A video key frame fingerprint checking method for illegal clipping identification

Abstract

The application relates to the field of video image recognition and discloses a video key frame fingerprint checking method for illegal clipping identification, which comprises the following steps: acquiring a video key frame sequence and extracting image feature points, constructing a space geometric grid covering a local area of the key frame through triangulation based on pixel coordinates; calculating the area ratio of adjacent triangles of a shared grid edge to generate an affine invariant feature vector; and determining a local tampering position by calculating a feature difference matrix and a time sequence deviation tensor between adjacent frames. The application utilizes the invariance principle of topological geometric structure, realizes substantial decoupling of global legal affine transformation and local malicious tampering, effectively shields coordinate drift disturbance caused by picture anti-shake and picture frame scaling, enhances the targeted recognition accuracy for picture splicing and object erasing, and reduces the dependence of the fingerprint checking process on pixel-level comparison calculation power.

Description

Technical Field

[0001] This invention relates to a video keyframe fingerprint verification method for identifying illegal editing, belonging to the field of video image recognition technology. Background Technology

[0002] Current video copyright protection and audiovisual content verification typically employ the extraction of spatial statistical features or deep convolutional features from video frames to generate perceptual hash fingerprints. This approach transforms the verification of massive amounts of video data into a distance metric between feature vectors by establishing feature vector representations of video images. This enables automated identification of duplicate content or large-scale content manipulation. In real video stream distribution, platforms commonly employ legitimate automated post-processing techniques such as electronic image stabilization, adaptive aspect ratio cropping, and dynamic resolution scaling to adapt to different resolution devices or optimize visual experience. Due to the global pixel coordinate drift and physical distance scaling resulting from these processes, existing fingerprint verification mechanisms are prone to misidentifying legitimate global geometric transformations as malicious local content tampering. Existing technical solutions rely excessively on the absolute spatial coordinates of feature points, making them highly susceptible to fingerprint mismatch due to nonlinear shifts in the coordinate system when faced with legitimate image editing actions.

[0003] If noise interference from global transformation is offset by adjusting the fuzzy threshold of feature comparison, it will mask the structural breaks in micro-regions within keyframes, leading to a decrease in the system's sensitivity to capturing covert tampering behaviors such as local image splicing or object erasure. Meanwhile, while using a pixel-by-pixel high-dimensional feature comparison path can improve accuracy, it incurs a huge computational load and storage overhead, failing to meet the stringent timeliness requirements of large-scale streaming media verification. The diverse adaptation of hardware terminals and image acquisition exhibit physical-layer constraints, and the software control methods at the data processing level also suffer from structural deficiencies. For example, Chinese invention patent CN114202723B discloses a method for image recognition... The intelligent editing application method, device, and medium calculate the pixel value of a preset position pixel in a specific video frame after bone stripping and linear registration, and combine it with a standard image to perform similarity scoring to extract keyframes. This mechanism relies heavily on the absolute spatial coordinates of the preset pixel position and the prior alignment of a fixed standard image. In real streaming media distribution, when faced with nonlinear coordinate drift caused by electronic image stabilization or resolution scaling, the core preset premise is fundamentally mismatched with the actual evolving scenario, resulting in baseline distortion of the fixed position pixel value and similarity discrimination. It is impossible to decouple global affine deformation and local structural tampering in terms of mechanism, which leads to the limitation of the detection performance of the verification network in non-ideal engineering environments.

[0004] Therefore, how to construct a spatial topological representation mechanism with affine invariance, so that it can shield global geometric transformation noise while maintaining high physical sensitivity to local tampering, is the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems in the background art, the present invention provides a video keyframe fingerprint verification method for illegal editing identification, comprising the following steps:

[0006] Step S101: Obtain the keyframe sequence of the video to be verified;

[0007] Step S102: Extract image feature points from the current keyframe and obtain the pixel coordinates of each image feature point in the current two-dimensional image coordinate system;

[0008] Step S103: Perform Delaunay triangulation on each image feature point according to the pixel coordinates to construct a spatial geometric mesh covering the local image region of the current keyframe.

[0009] Step S104: In the spatial geometric grid, retrieve adjacent triangle pairs with shared edges. Based on the pixel coordinates of each image feature point, calculate the first area of ​​the first triangle and the second area of ​​the second triangle in each adjacent triangle pair, and extract the ratio of the first area to the second area to determine the local topological ratio.

[0010] Step S105: Traverse all adjacent triangle pairs in the spatial geometric grid, arrange the local topological ratios according to the preset spatial index order, and generate the affine invariant feature vector corresponding to the current keyframe.

[0011] Step S106: Calculate the difference matrix of the affine invariant eigenvectors between two adjacent keyframes, calculate the pixel displacement deviation of each image feature point between two adjacent keyframes, and perform a logical AND operation between the distribution of non-zero elements in the difference matrix whose absolute value exceeds the preset fluctuation tolerance and the pulse jump points of the pixel displacement deviation to determine the local tampering location in the video to be checked.

[0012] Preferably, the step of determining the location of local tampering includes: obtaining the displacement difference components of each image feature point on the video time axis and the change in topological weight in the spatial geometric grid; determining the temporal deviation tensor characterizing the spatiotemporal topological continuity of the video sequence by linearly weighting the displacement difference components and the change in topological weight; and using the abnormal response value of the temporal deviation tensor to determine illegal editing.

[0013] Preferably, the steps for extracting image feature points include: obtaining the texture complexity of each pixel region in the current keyframe; establishing a non-linear mapping relationship between texture complexity and feature extraction threshold; adjusting the feature extraction threshold of different pixel regions using the non-linear mapping relationship; reducing the feature extraction threshold in regions where the texture complexity is below a preset lower limit to increase sampling density, and increasing the feature extraction threshold in regions where the texture complexity is above a preset upper limit to remove redundant feature points.

[0014] Preferably, the steps for performing Delaunay triangulation include: obtaining the convex hull boundary of each image feature point in the current two-dimensional image coordinate system; dividing the current keyframe into a set of non-overlapping triangles based on the convex hull boundary and the distribution position of each image feature point, and constructing a spatial geometric mesh that satisfies the empty circle property.

[0015] Preferably, the steps for determining illegal editing further include: establishing a sliding time window, statistically analyzing the mean and standard deviation of the temporal deviation tensor within the sliding time window; calculating the real-time deviation response value corresponding to the current keyframe; and determining that the video segment corresponding to the sliding time window has been tampered with by frame skipping or frame insertion when the real-time deviation response value deviates from the mean by more than 3 times the standard deviation.

[0016] Preferably, after generating the affine invariant feature vector, the method further includes: using a locality-sensitive hashing algorithm to perform dimensionality reduction projection on the affine invariant feature vector to generate a fixed-length binary fingerprint code.

