps:

[0008] (a) Watermark Embedding Process

[0009] Step 1-1: Video frame format conversion and embedding carrier selection.

[0010] Steps 1-2: Feature point extraction and filtering based on the SuperPoint network.

[0011] Steps 1-3: The Zernike moment amplitude is calculated using the fast recursive Zernike moment algorithm.

[0012] Steps 1-4: Watermark embedding based on improved MQIM.

[0013] Steps 1-5: Video frame reconstruction and encoding.

[0014] (II) Watermark Extraction Process

[0015] Step 2-1: Screen capture video preprocessing and geometric correction.

[0016] Step 2-2: Feature point matching and embedding region localization.

[0017] Steps 2-3: Zernike moment extraction and MQIM decoding.

[0018] Steps 2-4: Copyright verification.

[0019] In one possible implementation, step 1-1 is specifically performed as follows:

[0020] (1) Format conversion: Convert the input RGB format video frames to YUV color space.

[0021] (2) Carrier selection: U channel is selected as the watermark embedding carrier.

[0022] In one possible implementation, the feature point extraction and filtering are specifically as follows:

[0023] (1) Feature point extraction: The SuperPoint network is used to detect the corner features of the U-channel frame, generate a dense feature point set, output the feature point confidence heatmap, and include feature points with confidence ≥ 0.8 into the candidate set.

[0024] (2) Feature point selection: ① Boundary constraints: Assume the video frame resolution is A×B, the embedding region radius is k, and the feature point coordinates (x,y) must satisfy: x≥k, y≥k, Ax≥k, By≥k; ② Non-overlapping selection: Sort the candidate feature points in ascending order according to the x coordinate, and use the sliding window method (window step size=k) to retain points with a distance ≥k between adjacent feature points; ③ Deduplication optimization: Store the feature point coordinates through a hash table and remove duplicate coordinates.

[0025] In one possible implementation, the fast recursive Zernike moment algorithm uses a recursive strategy to optimize the numerical calculation of Zernike moments, as specifically implemented below:

[0026] (1) Definition of Zernike moment: Zernike moment is an orthogonal moment within the unit disk (x² + y² ≤ 1); radial basis functions , Let be the polar radius, representing the normalized distance from the point (x, y) to the center of the unit circle, satisfying 0 ≤ y ≤ 0. ≤1. Radial basis functions With angular direction function Composition of basis functions = , where n is the order and m is the degree of repetition, satisfying that n-|m| is even and |m|≤n;

[0027] (2) Recursive strategy:

[0028] Recursive relation: Derive a low-repetition polynomial from a radial polynomial of fixed order n and high repetition, as shown in the following formula:

[0029]

[0030] Among them, the recurrence coefficients of higher-order terms recurrence coefficients of intermediate terms Sum of coefficients of the fractional compensation term Defined as:

[0031]

[0032] Special case handling: When n=m or n=m+2, use a closed-form solution to avoid recursive divergence.

[0033] Complexity optimization: By sharing radial basis function values ​​between frames, the computational complexity is reduced from O(n^2) to O(n^2). 4 ) decreases to O(n 2 ).

[0034] (3) Normalization: Map the watermark embedding region to the unit disk and calculate the Zernike moment amplitude |A nm |

[0035] In one possible implementation, the watermark embedding is specifically as follows:

[0036] (1) Watermark preprocessing: The watermark is a 64-bit binary sequence, generated by encoding key copyright information according to a predefined format. An 8-bit CRC checksum is then appended to the 64-bit original watermark sequence, together forming a 72-bit watermark string. The key copyright information includes the copyright identifier, timestamp, and device identification code.

[0037] (2) Dynamic embedding intensity adjustment: Improved quantization index modulation MQIM is adopted, based on the Zernike moment amplitude |A nm The quantization step size S is dynamically adjusted using the following formula:

[0038]

[0039] in Let S be the normalized amplitude value (0≤a≤1), where S∈[0.01,0.03].

[0040] (3) Watermark embedding formula: Each watermark bit in the watermark string is embedded into the Zernike moment amplitude one by one. The amplitude after embedding is |A nm '| Satisfy:

[0041] When the watermark bit When =1, |A nm '|∈[kS+S / 2,(k+1)S).

