Video content protection method and device based on multi-channel image feature security mark

Abstract

The invention discloses a video content protection method and device based on a multi-channel image feature security mark, and relates to the technical field of video content security management, and the method comprises the steps: obtaining an original video, and checking the original video; acquiring a video frame image of the audited video, and performing fusion, dimension reduction, compression and encryption processing on multichannel high-order features of the video frame image through a perceptual hash algorithm, an interpolation hash algorithm, a compression coding principle and a national cipher SM4 algorithm to generate a dynamic security mark; embedding the dynamic security mark into the video frame image to generate an approved video; and extracting the dynamic security mark of the currently played video, and checking the consistency between the feature information extracted from the dynamic security mark and the feature information of the played video to judge whether the video is compliant or not. The video content is protected, the security of the video is improved, the compliance check of the video content is realized, and the accuracy and efficiency of compliance judgment of the video content are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video content security management, in particular to a video content protection method and device based on multi-channel image feature security marking. BACKGROUND

[0002] With the rapid development of digitalization, informatization and intelligentization technology, large-scale deployed electronic advertising screens are facing increasingly severe information security risks, which may be used by illegal persons to spread false advertisements, malicious links and other illegal content, seriously damaging the corporate image, social credibility and even endangering national security. Moreover, due to the extremely fast speed of video content creation and dissemination, infringement phenomena occur frequently, and the importance of copyright protection is increasingly prominent.

[0003] Therefore, in order to ensure the safety and compliance of video content, a video content protection method based on multi-channel image feature security marking is urgently needed. SUMMARY

[0004] The purpose of the present application is to provide a video content protection method and device based on multi-channel image feature security marking. The present application not only generates dynamic security markers, but also embeds dynamic security markers in videos to protect video content and improve video security.

[0005] To achieve the above purpose, the present application provides a video content protection method based on multi-channel image feature security marking, comprising the following steps: S1, obtaining an original video, auditing the original video, and obtaining a passed video; S2, obtaining a video frame image of the passed video; the video frame image has multi-channel high-order features; S3, fusing, reducing dimension, compressing and encrypting the multi-channel high-order features of the video frame image through a perceptual hash algorithm, an interpolation hash algorithm, a compression encoding principle and a SM4 algorithm, to generate a dynamic security marker; S4, embedding the dynamic security marker in the video frame image according to the discrete cosine transform principle, the relative value quantization mechanism, the multi-voting mechanism and the inverse discrete cosine transform principle, and encoding to generate an approved video; S5, extracting the dynamic security marker of the currently played video, and verifying the consistency of the feature information extracted from the dynamic security marker and the feature information of the currently played video, to determine whether the video is compliant.

[0006] Preferably, the original video is audited in S1 to obtain a passed video, specifically including: According to an AI algorithm model based on deep learning technology, whether the original video has sensitive, harmful, illegal and infringing content is pre-audited once to obtain a pre-auditing result: If the original video does not contain sensitive, bad, illegal and infringing content in the pre-audit result, the pre-audit is in compliance; otherwise, the pre-audit is not in compliance; Regardless of whether the pre-audit result is in compliance or not, a secondary artificial review of the pre-audit result is required; If the review is passed, it is determined that the video is in compliance; If the review is not passed, it is determined that the video is not in compliance.

[0007] Preferably, in S3, the multi-channel high-order features of the video frame image are fused, dimensionally reduced, compressed and encrypted by the perceptual hash algorithm, the interpolation hash algorithm, the compression encoding principle and the SM4 algorithm of the national cryptography, to generate a dynamic security mark, which specifically includes: The video frame image is subjected to binarization processing and filtering operation to obtain a video frame processing image, and the video frame processing image is equally divided into a plurality of block images; According to the perceptual hash algorithm, perceptual hash feature extraction is performed on the plurality of block images and the video frame image respectively, and the perceptual hash values of the plurality of block images and the perceptual hash value of the video frame image are calculated; According to the interpolation hash algorithm, interpolation hash feature extraction is performed on the video frame image, and the interpolation hash value of the video frame image is calculated; According to the compression encoding principle, the perceptual hash value and the interpolation hash value are combined to obtain a first combination value of the perceptual hash and the interpolation hash, and the first combination value is encoded into a two-dimensional code; According to the SM4 algorithm of the national cryptography, the two-dimensional code is encrypted to obtain a dynamic security mark.