[0017] Preferably, the step of determining the location of local tampering further includes: obtaining the reference fingerprint code of the original video; calculating the Hamming distance between the binary fingerprint code of the video to be checked and the reference fingerprint code; identifying the frame index position where the Hamming distance shows a non-continuous step change, and determining the tampering type in combination with the pixel displacement deviation.

[0018] Preferably, the tampering types include partial image splicing, object erasure, frame extraction, and frame insertion; wherein, partial image splicing and object erasure are identified by the numerical fluctuation of the local topology ratio, and frame extraction and frame insertion are identified by the numerical anomaly of the displacement differential component.

[0019] Preferably, after determining the location of the local tampering, the method further includes: retrieving the fingerprint database index corresponding to the reference fingerprint code, and outputting a verification report containing the timestamp information of the local tampering location and the coordinate information of the tampered area.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. In the verification of video keyframe fingerprints for illegal editing, this invention achieves logical decoupling of global geometric transformation and local structural tampering at the low-level feature expression stage by establishing a sequence of area ratios of adjacent triangles based on shared topological edges within the keyframe. Since the area ratio of triangles sharing the same edge has mathematical invariance under global affine transformations such as translation, rotation, proportional scaling, and conventional linear cropping, this technical solution enables video fingerprints to have natural immunity to legitimate post-processing conditions in the streaming media distribution process, avoiding feature fluctuations caused by image displacement or size adjustment. At the same time, once the video image encounters malicious operations such as local splicing or object erasure that cause a substantial break in the topological connection relationship, the area ratio will produce a discontinuous step change, thereby enabling the verification system to accurately lock local micro-tampering points with a low false alarm rate.

[0022] 2. This invention extracts the spatiotemporal distortion tensor between adjacent keyframes, transforming the static feature comparison of isolated frames into a dynamic evolution analysis of the spatiotemporal topological continuity of video sequences. This mechanism utilizes the synergistic change of the displacement difference of visual semantic anchors on the time axis and the spatial topological weights to construct dual physical constraints covering spatial geometry and temporal logic. This multi-dimensional feature coupling method enables the system to not only be sensitive to the pixel distribution of the image, but also to have a strong ability to capture the physical coherence of motion vectors between frames. Even for frame extraction or frame insertion operations within a short time window, deterministic identification can be achieved through abnormal pulses in the spatiotemporal distortion tensor, effectively making up for the detection blind spots of traditional global hash algorithms when dealing with temporal tampering.

[0023] 3. By adopting an anchor point adaptive screening mechanism based on local spatial information entropy density, this invention achieves a dynamic balance between computational complexity and recognition accuracy. The system dynamically adjusts the feature extraction threshold by evaluating the texture complexity of different regions of the key frame in real time. It increases the sampling density in sparse texture regions to ensure the connectivity of the topology graph, and removes redundant nodes in dense texture regions to reduce the computational load of subsequent local sensitive hashing. This strategy of actively adjusting the perception fineness according to the characteristics of image content not only improves the robustness of fingerprints under complex conditions, but also shortens the response cycle of feature retrieval and comparison in large-scale video copyright databases by compressing the feature dimension at the source, thereby improving the efficiency of engineering implementation. Attached Figure Description

[0024] Figure 1 This is a flowchart of the video keyframe fingerprint verification method for illegal editing identification according to the present invention;

[0025] Figure 2 This is a logic diagram for generating binary fingerprint encoding in this invention.

[0026] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0028] A method for verifying video keyframe fingerprints for identifying illegal editing includes the following steps:

[0029] Step S101: Obtain the keyframe sequence of the video to be verified;

[0030] Step S102: Extract image feature points from the current keyframe and obtain the pixel coordinates of each image feature point in the current two-dimensional image coordinate system;

[0031] Step S103: Perform Delaunay triangulation on each image feature point according to the pixel coordinates to construct a spatial geometric mesh covering the local image region of the current keyframe.

[0032] Step S104: In the spatial geometric grid, retrieve adjacent triangle pairs with shared edges. Based on the pixel coordinates of each image feature point, calculate the first area of ​​the first triangle and the second area of ​​the second triangle in each adjacent triangle pair, and extract the ratio of the first area to the second area to determine the local topological ratio.

[0033] Step S105: Traverse all adjacent triangle pairs in the spatial geometric grid, arrange the local topological ratios according to the preset spatial index order, and generate the affine invariant feature vector corresponding to the current keyframe.

[0034] Step S106: Calculate the difference matrix of the affine invariant eigenvectors between two adjacent keyframes, calculate the pixel displacement deviation of each image feature point between two adjacent keyframes, and perform a logical AND operation between the distribution of non-zero elements in the difference matrix whose absolute value exceeds the preset fluctuation tolerance and the pulse jump points of the pixel displacement deviation to determine the local tampering location in the video to be checked.

[0035] Preferably, the step of determining the location of local tampering includes: obtaining the displacement difference components of each image feature point on the video time axis and the change in topological weight in the spatial geometric grid; determining the temporal deviation tensor characterizing the spatiotemporal topological continuity of the video sequence by linearly weighting the displacement difference components and the change in topological weight; and using the abnormal response value of the temporal deviation tensor to determine illegal editing.

[0036] Preferably, the steps for extracting image feature points include: obtaining the texture complexity of each pixel region in the current keyframe; establishing a non-linear mapping relationship between texture complexity and feature extraction threshold; adjusting the feature extraction threshold of different pixel regions using the non-linear mapping relationship; reducing the feature extraction threshold in regions where the texture complexity is below a preset lower limit to increase sampling density, and increasing the feature extraction threshold in regions where the texture complexity is above a preset upper limit to remove redundant feature points.

[0037] Preferably, the steps for performing Delaunay triangulation include: obtaining the convex hull boundary of each image feature point in the current two-dimensional image coordinate system; dividing the current keyframe into a set of non-overlapping triangles based on the convex hull boundary and the distribution position of each image feature point, and constructing a spatial geometric mesh that satisfies the empty circle property.

[0038] Preferably, the steps for determining illegal editing further include: establishing a sliding time window, statistically analyzing the mean and standard deviation of the temporal deviation tensor within the sliding time window; calculating the real-time deviation response value corresponding to the current keyframe; and determining that the video segment corresponding to the sliding time window has been tampered with by frame skipping or frame insertion when the real-time deviation response value deviates from the mean by more than 3 times the standard deviation.

[0039] Preferably, after generating the affine invariant feature vector, the method further includes: using a locality-sensitive hashing algorithm to perform dimensionality reduction projection on the affine invariant feature vector to generate a fixed-length binary fingerprint code.

[0040] Preferably, the step of determining the location of local tampering further includes: obtaining the reference fingerprint code of the original video; calculating the Hamming distance between the binary fingerprint code of the video to be checked and the reference fingerprint code; identifying the frame index position where the Hamming distance shows a non-continuous step change, and determining the tampering type in combination with the pixel displacement deviation.