[0042] When the watermark bit When = 0, |A nm '|∈[kS,kS+S / 2).

[0043] Where k is an integer, the frame update is ultimately achieved through residual concatenation:

[0044]

[0045] Among them, f W (x,y) represents the pixel value of the U-channel frame at coordinates (x,y) after watermarking, f(x,y) represents the pixel value of the original U-channel frame at coordinates (x,y), and I is the embedding strength coefficient. This is a watermark residue.

[0046] In one possible implementation, the video frame reconstruction and encoding are as follows:

[0047] (1) Channel merging: Merge the U channel frame with embedded watermark with the original Y and V channel frames and convert them into a reconstructed frame in RGB format.

[0048] (2) Video encoding: The H.264 encoding standard is used to encode the reconstructed frames, and the bit rate fluctuation is controlled to be less than or equal to the set threshold. Finally, a watermarked video is generated, which consists of a series of reconstructed video frames.

[0049] In one possible implementation, the screen capture video preprocessing and geometric correction are as follows:

[0050] (1) Deblurring: Non-local mean filtering is used to process each frame of the screen capture video to eliminate motion blur caused by the screenshot while preserving the details of the screen edges;

[0051] (2) Border enhancement and segmentation: The temporal saliency of the screen border is enhanced by the directional gradient histogram feature, and then the Canny edge detection algorithm is used to generate multi-scale edge maps.

[0052] (3) Perspective Correction: ① Corner Detection: The FAST corner detection algorithm is used to quickly select candidate corners in the edge map. A random sampling consensus algorithm is introduced: a small number of candidate points are randomly sampled multiple times to calculate the perspective transformation model, and the interior points that satisfy the model are counted. The corner set corresponding to the model with the most interior points is regarded as the correct corner library. A matching error threshold is set, and points in the corner library whose reprojection error with the ideal model is greater than this threshold are regarded as exterior points and are removed, finally obtaining the optimized corner library. ② Perspective Transformation: The perspective transformation matrix is ​​calculated based on the corner coordinates in the corner library. Based on the perspective transformation matrix, the perspective transformation is performed on the screen-captured video frames with geometric distortion caused by the screen capture angle, correcting them into regular rectangular frames. The non-screen area is cropped by the slicing algorithm to restore the original video resolution.

[0053] In one possible implementation, the feature point matching and embedding region localization are specifically as follows:

[0054] (1) Feature point extraction: The grayscale image of the Y channel of the regular rectangular frame obtained after perspective correction is extracted and processed. The SuperPoint network is used to extract feature points. The parameter configuration of the SuperPoint network is the same as that of the embedding stage. On this basis, a confidence threshold is set for filtering: only feature points with a probability value not lower than this threshold are retained and included in the candidate set for subsequent matching.

[0055] (2) Nearest neighbor matching: The K-nearest neighbor matching algorithm is used to map the extracted feature points to the coordinate system during embedding.

[0056] (3) Region positioning: Based on the coordinates of the matched feature points, determine the watermark embedding region to ensure that it is consistent with the region position in the embedding stage.

[0057] In one possible implementation, steps 2-3 are specifically performed as follows:

[0058] (1) Zernike moment calculation: For the embedded region after localization, the Zernike moment amplitude |A is calculated using the fast recursive Zernike moment algorithm. nm ''|;

[0059] (2) MQIM inverse mapping decoding: by comparing |A nm The relationship between ''| and quantization step size S, decoding watermark bits:

[0060] When|A nm When |-kS<S / 2, the number of decoded bits =0;

[0061] When|A nm When |-kS≥S / 2, the number of decoded bits =1;

[0062] (3) Watermark verification: Perform CRC verification on the decoded 72-bit sequence. If the verification passes, extract the first 64 bits as the valid watermark information output. If the verification fails, use the watermark information of adjacent frames to perform interpolation and repair, and output the valid watermark information. The specific interpolation and repair is as follows: use the 64-bit watermark information that has been successfully extracted and verified in the adjacent frames, and repair the erroneous watermark bits of the current frame by majority decision or time-series interpolation.