[0008] Preferably, in S4, the dynamic security mark is embedded into the video frame image according to the discrete cosine transform principle, the relative value quantization mechanism and the inverse discrete cosine transform principle, and an approved video is generated by encoding, which specifically includes: The spatial domain luminance channel of the original video frame image is obtained; According to the discrete cosine transform principle, the spatial domain luminance channel is converted into frequency domain to obtain a Y channel DCT coefficient matrix; Based on the relative value quantization mechanism, a coefficient pair is selected in the Y channel DCT coefficient matrix according to the coefficient selection rule; Based on the relative value quantization mechanism, the coefficient values of the selected coefficient pair are modified, when embedding bit 1, the relative difference value between coefficient 1 and coefficient 2 of the coefficient pair is greater than or equal to the embedding intensity; when embedding bit 0, the relative difference value between coefficient 2 and coefficient 1 of the coefficient pair is greater than or equal to the embedding intensity; through the above operation, the embedding of the dynamic security mark is realized. In addition, a multi-voting mechanism is adopted to realize the redundant embedding of the dynamic security mark (three different coefficient pairs are selected), and finally the DCT coefficient matrix embedded with the dynamic security mark is obtained; Based on the principle of inverse discrete cosine transform, the frequency domain coefficient matrix with embedded dynamic security tags is transformed in the spatial domain to obtain the video frame image with embedded dynamic security tags. Video frames with embedded dynamic security tags are encoded to generate approved videos.

[0009] Preferably, the frequency domain expression of the luminance channel is: ; ; in, For frequency index, The frequency domain coefficient matrix, Image block size, The normalization coefficient is... , These are the pixel position coordinates in the spatial domain. In the spatial domain ( , The pixel value at the location. Using the basis function, the signal is decomposed into a superposition of cosine waves of different frequencies.

[0010] The spatial domain expression for the brightness channel of the video frame image with embedded dynamic security tags is: ; in, For embedding dynamic security tags in video frame image luminance channels ( , The pixel value at the location. The frequency domain coefficient matrix of the luminance channel of the video frame image to embed dynamic security tags. Image block size, The normalization coefficient is... These are basis functions.

[0011] Preferably, the coefficient pair selection rules specifically include: selecting mid-frequency coefficient pairs, coefficient pairs with a difference ratio range of (1~1.2), and mid-frequency coefficient pairs in the horizontal and vertical directions from the frequency domain coefficient matrix.

[0012] Preferably, in step S5, a dynamic security tag of the currently playing video is extracted, and the consistency between the feature information extracted from the dynamic security tag and the feature information of the currently playing video is verified to determine whether the video is compliant. This specifically includes: The program acquires the currently playing video frame image and divides it into blocks to obtain several video block images. For each video block image, according to the principle of discrete cosine transform, the video block image luminance channel (Y channel) is converted from spatial domain to frequency domain to obtain a Y channel DCT coefficient matrix; Based on the Y channel DCT coefficient matrix, according to the relative difference value of the embedded coefficient pair, the video dynamic security mark is extracted, and finally the redundant multiple dynamic security marks are voted and judged by the judgment mechanism to obtain the final extracted dynamic security mark information; The extracted dynamic security mark is decrypted to obtain a second combined value; the second combined value includes a global perceptual hash value and an interpolation hash value and a local perceptual hash value; According to the perceptual hash algorithm, the perceptual hash feature of the current playing video block image is extracted to obtain the perceptual hash value of the current playing video block image; According to the perceptual hash algorithm and the interpolation hash algorithm, the perceptual hash feature and the interpolation hash feature of the current playing video frame image are extracted to obtain the perceptual hash value and the interpolation hash value of the current playing video frame image; Based on the local perceptual hash value of the current playing video frame image extracted by the frame image and the local perceptual hash value extracted by the dynamic security mark, the Hamming distance of the local perceptual hash value is calculated as the first similarity according to the Hamming distance calculation formula; The first similarity includes the Hamming distance of several perceptual hash values; Based on the global perceptual hash value of the current playing video frame image extracted by the frame image and the global perceptual hash value extracted by the dynamic security mark, the Hamming distance of the global perceptual hash value is calculated as the second similarity according to the Hamming distance calculation formula; Based on the interpolation hash value of the current playing video frame image extracted by the frame image and the interpolation hash value extracted by the dynamic security mark, the Hamming distance of the global interpolation hash value is calculated as the third similarity according to the Hamming distance calculation formula; Based on the consistency judgment mechanism, according to the first similarity, the second similarity and the third similarity, the playing video is verified to judge whether the video is compliant.