[0041] Preferably, the tampering types include partial image splicing, object erasure, frame extraction, and frame insertion; wherein, partial image splicing and object erasure are identified by the numerical fluctuation of the local topology ratio, and frame extraction and frame insertion are identified by the numerical anomaly of the displacement differential component.

[0042] Preferably, after determining the location of the local tampering, the method further includes: retrieving the fingerprint database index corresponding to the reference fingerprint code, and outputting a verification report containing the timestamp information of the local tampering location and the coordinate information of the tampered area.

[0043] Example 1: The method claimed in this invention is applicable to digital rights protection scenarios requiring high-throughput automated verification of illegally edited video content. In real-world scenarios of streaming media distribution and video copyright library retrieval, video clips commonly face legitimate automated post-processing such as global translation and rotation caused by electronic image stabilization, adaptive aspect ratio cropping, and dynamic resolution scaling. These legitimate transformation operations cause significant overall drift and physical distance scaling of the pixel coordinates of keyframe image feature points, easily leading to global feature mismatch and false alarms in conventional perceptual hash verification mechanisms that rely on absolute coordinate displacement or Euclidean distance of feature points. To achieve substantial decoupling between globally legitimate geometric transformations and local malicious tampering from the underlying feature expression stage, this method reduces the dimensionality of video keyframe pixel domain changes to a geometric topology layer with specific semantic anchor points, and monitors common edges within the topology graph. The algebraic constancy of the ratio of triangle areas provides natural immunity to legal affine transformations, while maintaining high physical sensitivity to local microscopic tampering. The specific processing flow is as follows: Before acquiring the keyframe sequence of the video to be verified and extracting image feature points, the system needs to de-identify and de-sensitize the video image data to ensure the technical security of user privacy data. The processor, as the core computing hardware unit, receives the video sequence to be verified through an authorized interface and extracts the keyframe sequence according to a preset step size. In the initial state definition procedure for extracting image feature points from the current keyframe, the current keyframe, as the object of action, is a two-dimensional image pixel matrix after video decoding, with a resolution specification adapted to streaming media transmission standards. The processor, as the enabling environment, needs to meet the real-time decoding and floating-point matrix operations of multiple video streams.

[0044] The process judgment quantization procedure begins with the evaluation of the texture complexity of each pixel region in the current keyframe. To avoid random fragmentation of the topology map due to extremely uneven feature distribution in sparse or dense texture regions, the processor calculates the spatial information entropy density of each local region in the current keyframe as quantification data for monitoring texture complexity. A nonlinear mapping relationship is established between the spatial information entropy density and the feature extraction threshold. To eliminate the parameter black box caused by empirical settings, the nonlinear mapping relationship is determined through a real-time calibration formula. Based on the Shannon information entropy principle, the processor statistically analyzes the histogram of the gray-level distribution of the pixel matrix within the local image region. By calculating the negative weighted sum of the logarithmic products of the probabilities of occurrence of each gray level, the dimensionless spatial information entropy density is output. The processor performs nonlinear mapping calculations according to the negative exponential decay model: ,in, To extract a threshold for the output target features and constrain it to positive integer boundaries, The minimum basic sampling threshold supported by the current hardware system. To adjust the compensation gain for fluctuations in ambient illuminance, To control the slope decay constant of the feature screening, For the spatial information entropy density, the adjustment gain and attenuation constant in the formula are pre-determined through the following calibration process: The system performs feature extraction experiments on an untampered benchmark video set under standard ambient illumination, and counts the number of feature response points under different illumination change curves. After testing, the value range of the adjustment gain is limited to between 15 and 35. In this embodiment, a value of 25 is preferred to provide the most stable illumination robustness. The attenuation constant controlling the feature filtering slope is determined by performing exponential fitting regression analysis on the gradient distribution of the spatial information entropy density. Its reasonable value range is limited to between 0.4 and 1.2. In this embodiment, 0.8 is specifically selected. If the feature extraction threshold is below 0.4, it will lead to incomplete removal of redundant feature points in dense regions, causing computational overload. If it is above 1.2, it will cause excessive screening in sparse regions, thus destroying mesh connectivity. In sparse texture regions where the spatial information entropy density is below the preset lower limit, the processor automatically lowers the feature extraction threshold to increase the sampling point density, thereby ensuring the global connectivity of the subsequent topological mesh. In dense texture regions where the spatial information entropy density is above the preset upper limit, the processor automatically raises the feature extraction threshold to remove redundant feature points, compressing the feature dimension from the source and reducing the computational load of subsequent hash encoding. Using the above adaptively adjusted feature extraction threshold, the processor extracts the current key... In the keyframe, image feature points are obtained, and their pixel coordinates in the current two-dimensional image coordinate system are acquired. Specifically, the processor uses the FAST corner detection operator as the underlying feature response mechanism. By determining whether the gray-level difference of consecutive pixels in the circular neighborhood around the center pixel is greater than the feature extraction threshold, high-frequency local extrema points with physical corner attributes are selected. The processor obtains the convex hull boundary of each image feature point in the current two-dimensional image coordinate system, and divides the current keyframe into a set of non-overlapping triangles based on the convex hull boundary and the planar distribution position of each image feature point. The standard Delaunay triangulation algorithm is then executed to process the data. The processor constructs a spatial geometric mesh that covers the local image region of the current keyframe and satisfies the empty circle property. The topological connection edges in this spatial geometric mesh are attached with weight factors, which are negatively correlated with the planar Euclidean distance between nodes. This forms a soft spatial topological network of visual anchor points within a single frame. To cut off the global cascading flip caused by local node mutations, the processor divides the entire image into multiple independent local empty circle mesh sub-regions when performing Delaunay triangulation, using strong edge contours in the image as rigid constraint boundaries. When malicious tampering occurs in a sub-region, causing local topological reorganization, the rigid constraint boundary blocks the cross-region propagation of the mesh diagonal flip.

[0045] In the generated spatial geometric mesh, the processor retrieves adjacent triangle pairs with shared topological edges. For any retrieved adjacent triangle pair, the processor calculates the first area of ​​the first triangle and the second area of ​​the second triangle based on the two-dimensional pixel coordinates of its vertices, and calculates the ratio of the first area to the second area to determine the local topological ratio. According to the principle of projective geometric invariants, regardless of the global affine transformations such as translation, rotation, or proportional scaling of the video frame, the ratio of the areas of adjacent triangles with shared mesh edges always remains mathematically constant. However, once the video encounters malicious operations such as object erasure or local image splicing that cause a substantial break in the microscopic topological structure, the arrangement of visual semantic anchor points in a specific local area becomes heterogeneous, causing the local triangle connection relationships to be disrupted. Recombination causes a step-like discontinuous change in the local topological ratio. The processor traverses all adjacent triangle pairs in the spatial geometric grid, arranges the local topological ratios according to the preset spatial vertex index order, and combines them to generate an affine invariant feature vector corresponding to the current keyframe. In order to filter out high-frequency random noise caused by low bitrate encoding / decoding artifacts or slight image jitter, after generating the affine invariant feature vector, the processor applies a discrete cosine transform to the affine invariant feature vector, discards high-frequency components with spatial frequencies higher than the cutoff threshold, and extracts a low-frequency coefficient matrix representing the macroscopic topological deformation trend. The processor uses the locality-sensitive hashing algorithm to perform dimensionality reduction projection on the low-frequency coefficient matrix to generate a fixed-length binary fingerprint code, thereby outputting a topological verification fingerprint sequence with timestamp features.