[0063] In one possible implementation, steps 2-3 are specifically performed as follows:

[0064] The normalized correlation coefficient (NC) between the decoded watermark and the original watermark is calculated using the following formula:

[0065]

[0066] Where N represents the total length of the watermark sequence. This represents the value of the original watermark sequence at position j. This represents the mean of the original watermark sequence w. This represents the value of the watermark sequence extracted from the screen capture video at the j-th position. This represents the mean of the watermark sequence extracted from the screen-captured video. When NC ≥ 0.95, the video is considered legitimate; when NC < 0.95, it is considered pirated, and the leak tracing information read from the successfully extracted watermark bits is output.

[0067] The beneficial effects of this invention are as follows:

[0068] 1. A fast Zernike moment calculation method based on recursion strategy is proposed. While preserving the rotation, scaling and flip invariance of Zernike moments, it completely solves the numerical overflow problem in the calculation of higher-order moments and significantly improves the computational efficiency.

[0069] 2. A feature point filtering mechanism combining SuperPoint network and multiple constraints was designed. By using boundary constraints, non-overlapping filtering and hash deduplication, the spatial uniformity and stability of the watermark embedding area in the video frame are ensured, thereby effectively combating the local information loss that may be caused by screen capture.

[0070] 3. An innovative watermark embedding method based on U-channel carrier and dynamic improved quantization index modulation (MQIM) is proposed. This method adaptively adjusts the quantization step size and embedding intensity according to the texture characteristics of the embedding region, achieving an optimized balance between high visual quality (PSNR≥45dB) and strong robustness.

[0071] 4. An end-to-end video preprocessing and geometric correction workflow for screen capture attacks was constructed. By comprehensively utilizing nonlocal mean filtering, HOG border enhancement and perspective transformation techniques, typical screen capture distortions such as moiré patterns and lens distortion were effectively suppressed, laying the foundation for subsequent high-precision watermark extraction.

[0072] Compared to the closest existing solution mentioned in the background section, this invention achieves significant technological advancements. Regarding anti-attack capabilities, this invention excels in dealing with the complex interference unique to screen capture, achieving a bit error rate (BER) as low as 0.0011 under a 10-degree rotation attack, an improvement of approximately 98% compared to traditional DCT domain solutions. Even after screen capture attacks involving moiré patterns and geometric distortions, the average normalized correlation coefficient (NC) for watermark extraction is greater than 0.98, with an extraction success rate consistently above 95%. In terms of computational efficiency, thanks to the recursive fast algorithm of Zernike moments and the lightweight feature processing flow, this invention increases the speed of watermark embedding and extraction to 35 FPS, fully meeting the requirements of real-time processing. Regarding applicability, in cross-device environments consisting of different brands and models of displays and shooting equipment, the watermark extraction rate fluctuates within a range of less than 5%, demonstrating excellent compatibility and stability. Attached Figure Description

[0073] Figure 1 The embedded process is an embodiment of the present invention.

[0074] Figure 2 This is the extraction process in an embodiment of the present invention.

[0075] Figure 3 This is a feature point confidence heatmap obtained in an embodiment of the present invention. Detailed Implementation

[0076] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0077] This application provides a method for anti-screenshot video watermarking based on fast Zernike moments and SuperPoint networks. It mainly includes two steps: watermark embedding and watermark extraction. The core is to integrate the geometric invariance of Zernike moments with the high-precision feature localization of SuperPoint networks, combined with improved quantization index modulation (MQIM) to achieve anti-screenshot attack video watermarking processing. The specific steps are as follows:

[0078] (a) Watermark Embedding Process

[0079] Step 1: Video frame format conversion and embedding medium selection

[0080] (1) Format conversion: Convert the input RGB format video frames into YUV color space. Utilize the characteristic of YUV format to separate luminance (Y channel) and chrominance (U / V channel) to reduce the impact of watermark on human visual perception.

[0081] (2) Carrier selection: The human eye is only 1 / 4 to 1 / 3 as sensitive to the chroma channel (U / V) as to the luminance channel (Y). Furthermore, the mainstream video coding standard (H.264 / H.265) uses 4:2:0 chroma downsampling, and the resolution of the U channel is 1 / 4 of that of the Y channel. After the watermark is embedded, it is easy to be compressed and retained. Therefore, the U channel is selected as the watermark embedding carrier.