[0013] Preferably, based on the consistency judgment mechanism, according to the first similarity, the second similarity and the third similarity, the playing video is verified in real time to judge whether the video content is compliant, specifically including: According to the first similarity, the first similarity standard deviation of each playing video block image is calculated, and the first similarity of each playing video block image is sorted in ascending order to obtain a first similarity sequence; Get the first similarity standard deviation threshold, the second similarity threshold and the third similarity threshold, combine the first similarity standard deviation, the second similarity and the third similarity, and make a multi-level judgment to determine whether the video content is compliant.

[0014] Preferably, the first similarity standard deviation threshold value, the second similarity threshold value and the third similarity threshold value are obtained, and a multi-level judgment is performed in combination with the first similarity standard deviation, the second similarity and the third similarity to determine whether the video content is compliant, specifically including: Global anomaly judgment: If the judgment condition of the global anomaly is met, the video global anomaly is played, and the violation frame is reported to the management platform; the judgment condition of the global anomaly includes that the first similarity standard deviation is higher than the corresponding first similarity standard deviation threshold value, the second similarity is lower than the corresponding second similarity threshold value, and the third similarity is lower than the corresponding third similarity threshold value; If the judgment condition of the global anomaly is not met, the non-global anomaly of the played video is judged, and the local anomaly judgment is performed; Local anomaly judgment: If the judgment condition of the local anomaly is met, the local anomaly of the played video is judged, and the violation frame is reported to the management platform; the judgment condition of the local anomaly includes that the first similarity of the first video block image in the first similarity sequence is lower than the corresponding first similarity single-point threshold value, the first similarity of the first two video block images in the first similarity sequence is lower than the corresponding first similarity two-point threshold value, and the first similarity of the first four video block images in the first similarity sequence is lower than the corresponding first similarity four-point threshold value; If the judgment condition of the local anomaly is not met, the non-local anomaly of the played video is played, and the video content is determined to be compliant.

[0015] The application also provides a video content protection device based on a multi-channel image feature security mark, which comprises: A first acquisition module is configured to acquire an original video, and to obtain a passed video by auditing the original video; A second acquisition module is configured to acquire a video frame image of the passed video; the video frame image has multi-channel high-order features; A mark generation module is configured to fuse, reduce dimensions, compress and encrypt the multi-channel high-order features of the video frame image by using a perceptual hash algorithm, an interpolation hash algorithm, a compression encoding principle and a SM4 algorithm, and to generate a dynamic security mark; A mark embedding module is configured to embed the dynamic security mark into the video frame image according to a discrete cosine transform principle, a relative value quantization mechanism and a discrete cosine inverse transform principle, and to encode the video frame image to generate an approved video; A verification module is configured to extract the dynamic security mark of a played video, and to verify the consistency between the feature information extracted from the dynamic security mark and the feature information of the played video, so as to determine whether the video is compliant.