[0046] In the temporal evolution analysis for determining the location of local tampering, the processor extracts keyframes from two adjacent frames along the video timeline, calculates the feature difference matrix between the first affine invariant eigenvector corresponding to the current keyframe and the second affine invariant eigenvector corresponding to the next keyframe. Simultaneously, the processor obtains the motion vectors of each image feature point between adjacent keyframes through optical flow tracing, calculates the pixel displacement deviation between adjacent keyframes, and establishes a sliding time window to achieve accurate target identification and eliminate interference from isolated high-frequency noise. The processor statistically analyzes the mean and standard deviation of the temporal deviation tensor within the sliding time window and calculates the real-time deviation response value corresponding to the current keyframe. Before executing subsequent logical operations, the processor, based on the spatial vertex index order when constructing the spatial geometric mesh, maps the two-dimensional pixel displacement deviation of the identified pulse jump point back to the topological node index position corresponding to the one-dimensional floating-point array, thereby generating a jump based on the one-dimensional coordinate system distribution of unified feature nodes. In the alignment and mapping implementation steps, since each non-zero element in the feature difference matrix uniquely corresponds to a pair of adjacent triangles with common edges in the spatial geometric grid, and each pair of adjacent triangles with common edges is topologically composed of 4 feature points as vertices, the processor executes the following association mapping rule: For each non-zero element position in the feature difference matrix whose absolute value exceeds the preset fluctuation tolerance, the processor automatically retrieves the global index of the 4 vertices contained in the corresponding pair of adjacent triangles with common edges, and synchronously retrieves the level state at the index of these 4 vertices in the jump mask. As long as at least one of these 4 vertices has a mask state that represents the instruction code 1 for pulse jump, the processor determines that the pulse jump point of the feature difference matrix element and the pixel displacement deviation amount is aligned in the spatial topology, and outputs the result of its logical AND operation as 1; otherwise, it outputs 0. In this way, the cross-dimensional point and surface feature precise collaborative adjudication is completed in the matrix element dimension of the difference matrix.The processor performs a logical AND operation on the distribution of non-zero elements in the feature difference matrix whose absolute values ​​exceed a preset fluctuation tolerance, and the pulse jump points of the pixel displacement deviation. This eliminates global synchronization displacement caused by legitimate image stabilization, thus determining the location of local tampering in the video to be checked. When the real-time deviation response value deviates from the mean by more than three times the standard deviation, the system generates a non-linear abnormal response, determining that the video segment corresponding to the sliding time window has been tampered with by frame skipping or frame insertion. After determining the location of the local tampering, the processor retrieves the reference fingerprint code of the original video and calculates the Hamming distance between the binary fingerprint code of the video to be checked and the reference fingerprint code. By identifying frame index positions where discontinuous step changes occur in Hamming distance and combining this with the distribution pattern of pixel displacement deviation, the timestamp information of the local tampering position and the coordinate information of the tampered area are structurally integrated to output an illegal editing judgment type verification report. This achieves automated locking of local image splicing, object erasure, frame extraction, and frame insertion. The logical decoupling and collaborative operation of the feature difference matrix and the temporal deviation tensor transform the microstructural fractures caused by local video splicing variations into deterministic abnormal response pulses. The video content anti-counterfeiting verification system enters a stable and highly sensitive locking operation state when facing high false alarm interference.

[0047] Example 2: This example provides a test group using the method of the present invention to conduct engineering verification of a video keyframe fingerprint verification method for illegal editing identification. The test group built a fingerprint verification platform based on a specific video processing chip. The processor used was a central processing unit with a core frequency of 2.5GHz and 8 physical cores, configured with 32GB of random access memory. The test stream was deployed on a general-purpose workstation with a 64-bit operating system. The test video set contained 1000 streaming video files with a duration of 30 minutes, a resolution of 1920×1080, and a frame rate of 30 frames per second, covering different technical types of image features such as motion, scenery, and interviews. The test group divided the test video set into an original group, a legal editing group, and a malicious editing group. The legal editing group scaled the original video proportionally, with the scaling ratio set to 0.5 to 2.0, the rotation angle set to 1 degree to 15 degrees, and applied image stabilization filtering. The malicious editing group erased objects or spliced ​​foreign image pixel blocks in local areas of the video, with the local tampering area accounting for 0.5% to 5.0% of the total area of ​​the current keyframe.

[0048] The specific steps of the processor executing the fingerprint verification method are as follows: The processor reads the keyframe sequence of the video to be verified. For the current keyframe, the processor extracts image feature points and obtains the pixel coordinates of each image feature point in the current two-dimensional image coordinate system. The number of image feature points extracted is dynamically adjusted according to the image texture complexity and is set at... One to Between each keyframe, the processor performs Delaunay triangulation on each image feature point based on pixel coordinates, constructing a spatial geometric mesh covering the local image region. The processor then retrieves adjacent triangle pairs with shared mesh edges within the spatial geometric mesh. Based on the pixel coordinates of each image feature point, it calculates the first area of ​​the first triangle and the second area of ​​the second triangle in each adjacent triangle pair, and calculates the ratio of the first area to the second area to generate a local topological ratio, denoted as [missing value]. The processor traverses all adjacent triangle pairs in the spatial geometric mesh, and according to the preset spatial vertex index order, concatenates all generated local topological ratios into a one-dimensional floating-point array to generate the affine invariant feature vector corresponding to the current keyframe, denoted as . In the legal editing group test, although the video frame underwent overall scaling or rotation, causing a linear transformation of the pixel coordinates of feature points, adjacent triangle pairs shared a grid edge, and their first and second areas changed proportionally under the affine transformation. Therefore, the local topological ratio calculated by the processor remained the same. Keeping numerical values ​​constant, affine invariant eigenvectors The cross-correlation coefficient is greater than No local tampering warning was triggered. In the test by the malicious editing group, the physical texture continuity of the local area was destroyed due to the erasure of local objects or the splicing of the screen. This caused the topological connection relationship of the spatial geometric mesh generated by Delaunay triangulation to break. The vertex coordinates of adjacent triangle pairs in the tampered area produced nonlinear relative displacement, causing the ratio of the first area to the second area to change abruptly.