[0082] Step 2: Feature point extraction and filtering based on the SuperPoint network (Self-Supervised Interest Point Detection and Description Network).

[0083] (1) Feature point extraction: The SuperPoint network is used to detect the corner features of the U-channel frame, generate a dense feature point set, and output a feature point confidence heatmap (e.g. Figure 3 As shown, the upper part is the original experimental image, and the lower part is the corresponding feature point confidence heatmap. The value stored at each pixel location represents the probability estimate that the pixel is an image feature point, i.e., the "confidence" of that point. The higher the confidence, the greater the possibility that the point is a stable feature point (and the more obvious the point is in the heatmap). Feature points with a confidence of ≥0.8 are included in the candidate set.

[0084] (2) Feature point selection: ① Boundary constraint: Assume the video frame resolution is A×B (e.g., 1920×1080), the embedding region radius k=128 pixels (experiments have verified that this size can balance embedding capacity and robustness), and the feature point coordinates (x,y) must satisfy: x≥k, y≥k, Ax≥k, By≥k, to avoid feature points being close to the frame boundary, resulting in incomplete embedding regions; ② Non-overlapping selection: Sort candidate feature points in ascending order according to x coordinate, and use the sliding window method (window step size=k) to retain points with a distance ≥k between adjacent feature points, to avoid overlapping embedding regions causing watermark interference; ③ Deduplication optimization: Store feature point coordinates through a hash table, and remove duplicate coordinates (in the embodiment, the ratio of the number of duplicate feature points to the total number of candidate feature points, i.e., the duplication rate, is controlled to be ≤5%), shortening the subsequent calculation time and ensuring that the finally determined watermark embedding region is evenly distributed within the frame (20-30 effective feature points are retained per frame).

[0085] Step 3: Calculate the Zernike moment amplitude using the fast recursive Zernike moment algorithm.

[0086] To address the issue in traditional Zernike moment numerical calculations where the radial polynomial R is directly calculated... nmThe high risk of numerical overflow and high computational complexity caused by the high-order exponentiation and factorial operations involved in (ρ) are problems. To address this, this invention proposes a fast recursive Zernike moment algorithm, employing a recursive strategy to optimize the numerical calculation of Zernike moments. This recursive strategy is the core of the algorithm's "fast" implementation. It completely avoids complex factorial operations in each step of the calculation through recursive relations, thus solving the numerical overflow problem. Simultaneously, it reduces the computational complexity from the traditional O(n^2) to O(ρ). 4 The value was significantly reduced to O(n). 2 The specific implementation is as follows:

[0087] (1) Definition of Zernike moment: Zernike moment is an orthogonal moment within the unit disk (x² + y² ≤ 1); radial basis functions , Let be the polar radius, representing the normalized distance from the point (x, y) to the center of the unit circle, satisfying 0 ≤ y ≤ 0. ≤1. Radial basis functions With angular direction function Composition of basis functions = , where n is the order (positive integer or 0), m is the repetition degree, and n-|m| is even and |m|≤n;

[0088] (2) Recursive strategy:

[0089] Recursive relation: Derive a low-repetition polynomial from a radial polynomial of fixed order n and high repetition, as shown in the following formula:

[0090]

[0091] Among them, the recurrence coefficients of higher-order terms recurrence coefficients of intermediate terms Sum of coefficients of the fractional compensation term Defined as:

[0092]

[0093] Special case handling: When n=m or n=m+2, use a closed-form solution (e.g., when n=m, = Avoid recursive divergence.

[0094] Complexity optimization: By sharing radial basis function values ​​between frames, the computational complexity is reduced from O(n^2) to O(n^2). 4 ) decreases to O(n 2 This meets the real-time requirements.

[0095] (3) Normalization: Map the watermark embedding region (128×128) to the unit disk and calculate the Zernike moment amplitude |A nm(Amplitude has rotational and scaling invariance, and is the core carrier for watermark embedding).

[0096] Step 4: Watermark embedding based on improved MQIM

[0097] (1) Watermark preprocessing: The watermark is a 64-bit binary sequence (this sequence is generated by encoding key copyright information such as copyright logo, timestamp and device identification code (ID) according to a predefined format). An 8-bit CRC check code is generated and appended to the 64-bit original watermark sequence to form a 72-bit watermark string, which improves the fault tolerance of extraction.