[0016] In summary, compared with the prior art, the video content protection method and device based on the multi-channel image feature security mark has the following beneficial effects: (1) The dynamic security mark is generated by using the perceptual hash algorithm, the interpolation hash algorithm, the compression encoding principle and the SM4 algorithm, and then the dynamic security mark is embedded in the original video and encoded to form the approved video according to the discrete cosine transform principle, the relative value quantization mechanism and the inverse discrete cosine transform principle, so that the security mark information is difficult to crack under limited cost, the protection of the video content is realized, the security of the video is improved, the content security management of the public electronic screen can be realized, the security risks such as vulnerability exploitation, malicious replacement, virus infection, wiretapping attack, near-field screen projection and the like can be effectively resisted, and the video content can be managed in the whole life cycle such as video generation, transmission, use and destruction.

[0017] (2) The dynamic security mark is extracted and the similarity degree measurement decision mechanism is used to realize the compliance inspection of the video content, and the accuracy and efficiency of the video content compliance decision are greatly improved.

[0018] (3) The embedding strategy combining the relative value quantization mechanism and the voting mechanism can effectively resist the attacks such as cropping, compression and rotation, and eliminate the error accumulation caused by the transmission and multiple encoding and decoding operations, so that the correct extraction of the dynamic security mark is ensured, and the reliability of the video content protection system is improved.

[0019] The technical method of the present application will be further described in detail below by means of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the video content protection method based on the multi-channel image feature security mark of the present application is shown in the figure. Figure 2 The generation and embedding flowchart of the dynamic security mark of the present application is shown in the figure. Figure 3 The extraction and verification flowchart of the dynamic security mark of the present application is shown in the figure. Figure 4 The video content consistency decision logic diagram of the present application is shown in the figure. Figure 5 The module diagram of the video content protection device based on the multi-channel image feature security mark of the present application is shown in the figure. DETAILED DESCRIPTION

[0021] The technical method of the present application will be further described in detail below by means of the accompanying drawings and examples. It should be noted that: unless otherwise specifically stated, the relative arrangement, numerical expression and numerical value of the components and steps set forth in these examples do not limit the scope of the present application.

[0022] The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the application or its application and uses.

[0023] Techniques, systems, and devices known to those of ordinary skill in the relevant art can not be discussed in detail, but should be considered as part of the specification where appropriate.

[0024] In all of the examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as a limitation. Thus, other examples of the exemplary embodiments can have different values.

[0025] Technical terms used in the present application are explained as follows: Discrete Cosine Transform (DCT) is a mathematical transform method that converts signals from time domain (or spatial domain) to frequency domain, mainly used in data compression and feature extraction fields.

[0026] Inverse Discrete Cosine Transform (IDCT) is a core algorithm that reconstructs signals from frequency domain back to time domain / spatial domain. As the inverse operation of DCT, IDCT realizes the visual reconstruction of compressed data in image / video decoding.

[0027] Perceptual Hashing (pHash) is a frequency domain-based perceptual hashing algorithm that maps multimedia content (such as images, audio) into compact hash values. It generates a unique "fingerprint" by capturing the perceptual features of the content, which is used to efficiently compare the similarity of the content. Its core feature is perceptual robustness - it keeps the hash value stable for operations that do not affect subjective perception (such as scaling, compression), while being sensitive to content tampering.

[0028] Interpolation Hash (iHash) is a spatial domain-based perceptual hashing algorithm. In contrast to pHash, its core feature is to generate hash values directly through image scaling interpolation and pixel-level comparison, with a significant computational efficiency advantage. It is sensitive to small changes in image content (such as edges, textures), but slightly less robust than pHash.

[0029] Quick Response Code (QR code) can store digital, alphabetic, binary data, and even Chinese characters and other information. QR code has the characteristics of high capacity storage, fast scanning, strong error correction ability, and multi-directional readability.

[0030] Majority Voting Mechanism refers to a robustness enhancement strategy, which repeatedly embeds the same watermark information into multiple transform domain coefficients of the host media, and determines the final watermark by counting the high frequency occurrence value of the decoding results at each position during extraction. The core is to improve the attack resistance through spatial / frequency redundancy.