[0049] Local topology ratios collected by the processor The resulting jump bias corresponds to the affine invariant eigenvector. In step five, when data anomalies occur in a specific index segment, the processor calculates the feature difference matrix between adjacent frames and extracts the temporal deviation tensor to determine the precise location of the local tampering. The test group statistically analyzes the operating indicators of the fingerprint verification platform on the full set of test videos. Experimental data shows that for the legitimate editing group, the false alarm rate of the system is 0.32%; for the malicious editing group, the system's targeted recognition accuracy rate is 99.15% for local image splicing and object erasure with an area greater than or equal to 1.0%. In the entire verification process, since the processor only performs linear algebraic operations on the topological grid area ratio of feature points, it does not need to perform pixel-by-pixel overlap domain comparison. The average time consumption for single-frame fingerprint generation and comparison is 4.15ms, which meets the high-throughput automated verification requirements of streaming media platform content review. The test results confirm the feasibility of the spatial topological invariance principle in isolating global geometric deformation noise and capturing micro-editing variations.

[0050] Example 3: This example combines Figures 1 to 2 This paper describes a video keyframe fingerprint verification method for illegal editing identification, such as... Figure 1 As shown, step S101 obtains the keyframe sequence of the video to be verified, then proceeds to step S102 to extract image feature points from the current keyframe and obtain the pixel coordinates of each image feature point in the current two-dimensional image coordinate system. Then, in step S103, Delaunay triangulation is performed on each image feature point based on the pixel coordinates to construct a spatial geometric mesh covering the local image region of the current keyframe. Finally, in step S104, adjacent triangle pairs with shared edges are retrieved from the spatial geometric mesh. Based on the pixel coordinates of each image feature point, the first area of ​​the first triangle and the second area of ​​the second triangle in each adjacent triangle pair are calculated. The local topological ratio is determined by extracting the ratio of the first area to the second area. Step S105 involves traversing all adjacent triangle pairs in the spatial geometric grid and arranging the local topological ratios according to the preset spatial index order. This generates the affine invariant feature vector corresponding to the current key frame. Finally, step S106 calculates the difference matrix of the affine invariant feature vectors between two adjacent key frames. The pixel displacement deviation of each image feature point between two adjacent key frames is calculated. The non-zero elements in the difference matrix whose absolute values ​​exceed the preset fluctuation tolerance are distributed and logically ANDed with the pulse jump points of the pixel displacement deviation to determine the local tampering position in the video to be checked.

[0051] like Figure 2 As shown, the low-frequency coefficient matrix is ​​converted into a 256-element feature vector and input to a linear centripetal projection layer. Inside this linear centripetal projection layer, the orthogonal random transformation matrix and the input data are combined to perform an inner product operation to project to a low-dimensional affine feature subspace, thereby outputting an intermediate projection vector of 32 floating-point numbers. This intermediate projection vector of 32 floating-point numbers enters the adaptive binary quantization layer to receive the threshold decision instruction sequence. Through conditional branch judgment, if the value of the current floating-point number is greater than or equal to zero, a high-level signal is output and interpreted as technical instruction code 1; if the value of the current floating-point number is less than zero, a low-level signal is output and interpreted as technical instruction code 0. The above quantization branch signal combination outputs a fixed-length 32-bit binary fingerprint code.

[0052] Example 4: When the method claimed in this invention faces non-uniform frame drops induced by drastic fluctuations in channel bandwidth of streaming media servers, as well as transient motion artifacts accompanying the video source itself, the system's timing deviation tensor is prone to random baseline drift, leading to false alarms. To improve the engineering adaptability of the fingerprint verification method, the processor needs to configure a specific physical carrier for the information processing stream. By controlling the memory addressing and instruction sequence within the computer system, the abstract judgment rules are transformed into discretized execution steps for targeted repair of spatiotemporal distortions. After the key frame sequence of the video to be verified is imported into memory, the processor allocates an independent storage node as the input to the initial state definition procedure. The object of the operation is a one-dimensional floating-point array containing the planar coordinates of adjacent triangles output by the previous steps. The functional specifications of the enabling environment require the processor's floating-point operation capability to be greater than or equal to 10 billion times per second to support millisecond-level real-time differential transformation. The first step of the execution logic of the process judgment quantization procedure is engineering timing decomposition. The processor reads the current key frame sequence according to the address order continuously allocated in the memory. The first affine invariant eigenvector corresponding to the keyframe and the second affine invariant eigenvector corresponding to the next keyframe are subjected to matrix subtraction in the hardware processing unit to generate a feature difference matrix. Simultaneously, the processor calls multi-threaded parallel computing instructions to initiate local feature block matching on the two-dimensional image pixel matrix of two adjacent keyframes, calculates the displacement vector of each pixel block along the horizontal and vertical axes of the image, and accumulates them to obtain the pixel displacement deviation between two adjacent keyframes. The second step is the tracing of data lineage and white-box mapping. The processor constructs a three-dimensional spatial geometric temporal tensor. The first and second dimensions of this tensor correspond to the row index and column index of the feature difference matrix, respectively, and the third dimension corresponds to the temporal addressing index of the continuous video frames. In order to transform this multi-dimensional composite feature into a single quantization conclusion for scalar determination, the processor executes the feature dimensionality reduction transformation rule: extract all coefficients of the three-dimensional spatial geometric temporal tensor within a preset sliding time window, calculate the statistical second moment of these coefficients deviating from the zero-level axis, and obtain the temporal deviation tensor of the current frame. This temporal deviation tensor represents the temporal evolution smoothness of the visual semantic anchor under the spatial topological common edge constraint.