[0098] (2) Dynamic embedding intensity adjustment: Improved quantization index modulation (MQIM) is adopted, based on the Zernike moment amplitude |A nm The quantization step size S is dynamically adjusted using the following formula:

[0099]

[0100] in The amplitude is a normalized value (0≤a≤1), S∈[0.01,0.03] (experiments have verified that this range can balance imperceptibility and robustness).

[0101] (3) Watermark embedding formula: Each watermark bit in the watermark string is embedded into the Zernike moment amplitude one by one. The amplitude after embedding is |A nm '| Satisfy:

[0102] When the watermark bit When =1, |A nm '|∈[kS+S / 2,(k+1)S).

[0103] When the watermark bit When = 0, |A nm '|∈[kS,kS+S / 2).

[0104] Where k is an integer, the frame update is ultimately achieved through residual concatenation:

[0105]

[0106] Among them, f W (x,y) represents the pixel value of the U-channel frame after watermarking at coordinates (x,y), f(x,y) represents the pixel value of the original U-channel frame at coordinates (x,y), and I is the embedding strength coefficient (I is 0.8 in this embodiment). This is a watermark residue.

[0107] Step 5: Video Frame Reconstruction and Encoding

[0108] (1) Channel merging: Merge the U channel frame with embedded watermark with the original Y and V channel frames and convert them into a reconstructed frame in RGB format.

[0109] (2) Video encoding: The H.264 encoding standard is used to encode the reconstructed frames, and the bit rate fluctuation is controlled to be less than or equal to a set threshold. In this embodiment, it is preferably 5% to ensure that the bit rate of the watermarked video is consistent with that of the original video and to avoid the loss of the watermark due to encoding. Finally, a watermarked video is generated, which consists of a series of reconstructed video frames.

[0110] (II) Watermark Extraction Process

[0111] Step 1: Preprocessing and geometric correction of screen capture video (referring to the video obtained by recording the screen of a watermarked video being played using a shooting device).

[0112] (1) Deblurring: Non-local mean filtering is used to replace traditional Gaussian filtering to process each frame of the screen capture video, eliminating motion blur caused by the screenshot, while preserving screen edge details (improving edge clarity by 20%).

[0113] (2) Border enhancement and segmentation: The temporal saliency of the screen border is enhanced by histogram of oriented gradients (HOG) features. Then, the Canny edge detection algorithm (adaptive threshold set to 30 pixels for the inner border and 1 pixel for the outer border in the embodiment) is used to generate a multi-scale edge map to distinguish the real screen border from environmental noise.

[0114] (3) Perspective Correction: ① Corner Detection: The FAST corner detection algorithm is used to quickly select candidate corners in the edge map. To eliminate false detections and obtain accurate corner coordinates, the RANSAC (Random Sample Consensus) algorithm is introduced: a small number of candidate points are randomly sampled multiple times to calculate the perspective transformation model, and the interior points that satisfy the model are counted. The set of corner points corresponding to the model with the most interior points (i.e., the four corner points of the screen border) is regarded as the correct corner point library. A matching error threshold is set (in this example, it can be 1 pixel). Points in the corner point library whose reprojection error with the ideal model is greater than this threshold are regarded as exterior points and are eliminated. Finally, the optimized corner point library is obtained. ② Perspective Transformation: The perspective transformation matrix is ​​calculated based on the corner coordinates in the corner point library. Based on the perspective transformation matrix, the perspective transformation is performed on the screen-captured video frame with geometric distortion caused by the screen capture angle, correcting it into a regular rectangular frame. The non-screen area is cropped by the slicing algorithm to restore the original video resolution (e.g., 1920×1080).

[0115] Step 2: Feature point matching and embedding region localization

[0116] (1) Feature point extraction: For the regular rectangular frame obtained after perspective correction in step 1, the grayscale image of its Y channel is extracted and processed (the feature stability of the Y channel is better than that of the U channel). The SuperPoint network is used to extract feature points. The SuperPoint network uses the parameter configuration of the embedding stage. On this basis, a confidence threshold (≥0.8 in this embodiment) is set for filtering: only feature points with a probability value not lower than this threshold are retained and included in the candidate set for subsequent matching.