[0031] Mid-frequency region: the middle region of the DCT coefficient matrix between the low-frequency region and the high-frequency region, which reflects the structured texture and medium-scale features in the image (such as regular texture and medium-contrast edges), and its energy contribution and perceptual importance are between low-frequency and high-frequency. Embedding watermark in the mid-frequency region can balance robustness and imperceptibility.

[0032] High-frequency region: the lower right corner of the DCT coefficient matrix, which saves detailed texture (such as edges and noise), and is often quantized and discarded.

[0033] Low-frequency region: the upper left corner of the DCT coefficient matrix, which saves the main structure and contour information of the image.

[0034] As shown in Figure 1 The present application provides a video content protection method based on multi-channel image feature security marking, which comprises the following steps: Step S1, obtaining an original video, auditing the original video, and obtaining a passed video.

[0035] Further, the original video is audited in step S1 to obtain a passed video, and the specific content includes: According to the AI algorithm model based on deep learning technology, whether the original video contains sensitive, harmful, illegal and infringing content is pre-audited once to obtain a pre-auditing result: If the original video does not contain sensitive, harmful, illegal and infringing content in the pre-auditing result, the pre-auditing is in compliance; otherwise, the pre-auditing is not in compliance. Whether the pre-auditing result is in compliance or not, the pre-auditing result needs to be audited twice by artificial; If the recheck is passed, it is determined as a compliant video; If the recheck is not passed, it is determined as a non-compliant video.

[0036] It can be understood that the pre-auditing can reduce the workload of the auditing personnel. Although the pre-auditing result needs to be audited twice by artificial, the auditing effort is different. The video of pre-auditing in compliance can be easily audited, and the video of pre-auditing not in compliance needs to be discriminated against the non-compliant result set in the video to determine whether it is illegal.

[0037] In this invention, when a user uploads an original video file to the intelligent content review server, the intelligent content review server uses a large model and dedicated algorithms to check the content of each frame of the video. It supports multiple AI detection algorithm models such as inappropriate images, sensitive text, and sensitive faces. Then, the intelligent content review server returns the intelligent pre-review result. Subsequently, a professional review team reviews the AI ​​review result and gives the final review result based on national laws and regulations, industry standards, and the platform's own rules.

[0038] Step S2: Obtain video frame images of the approved videos. These video frame images have multi-channel high-order features.

[0039] like Figure 2 As shown, the generation and embedding process of the dynamic security mark in this invention is as follows: Upload the original video, read the original video frame, generate a hidden image (encrypted dynamic security mark), perform DCT transformation to convert the current image frame into a frequency domain matrix, embed the hidden image (encrypted dynamic security mark) into the DCT coefficient matrix, then perform inverse DCT transformation, write it into the image matrix, then convert it back into an image frame and write it into the video to generate a new approved video, and provide a download link. The detailed process is as follows: Steps 3-4.

[0040] Step S3: Through perceptual hashing algorithm, interpolation hashing algorithm, compression coding principle and national cryptographic SM4 algorithm, the multi-channel high-order features of video frame image are fused, reduced in dimension, compressed and encrypted to generate dynamic security tags.

[0041] Furthermore, the specific content of step S3 includes: Step S31: Perform binarization and filtering operations on the video frame image to obtain the processed video frame image, and divide the processed video frame image into several block images.

[0042] In one specific embodiment of the present invention, the video frame processing image is divided into nine blocks.

[0043] Step S32: Based on the perceptual hash algorithm, perform perceptual hash feature extraction on several block images and video frame images respectively, and calculate the perceptual hash values ​​of several block images and video frame images.

[0044] Step S33: Extract interpolation hash features from the video frame image according to the interpolation hash algorithm, and calculate the interpolation hash value of the video frame image.

[0045] Step S34: Based on the compression coding principle, combine the perceptual hash value and the interpolated hash value to obtain the first combined value of the perceptual hash and the interpolated hash, and encode the first combined value into a QR code.