[0053] The processor inputs the low-frequency coefficient matrix after discrete cosine transform into a locality-sensitive hash projection model. The specific internal topology and inference mechanism of this model are disclosed below. The model includes a linear centripetal projection layer and an adaptive binary quantization layer. In the linear centripetal projection layer, the low-frequency coefficient matrix is ​​flattened into a feature vector containing 256 elements and then performs an inner product operation with an orthogonal random transformation matrix pre-loaded into the cache. This orthogonal random transformation matrix is ​​generated by a preset pseudo-random algorithm during system initialization. Its pseudo-random generation seed is bound to the file hash feature value generated after decoding and de-identification of the original video to be verified, ensuring that multiple independent verification tasks for the same physical video can be accurately projected to an absolutely consistent affine feature subspace. The row vector dimension of the orthogonal random matrix is ​​completely aligned with the input feature vector. After the inner product operation, the topological relationships in the original high-dimensional space are mapped to the low-dimensional affine feature subspace, outputting an intermediate projection vector containing 32 floating-point numbers. In the actual verification process, through... The hash feature value is obtained through a 1:1 copyright verification or multi-target initial screening retrieval mechanism. When the video data stream to be verified is accessed by the system, the user terminal or streaming media platform will concurrently transmit the copyright file index tag declared by the video. The processor directly retrieves the corresponding original video metadata from the local fingerprint database based on the index tag, and synchronously reads the file hash feature value that was calculated and solidified when the original video was first registered and entered into the database. As the benchmark seed for the current verification task, the processor loads the file hash feature value into memory and inputs it into a preset pseudo-random algorithm. Thus, in the current session stage of system initialization, the orthogonal random transformation matrix is ​​dynamically reconstructed, realizing the absolute binding between the verification matrix and the target copyright video. In the adaptive binary quantization layer, the processor calls the threshold adjudication instruction sequence to check each floating-point number in the intermediate projection vector one by one. If the value of the current floating-point number is greater than or equal to zero, the corresponding register bit is driven to a high-level signal and parsed as the technical instruction code 1 in the business scenario.If the current floating-point number is less than zero, the corresponding register bit is driven low, parsed as the technical instruction code 0 in the business scenario. After traversal, the 32 register bits are combined to output a fixed-length 32-bit binary fingerprint code, completing the transition from continuous physical geometric features to a discrete digital fingerprint white-box structure. Addressing the blindness of threshold setting in frame extraction and insertion tampering judgment, this invention introduces an endogenous threshold calibration and dynamic evolution function. The judgment criterion is based on the statistical regularity of the sliding time window inherent in the system. During operation, the processor periodically updates the timestamp interval and motion vector of 10 consecutive video frames. Variance envelope calculation is performed to calibrate the intrinsic perturbation mean and standard deviation of the system in the current channel environment in real time. When the system faces transient spatiotemporal discontinuities caused by illegal editing and deliberate cut-off of images, the sudden change in motion vector generated by optical flow tracing at the break point will drive the timing deviation tensor to undergo a step, causing the real-time deviation response value calculated by the processor to produce a pulse jump. The processor compares the real-time deviation response value with the system's intrinsic standard deviation using a dynamic function. The trigger condition control logic is as follows: the processor determines whether the real-time deviation response value is greater than the preset fluctuation tolerance. The preset fluctuation tolerance is updated in real time according to the update formula, and the specific update formula is as follows: ,in, To preset the fluctuation tolerance, This represents the mean intrinsic perturbation value statistically obtained by the processor within the current sliding time window. The corresponding standard deviation is... The compensation coefficient is dynamically and adaptively adjusted. To prevent sudden tampering frames from entering the sliding time window and causing data pollution to the subsequent normal mean and standard deviation, thus triggering baseline adaptive drift, this invention configures a buffer snapshot isolation logic with anomaly blocking function in the processor memory. When the real-time deviation response value of a key frame is determined to be greater than the current preset fluctuation tolerance and triggers an anti-counterfeiting warning, the processor immediately locks and suspends the statistical update of the sliding time window. The abnormal response data corresponding to the tampered frame is forcibly prohibited from being written into the circular queue for calculating the mean and standard deviation. At this time, the sliding time window automatically copies and reuses the historical statistical snapshot within the previous safe steady-state period until the abnormal response pulse completely disappears and the real-time deviation response values ​​of three consecutive subsequent key frames all fall back to within the safe baseline. Only then does the processor release the suspended state and resume the data sliding statistics of the window, thereby eliminating the continuous blinding of the algorithm caused by malicious tampering in terms of mechanism.

[0054] To resolve the discrepancy between normal artifacts caused by rapid transient motion in video and malicious frame dropping / manipulation, a compensation coefficient is used. The global motion velocity is monotonically correlated with the average global displacement velocity in the current video sequence. The processor quantifies this global motion velocity by statistically analyzing the average pixel displacement amplitude of optical flow tracing at feature points across the entire image. Based on Monte Carlo measurements of a standard, untampered video library, the lower bound of the adjustment coefficient corresponding to the physical limit of the system hardware white noise under static or extremely low-speed backgrounds is determined. Simultaneously, the maximum tolerance extreme value that can effectively absorb nonlinear jitter errors under high-speed motion is derived as the upper bound of the adjustment coefficient. When the processor detects that the global motion velocity of the image is high, to avoid false alarms caused by stress noise, the compensation coefficient is adjusted. The value is adaptively and slightly adjusted upwards to 3.5; when the overall motion speed of the screen is static or low-speed, in order to capture single-frame-level micro-frame skipping, the compensation coefficient is adjusted upwards. The value was adjusted downwards to 2.8. Under the current operating conditions, the average value is... The measured value is 5.2, and the standard deviation is... The measured value is 1.5, which is the compensation coefficient after global velocity weighting. The value is set to 3.0, and the current preset fluctuation tolerance is calculated. The exact value is 9.7. When a video frame containing non-uniform frame dropping interference passes through, the real-time deviation response value it causes is 7.4, which does not exceed the current preset fluctuation tolerance of 9.7. The processor determines it as legitimate channel dissipation and performs downlink suppression, without releasing an alarm work order. However, when a video segment is maliciously tampered with by frame insertion, local topology reorganization and timeline breakage cause the real-time deviation response value to jump to 18.6, directly exceeding the preset fluctuation tolerance of 9.7. The system then releases a non-linear abnormal response signal and generates an anti-counterfeiting warning prompt on the background management terminal. This solves the interference of network fluctuation noise on the accuracy of tamper identification and completes the complete application closed loop from raw data processing at the sensing end to decision support at the execution end. When the processor traverses all adjacent triangle pairs in the spatial geometric grid and arranges the local topology ratios according to the preset spatial vertex index order to combine and generate the affine invariant feature vector corresponding to the current key frame, in order to avoid the video image undergoing global translation... After rotation transformation, the spatial vertex arrangement order becomes disordered and misaligned due to the reconstruction of the absolute pixel coordinate axis direction. The preset spatial vertex index order adopts the relative radial angle sorting rule based on the counterclockwise outward radiation of the convex hull topological center. The processor calculates the geometric center coordinates of the convex hull boundary based on the pixel coordinates of all extracted image feature points, establishes a local polar coordinate system with the geometric center coordinates as the origin, and calculates the polar angle and polar radius of the geometric centroid of each triangle in the local polar coordinate system. The processor prioritizes performing one-dimensional linear sorting on all adjacent triangle pairs in ascending order of polar angle. When encountering a special degenerate state with the same polar angle, the secondary sorting rule based on ascending polar radius is unconditionally activated. This locks the arrangement order of each local grid unit inside the grid onto the inherent geometric topological structure that changes synchronously with the image, so that the local topological ratios of each dimension in the final combined feature vector maintain character-level spatial position alignment before and after continuous affine transformation.