[0117] (2) Nearest neighbor matching: In order to deal with the complex geometric transformations that may be introduced during the screen capture process, which may cause the feature points to shift (such as ±15% scaling, ±20° rotation), the K-nearest neighbor (K=3) matching algorithm is used to map the extracted feature points to the coordinate system during embedding, which can achieve excellent matching accuracy >95%.

[0118] (3) Region positioning: Based on the coordinates of the matched feature points, determine the watermark embedding region (128×128) to ensure that it is consistent with the region position in the embedding stage.

[0119] Step 3: Zernike Moment Extraction and MQIM Decoding

[0120] (1) Zernike moment calculation: For the embedded region after localization, the Zernike moment amplitude |A is calculated using the fast recursive Zernike moment algorithm. nm ''|;

[0121] (2) MQIM inverse mapping decoding: by comparing |A nm The relationship between ''| and quantization step size S, decoding watermark bits:

[0122] When|A nm When |-kS<S / 2, the number of decoded bits =0;

[0123] When|A nm When |-kS≥S / 2, the number of decoded bits =1;

[0124] (3) Watermark verification: Perform CRC verification on the decoded 72-bit sequence. If the verification passes, extract the first 64 bits as the valid watermark information output. If the verification fails, use the watermark information of adjacent frames to perform interpolation and repair, and output the valid watermark information. The specific interpolation and repair is as follows: use the 64-bit watermark information that has been successfully extracted and verified in the adjacent frames, and repair the erroneous watermark bits of the current frame by majority decision or time-series interpolation.

[0125] Step 4: Copyright Verification

[0126] Calculate the normalized correlation coefficient (NC) between the decoded watermark and the original watermark.

[0127]

[0128] Where N represents the total length of the watermark sequence. This represents the value of the original watermark sequence at position j. This represents the mean of the original watermark sequence w. This represents the value of the watermark sequence extracted from the screen capture video at the j-th position. This represents the mean of the watermark sequence extracted from the screen-captured video. When NC ≥ 0.95, the video is considered legitimate; when NC < 0.95, it is considered pirated. The output is the leak tracing information (shooting device ID, timestamp) read from the successfully extracted watermark bits.

[0129] In terms of technical implementation, this invention also considers potential alternatives and expansion directions. If computational resources are extremely limited in practical applications, geometrically invariant fusion orthogonal moments (such as Pseudo-Zernike moments, Orthogonal Fourier-Mellin moments, and Zernike-Fourier moments) can be used to replace Zernike moments, but their fast computation strategies need to be redesigned accordingly. In the feature extraction stage, the SuperPoint network can be replaced with a traditional SIFT feature combined with RANSAC matching scheme when necessary, but this may sacrifice processing speed to some extent. From an application expansion perspective, this invention can be adapted to AR / VR scenarios, enhancing the robustness of watermarks in 3D immersive content by integrating multi-view feature synchronization technology. Furthermore, the watermark summary information generated by this solution can be combined with blockchain technology to achieve tamper-proof evidence storage and traceability of copyright information, further expanding its commercial value in the field of digital asset protection.

[0130] This invention possesses clear commercial application prospects and market value. Its core market positioning lies in digital content copyright protection, effectively preventing illegal screen recording and unauthorized access to films and television shows during theatrical releases or online streaming. It can also be used to protect sensitive video content from recording and leakage in scenarios such as online meetings and internal training. In terms of derivative markets, this technology can provide short video platforms and online education institutions with low-cost, high-efficiency video copyright authentication and source tracking services. Simultaneously, its strong anti-attack and real-time processing capabilities make it highly suitable for the field of intelligent security monitoring, enabling proactive detection and alarm for tampering of surveillance videos.