[0046] In a specific embodiment of the present invention, the first combined value includes: 10 pHash values ​​and 1 iHash value.

[0047] Step S35: Apply the national cryptographic SM4 algorithm and use a preset key to encrypt the QR code to obtain a dynamic security tag.

[0048] Step S4: Based on the principles of discrete cosine transform, relative value quantization mechanism, multi-voting mechanism, and inverse discrete cosine transform, embed dynamic security tags into video frame images and encode them to generate approved videos.

[0049] Furthermore, the specific content of step S4 includes: Step S41: Obtain the spatial domain luminance channel (Y channel) of the original video frame image.

[0050] Step S42: Based on the principle of discrete cosine transform, the spatial domain luminance channel is transformed into the frequency domain to obtain the DCT coefficient matrix of the luminance channel (Y channel). This invention uses the frequency domain representation of the Y channel as the target for embedding dynamic security markers.

[0051] The frequency domain expression for the luminance channel is: ; ; in, For frequency index, The frequency domain coefficient matrix, Image block size, The normalization coefficient is... , These are the pixel position coordinates in the spatial domain. In the spatial domain ( , The pixel value at the location. Using the basis function, the signal is decomposed into a superposition of cosine waves of different frequencies.

[0052] Step S43: Based on the relative value quantization mechanism, according to the coefficient pair selection rules, select coefficient pairs in the Y channel DCT coefficient matrix as shown in Table 1, for example: coefficient pair (5,0)-(2,3).

[0053] The specific rules for selecting coefficient pairs include: (1) Focusing on the mid-frequency region: Selecting mid-frequency coefficient pairs in the frequency domain coefficient matrix, avoiding high-frequency regions (which are easily compressed and discarded) and low-frequency regions (which are visually sensitive); (2) Energy stability: Selecting coefficient pairs with similar quantization step sizes (such as coefficient pairs with a difference ratio between 1 and 1.2) to ensure that the relative size relationship is stable after compression; (3) To complementarity: Selecting mid-frequency coefficient pairs in the horizontal and vertical directions, such as (3,4) and (4,3), to improve the ability to resist directional attacks (such as rotation).

[0054] Table 1 DCT standard luminance quantization matrix matrix

[0055] Step S44, based on the relative value quantization mechanism, modifying the coefficient values of the selected coefficient pair, when embedding bit 1, making the relative difference value of coefficient pair coefficient 1 and coefficient 2 >= embedding strength; when embedding bit 0, making the relative difference value of coefficient pair coefficient 2 and coefficient 1 >= embedding strength; through the above operation, the embedding of dynamic security mark is realized. In addition, the redundant embedding of dynamic security mark is realized by adopting the multi-voting mechanism (selecting three different coefficient pairs), and finally the DCT coefficient matrix embedded with dynamic security mark is obtained.

[0056] In the present application, the relative difference value represents the relative value size relationship of the coefficient pair, and expresses the dynamic security mark. In addition, by redundantly embedding the dynamic security mark information and adopting the multi-voting mechanism for security mark information extraction, the goals of low error rate, anti-geometric attack and enhanced invisibility can be realized in the case of existing quantization value difference. The embedding strategy combining the frequency domain relative value quantization mechanism and the voting mechanism realizes the invisibility, robustness and reversibility of the security mark information.

[0057] In a specific embodiment of the present application, the embedding strength value is 200, and taking the coefficient pair (5, 0)-(2, 3) as an example, if the embedded bit is "1", the corresponding DCT coefficient matrix coefficient value is modified, so that DCT(5, 0)-DCT(2, 3) is greater than or equal to 200, which meets the embedding strength requirement.

[0058] Step S45, according to the principle of inverse discrete cosine transform, the frequency domain coefficient matrix embedded with dynamic security mark is converted into spatial domain to obtain the video frame image embedded with dynamic security mark.

[0059] The present application effectively resolves the risk of code stealing and image local replacement by closely binding the collected multi-channel frame image features with the security mark.