[0055] Example 5: Based on the keyframe fingerprint verification architecture of the video to be verified determined in the above examples, in order to improve the overall anti-counterfeiting detection accuracy and anti-interference stability of the video keyframe fingerprint verification method in dealing with real streaming media distribution, complex network transmission, and malicious local tampering, the method claimed in this invention is configured as a spatiotemporal topology verification procedure with defensive capabilities when implemented in actual engineering. This procedure is used to deeply decouple and target the global legitimate deformation and local malicious splicing caused by imperfect underlying data, transmission jitter, or video post-processing during high-throughput automated verification of the video to be verified. When constructing and comparing the intra-frame spatial geometric grid and affine invariant feature vectors of the video to be verified, it addresses the image local distortion caused by differences in video acquisition hardware or dynamic bitrate encoding of streaming media. To mitigate noise disturbances, before performing Delaunay triangulation, the processor performs a pre-processing operation of visual semantic anchor point purification and spatial mesh calibration. By extracting local features from the image pixel matrix of the current keyframe, the processor uses a preset non-maximum suppression rule to remove pseudo-feature points within a specific spatial neighborhood caused by sudden changes in illumination, retaining only visual semantic anchor points with stable planar physical coordinates in the current two-dimensional image coordinate system. After completing Delaunay triangulation and generating an intra-frame spatial topology map, the processor verifies the geometric topological continuity of the mesh by calculating the normal vector deviation between adjacent mesh units. This effectively decouples the global physical coordinate synchronization drift caused by overall image scaling from the topological connection line breakage caused by local object erasure during the low-level feature expression stage.

[0056] When calculating the local topological ratio of adjacent triangle pairs and generating affine-invariant topological feature vectors, to prevent computational overflow or dimensional breakage due to local polygon degradation, the processor employs an area feature optimization and normalization control logic based on shared topological edges. This logic defines any adjacent triangle pair in the spatial geometric mesh as including a first triangle and a second triangle, with their shared topological edge denoted as... The processor obtains the edges in the first triangle that do not belong to the shared topology. The first vertex's planar physical coordinates in the current two-dimensional image coordinate system At the same time, obtain the edges in the second triangle that do not belong to the shared topology. The second vertex's planar physical coordinates in the current two-dimensional image coordinate system The processor calculates the edge from the first vertex to the shared topology. First vertical distance And calculate the edge from the second vertex to the shared topology. The second vertical distance At this point, the first area of ​​the first triangle The second area of ​​the second triangle The ratio is converted to the first vertical distance. The second vertical distance The processor will use the first-order linear ratio to determine the local topological ratio. Expressed as: ,in, This represents the local topological ratio of the current pair of adjacent triangles. Indicates the edge from the first vertex to the shared topology. The first vertical distance, Indicates the edge from the second vertex to the shared topology. The second vertical distance, the processor's ratio of the extracted local topology. Perform boundary constraint checks, when detected When the value approaches 0, indicating the occurrence of mesh collinearity degradation, the processor will change the local topology ratio. The output is locked to a preset upper limit saturation constant, thereby avoiding dangling variables or non-numerical exceptions when generating a one-dimensional floating-point array.

[0057] When determining the location of local tampering by calculating the feature difference matrix and temporal deviation tensor between adjacent frames, to address the disruption to the continuity of the timeline caused by malicious frame extraction or dynamic local splicing, the method claimed in this invention extracts the spatiotemporal distortion tensor between adjacent keyframes. This transforms the static spatial feature comparison of isolated keyframes into a spatiotemporal topological evolution analysis of the video sequence on the timeline. After calculating the feature difference matrix between adjacent frames based on affine invariant topological feature vectors, the processor obtains the motion vectors of each visual semantic anchor point in the current keyframe sequence over time. Furthermore, it couples the spatial topological weights of the feature difference matrix with the temporal displacement difference in a multidimensional way to generate the spatiotemporal distortion tensor. Based on the principle of state superposition in spatiotemporal kinematics, the processor extracts the displacement difference vector between adjacent frames for each visual semantic anchor point. Simultaneously extract the topological weight change vector corresponding to the feature difference matrix. The processor performs coupled solution using a linear tensor outer product weighted algorithm: ,in, To characterize the continuous spatiotemporal distortion tensor, To reflect the pixel displacement difference vector that reflects the physical motion characteristics, To deduce the scalar topological weight change vector based on the area ratio of adjacent triangles, For tensor outer product operators, and The proportional weight constants of the fundamental motion vector response constrained by topological variations all have values ​​between 0 and 1, representing the spatiotemporal distortion tensor. Each component corresponds to the degree of geometric nonlinear variation of a specific local image region under spatiotemporal constraints. During the algorithm's execution, to eliminate the dimensional mismatch problem when directly adding a first-order vector and a second-order tensor, the processor first diagonally maps and expands the pixel displacement difference vector reflecting physical motion characteristics in the latter term before performing the addition operation. Specifically, it constructs a diagonal matrix with the same dimensions as the tensor outer product result of the former term, and sequentially fills each one-dimensional value of the pixel displacement difference vector into the main diagonal index position of this diagonal matrix. All other off-diagonal elements of the diagonal matrix are unconditionally cleared to zero, thus smoothly transforming and normalizing the first-order vector features into second-order diagonal matrix features. The processor then performs linear addition of the matrix generated after the outer product of the former term and the transformed diagonal matrix of the latter term at corresponding element positions, thereby completing the same-dimensional matrix-level fusion operation of the spatiotemporal constraints features. The processor presets a hardening trigger threshold for locking local tampering points. The hardening trigger threshold Instead of relying on fixed values ​​based on experience, the processor determines the values ​​based on the frame rate of the video being calibrated. With average image entropy The calculation rules for dynamic adaptive calibration are as follows: ,in, Indicates the hardening trigger threshold. This represents the preset scaling factor. This represents the average image entropy of the current keyframe. The frame rate of the video to be verified is represented by the preset proportional coefficient in the formula, which is a dimensionless control parameter. Its value is determined through a baseline tamper sensitivity test experiment under various common distribution frame rate conditions. The experimental procedure is as follows: In a baseline video segment with a frame rate of 30 frames per second and an average image entropy of 4.5, a local image splicing tampering with a proportion of 1% is injected. Through bidirectional approximation convergence test, the physical effective value range of this proportional coefficient is measured to be 1.2 to 2.8. In this embodiment, it is specifically fixed to 1.8. If the coefficient is lower than 1.2, it will cause excessive sensitivity to the small distortions generated by conventional electronic image stabilization and frequent false alarms. If the coefficient is higher than 2.8, it will make the spatiotemporal distortion threshold too high, causing the system to miss the hidden frame-skipping tampering within a short time window. The processor will convert the spatiotemporal distortion tensor into a variable. The components of each region in the hardening trigger threshold A comparison is performed one by one, and when the spatiotemporal distortion component of a specific local region exceeds the hardening trigger threshold... When this occurs, it indicates that the spatial geometric topological fracture and temporal trajectory of the region have produced a nonlinear step abrupt change that cannot be explained by affine transformation. The processor triggers the tampering warning mechanism and outputs a clear decision conclusion to the control backend of the video streaming platform's content review center. This decision conclusion specifies the local splicing position where the micro-editing variation occurs and the corresponding video timestamp range, thereby providing high-throughput automated verification decision support.