Claims

1. A method for anti-screen capture video watermarking based on fast Zernike moments and SuperPoint networks, characterized in that, The steps include the following: (a) Watermark Embedding Process Step 1-1: Video frame format conversion and embedding medium selection; Steps 1-2: Feature point extraction and filtering based on the SuperPoint network; Steps 1-3: Calculate the Zernike moment amplitude using the fast recursive Zernike moment algorithm; Steps 1-4: Watermark embedding based on improved MQIM; Steps 1-5: Video frame reconstruction and encoding; (II) Watermark Extraction Process Step 2-1: Screen capture video preprocessing and geometric correction; Step 2-2: Feature point matching and embedding region localization; Steps 2-3: Zernike moment extraction and MQIM decoding; Steps 2-4: Copyright verification.

2. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 1, characterized in that, The specific operation of step 1-1 is as follows: (1) Format conversion: Convert the input RGB format video frames to YUV color space; (2) Carrier selection: U channel is selected as the watermark embedding carrier.

3. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 2, characterized in that, The feature point extraction and filtering are as follows: (1) Feature point extraction: The SuperPoint network is used to detect the corner features of the U-channel frame, generate a dense feature point set, output the feature point confidence heatmap, and include feature points with confidence ≥ 0.8 into the candidate set; (2) Feature point selection: ① Boundary constraints: Assume the video frame resolution is A×B, the embedding region radius is k, and the feature point coordinates (x,y) must satisfy: x≥k, y≥k, Ax≥k, By≥k; ② Non-overlapping selection: Sort the candidate feature points in ascending order according to the x coordinate, and use a sliding window to retain points with a distance ≥k between adjacent feature points; ③ Deduplication optimization: Store the feature point coordinates through a hash table and remove duplicate coordinates.

4. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 3, characterized in that, The fast recursive Zernike moment algorithm uses a recursive strategy to optimize the numerical calculation of Zernike moments, and the specific implementation is as follows: (1) Definition of Zernike moment: Zernike moment is an orthogonal moment within the unit disk; radial basis functions , Let be the polar radius, representing the normalized distance from the point (x, y) to the center of the unit circle, satisfying 0 ≤ y ≤ 0. ≤1; Radial basis functions With angular direction function Composition of basis functions = , where n is the order and m is the degree of repetition, satisfying that n-|m| is even and |m|≤n; (2) Recursive strategy: Recursive relation: Derive a low-repetition polynomial from a radial polynomial of fixed order n and high repetition, as shown in the following formula: Among them, the recurrence coefficients of higher-order terms recurrence coefficients of intermediate terms Sum of coefficients of the fractional compensation term Defined as: Special case handling: When n=m or n=m+2, use a closed-form solution to avoid recursive divergence; Complexity optimization: By sharing radial basis function values ​​between frames, the computational complexity is reduced from O(n^2) to O(n^2). 4 ) decreases to O(n 2 ); (3) Normalization: Map the watermark embedding region to the unit disk and calculate the Zernike moment amplitude |A nm | 5. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 4, characterized in that, The watermark embedding is specifically as follows: (1) Watermark preprocessing: The watermark is a 64-bit binary sequence, generated by encoding key copyright information according to a predefined format, and an 8-bit CRC check code is generated and appended to the 64-bit original watermark sequence to form a 72-bit watermark string; the key copyright information includes copyright identifier, timestamp and device identification code; (2) Dynamic embedding intensity adjustment: Improved quantization index modulation MQIM is adopted, based on the Zernike moment amplitude |A nm The quantization step size S is dynamically adjusted using the following formula: in Let S be the normalized amplitude value (0≤a≤1), where S∈[0.01,0.03]. (3) Watermark embedding formula: Each watermark bit in the watermark string is embedded into the Zernike moment amplitude one by one. The amplitude after embedding is |A nm '| Satisfy: When the watermark bit When =1, |A nm '|∈[kS+S / 2,(k+1)S); When the watermark bit When = 0, |A nm '|∈[kS,kS+S / 2); Where k is an integer, the frame update is ultimately achieved through residual concatenation: Among them, f W (x,y) represents the pixel value of the U-channel frame at coordinates (x,y) after watermarking, f(x,y) represents the pixel value of the original U-channel frame at coordinates (x,y), and I is the embedding strength coefficient. This is a watermark residue.

6. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 5, characterized in that, The video frame reconstruction and encoding are detailed as follows: (1) Channel merging: Merge the U channel frame with embedded watermark with the original Y and V channel frames and convert it into a reconstructed frame in RGB format; (2) Video encoding: The H.264 encoding standard is used to encode the reconstructed frames, and the bit rate fluctuation is controlled to be less than or equal to the set threshold. Finally, a watermarked video is generated, which consists of a series of reconstructed video frames.

7. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 1, characterized in that, The specific preprocessing and geometric correction of the screen-captured video are as follows: (1) Deblurring: Non-local mean filtering is used to process each frame of the screen capture video to eliminate motion blur caused by the screenshot while preserving the details of the screen edges; (2) Border enhancement and segmentation: The temporal saliency of the screen border is enhanced by the directional gradient histogram feature, and then the Canny edge detection algorithm is used to generate multi-scale edge maps; (3) Perspective correction: ① Corner detection: The FAST corner detection algorithm is used to quickly select candidate corners in the edge map; a random sampling consensus algorithm is introduced: a small number of candidate points are randomly sampled multiple times to calculate the perspective transformation model, and the interior points that satisfy the model are counted. The corner set corresponding to the model with the most interior points is regarded as the correct corner library; a matching error threshold is set, and the points in the corner library whose reprojection error with the ideal model is greater than this threshold are regarded as exterior points and are removed, and finally the optimized corner library is obtained; ② Perspective transformation: the perspective transformation matrix is ​​calculated according to the corner coordinates in the corner library. Based on the perspective transformation matrix, the perspective transformation is performed on the screen-captured video frame with geometric distortion caused by the screen capture angle, and it is corrected into a regular rectangular frame. The non-screen area is cropped by the slicing algorithm to restore the original video resolution.

8. The anti-screen capture video watermarking method based on fast Zernike moments and SuperPoint networks according to claim 7, characterized in that, The feature point matching and embedding region localization are as follows: (1) Feature point extraction: For the regular rectangular frame obtained after perspective correction, the grayscale image of its Y channel is extracted and processed; the SuperPoint network is used to extract feature points. The SuperPoint network uses the parameter configuration of the embedding stage. On this basis, a confidence threshold is set for filtering: only feature points with a probability value not lower than this threshold are retained and included in the candidate set for subsequent matching. (2) Nearest neighbor matching: The K-nearest neighbor matching algorithm is used to map the extracted feature points to the coordinate system during embedding; (3) Region positioning: Based on the coordinates of the matched feature points, determine the watermark embedding region to ensure that it is consistent with the region position in the embedding stage.

9. A method for anti-screen capture video watermarking based on fast Zernike moments and SuperPoint networks according to claim 8, characterized in that, The specific operations for steps 2-3 are as follows: (1) Zernike moment calculation: For the embedded region after localization, the Zernike moment amplitude |A is calculated using the fast recursive Zernike moment algorithm. nm ''|; (2) MQIM inverse mapping decoding: by comparing |A nm The relationship between ''| and quantization step size S, decoding watermark bits: When|A nm When |-kS<S / 2, the number of decoded bits =0; When|A nm When |-kS≥S / 2, the number of decoded bits =1; (3) Watermark verification: Perform CRC verification on the decoded 72-bit sequence. If the verification passes, extract the first 64 bits as the valid watermark information output. If the verification fails, use the watermark information of adjacent frames to perform interpolation and repair, and output the valid watermark information. The specific interpolation and repair is as follows: use the 64-bit watermark information that has been successfully extracted and verified in the adjacent frames, and repair the erroneous watermark bits of the current frame by majority decision or time-series interpolation.

10. A method for anti-screen capture video watermarking based on fast Zernike moments and SuperPoint networks according to claim 9, characterized in that, The specific operations for steps 2-3 are as follows: The normalized correlation coefficient NC between the decoded watermark and the original watermark is calculated using the following formula: Where N represents the total length of the watermark sequence. This represents the value of the original watermark sequence at position j. This represents the mean of the original watermark sequence w. This represents the value of the watermark sequence extracted from the screen capture video at the j-th position. This represents the mean of the watermark sequence extracted from the screen capture video; when NC≥0.95, it is determined to be a legitimate video; when NC<0.95, it is determined to be a pirated video, and the leak tracing information read from the successfully extracted watermark bits is output.

Images