[0058] Example 6: When the fingerprint verification system faces a multi-source deployment scenario with different batches of video sources to be verified and heterogeneous distribution network nodes, the discreteness difference of the video sources in the early compression and encoding process causes non-uniform deviation between the frame rate and the initial motion vector baseline. In order to ensure that the fingerprint verification method has a definite starting point before being applied to the new deployment environment, the processor starts a standardized initial position parameter fine-tuning procedure before accessing the external video anti-counterfeiting verification data stream. During the power-on self-test phase, the processor loads a local verification package containing a standard motion calibration sequence into the memory address space. The standard motion calibration sequence consists of a matrix of 500 desensitized two-dimensional test frames that cover standard translation and frame rate changes. The internal time-series variance calculation instruction is called to calculate the virtual difference sequence of the entire process of the standard motion calibration sequence. Based on this, the latency residual of the system under this specific hardware architecture is extracted to complete the initial parameter alignment for field deployment.

[0059] The processor establishes a sliding time window to capture the output response of the standard motion calibration sequence after undergoing simulated affine deformation. By monitoring the convergence curve of the transform domain coefficients, the static sensitivity envelope of the fingerprint hash model is calibrated. The second-order time delay lag induced by register addressing delay during the operation is quantized into a constant operator. The constant operator is substituted into the threshold adjustment function as the basic background value. The mean and standard deviation of the disturbance obtained by the virtual difference sequence operation are retrieved to calculate the initial preset fluctuation tolerance of the system in the pre-calibration state. This is then solidified as the physical metric benchmark in the current deployment environment. The physical metric benchmark is used to provide a zero-point reference baseline when the feature difference matrix and the timing deviation tensor flow, and to eliminate hardware noise parasitic deviations. After the calibration procedure is completed, the system releases a low-level control signal indicating that the self-test has passed, automatically locking the calibration platform within the current parameter window. This ensures that the subsequently accessed video data stream follows the physical metric benchmark, guaranteeing the steady state of the automated anti-counterfeiting locking loop when reused across platforms.

[0060] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A video keyframe fingerprint verification method for illegal editing identification, characterized in that, Includes the following steps: Step S101: Obtain the keyframe sequence of the video to be verified; Step S102: Extract image feature points from the current keyframe and obtain the pixel coordinates of each image feature point in the current two-dimensional image coordinate system; Step S103: Perform Delaunay triangulation on each image feature point according to the pixel coordinates to construct a spatial geometric mesh covering the local image region of the current keyframe. Step S104: In the spatial geometric grid, retrieve adjacent triangle pairs with shared edges. Based on the pixel coordinates of each image feature point, calculate the first area of ​​the first triangle and the second area of ​​the second triangle in each adjacent triangle pair, and extract the ratio of the first area to the second area to determine the local topological ratio. Step S105: Traverse all adjacent triangle pairs in the spatial geometric grid, arrange the local topological ratios according to the preset spatial index order, and generate the affine invariant feature vector corresponding to the current keyframe. Step S106: Calculate the difference matrix of the affine invariant eigenvectors between two adjacent keyframes, calculate the pixel displacement deviation of each image feature point between two adjacent keyframes, and perform a logical AND operation between the distribution of non-zero elements in the difference matrix whose absolute value exceeds the preset fluctuation tolerance and the pulse jump points of the pixel displacement deviation to determine the local tampering location in the video to be checked.

2. The video keyframe fingerprint verification method for illegal editing identification according to claim 1, characterized in that, The steps to determine the location of local tampering include: obtaining the displacement difference components of each image feature point on the video time axis and the change in topological weight in the spatial geometric grid; determining the temporal deviation tensor, which characterizes the spatiotemporal topological continuity of the video sequence, by linearly weighting the displacement difference components and the change in topological weight; and using the abnormal response value of the temporal deviation tensor to determine illegal editing.

3. The video keyframe fingerprint verification method for illegal editing identification according to claim 1, characterized in that, The steps for extracting image feature points include: obtaining the texture complexity of each pixel region in the current keyframe; establishing a non-linear mapping relationship between texture complexity and feature extraction threshold; adjusting the feature extraction threshold of different pixel regions using the non-linear mapping relationship; reducing the feature extraction threshold to increase sampling density in regions where texture complexity is below a preset lower limit, and increasing the feature extraction threshold to remove redundant feature points in regions where texture complexity is above a preset upper limit.

4. The video keyframe fingerprint verification method for illegal editing identification according to claim 1, characterized in that, The steps for performing Delaunay triangulation include: obtaining the convex hull boundary of each image feature point in the current two-dimensional image coordinate system; dividing the current keyframe into a set of non-overlapping triangles based on the convex hull boundary and the distribution position of each image feature point, and constructing a spatial geometric mesh that satisfies the empty circle property.

5. A video keyframe fingerprint verification method for illegal editing identification according to claim 2, characterized in that, The steps for determining illegal editing also include: establishing a sliding time window, calculating the mean and standard deviation of the temporal deviation tensor within the sliding time window; calculating the real-time deviation response value corresponding to the current keyframe; and determining that the video segment corresponding to the sliding time window has been tampered with by frame skipping or interpolation when the real-time deviation response value deviates from the mean by more than 3 times the standard deviation.

6. The video keyframe fingerprint verification method for illegal editing identification according to claim 1, characterized in that, After generating the affine invariant feature vector, the process also includes: using the locality-sensitive hashing algorithm to perform dimensionality reduction projection on the affine invariant feature vector to generate a fixed-length binary fingerprint code.

7. A video keyframe fingerprint verification method for illegal editing identification according to claim 6, characterized in that, The steps for determining the location of local tampering also include: obtaining the reference fingerprint code of the original video; calculating the Hamming distance between the binary fingerprint code of the video to be checked and the reference fingerprint code; identifying the frame index position where the Hamming distance shows a non-continuous step change, and determining the tampering type in combination with the pixel displacement deviation.

8. A video keyframe fingerprint verification method for illegal editing identification according to claim 7, characterized in that, The types of tampering include partial image splicing, object erasure, frame extraction, and frame insertion; among them, partial image splicing and object erasure are identified by the numerical fluctuation of the local topology ratio, while frame extraction and frame insertion are identified by the numerical anomaly of the displacement differential component.

9. A video keyframe fingerprint verification method for illegal editing identification according to claim 7, characterized in that, After determining the location of the local tampering, the process also includes: retrieving the fingerprint database index corresponding to the reference fingerprint code, and outputting a verification report containing the timestamp information of the local tampering location and the coordinate information of the tampered area.

Images