[0060] The Y channel spatial domain expression of the video frame image embedded with dynamic security mark is: ; Wherein, is the pixel value of the Y channel of the video frame image embedded with dynamic security mark, , is the frequency domain coefficient matrix of the luminance channel of the video frame image embedded with dynamic security mark, is the image block size, is the normalization coefficient, is the base function. is the base function.

[0061] Step S46: Encode the video frame image with embedded dynamic security tags to generate an approved video.

[0062] Step S5: Extract the dynamic security tag of the currently playing video and verify the consistency between the feature information extracted from the dynamic security tag and the feature information of the currently playing video to determine whether the video is compliant.

[0063] This invention Figure 3 The specific process for watermark extraction and verification is as follows: The terminal device collects the video content being played, obtains video frames, then segments the video frames into blocks, converts the luminance channel of each block to the frequency domain to obtain the DCT coefficient matrix, then extracts and decrypts the security mark information by embedding the relative value relationship of the coefficient pairs. If no security mark is extracted, a violation frame is reported to the management platform; if extraction is successful, local and global image features are extracted from the video frame image, and image feature verification is performed by calculating the similarity between the feature information in the security mark and the extracted feature information. If verification fails, a violation frame is reported to the management platform; if verification is successful, the next frame of the playing video is collected and verified.

[0064] Furthermore, such as Figure 3 The specific content of step S5 shown includes: Step S51: Obtain the playback video frame image and perform block processing on the playback video frame image to obtain several video block images.

[0065] Step S52: For each video block image, according to the principle of discrete cosine transform, the Y channel of the video block image is transformed from the spatial domain to the frequency domain to obtain the Y channel DCT coefficient matrix.

[0066] Step S53: Based on the Y-channel DCT coefficient matrix, extract dynamic security tags according to the relative differences between the embedded coefficient pairs. If dynamic security tag extraction fails, report the violation frame.

[0067] Step S54: Decrypt the dynamic security tag to obtain the second combined value. The second combined value includes: the perceptual hash value and the interpolation hash value of the approved video frame image. The perceptual hash value of the approved video frame image includes: a global perceptual hash value and a local perceptual hash value.

[0068] Step S55: Based on the perceptual hash algorithm, perform perceptual hash feature extraction on the video block images to obtain the perceptual hash value of the video block images.

[0069] Step S56: Based on the perceptual hash algorithm and the interpolation hash algorithm, perform perceptual hash feature extraction and interpolation hash feature extraction on the video frame image to obtain the perceptual hash value and interpolation hash value of the video frame image.

[0070] Step S57, based on the local perception hash value and the local perception hash value extracted from the dynamic security mark, the Hamming distance of the local perception hash value is calculated as the first similarity according to the Hamming distance calculation formula. The first similarity includes the Hamming distance of several perception hash values.

[0071] Step S58, based on the global perception hash value and the global perception hash value extracted from the dynamic security mark, the Hamming distance of the global perception hash value is calculated as the second similarity according to the Hamming distance calculation formula.

[0072] Step S59, based on the interpolation hash value and the interpolation hash value extracted from the dynamic security mark, the Hamming distance of the global interpolation hash value is calculated as the third similarity according to the Hamming distance calculation formula.

[0073] It should be noted that the perception hash and interpolation hash in the application obtain the similarity by calculating the Hamming distance. The Hamming distance calculation formula between hash values is as follows: ; Among them, dis HamDistance() is the Hamming distance function between hash values, h, h' and h are the hash values for comparison, h′ is the similarity. The calculation formula shows that when the Hamming distance is less than or equal to 10, it is considered to be the same target. sim

[0074] Step S510, based on the consistency determination mechanism, the first similarity, the second similarity and the third similarity are used to verify the played video, and it is judged whether the played video content is compliant.

[0075] The application determines whether the video content is tampered by the consistency determination mechanism.

[0076] Further, the specific content of step S510 includes: According to the first similarity, the first similarity standard deviation of each played video block image is calculated, and the first similarity of all video block images is sorted in ascending order to obtain a first similarity sequence.

[0077] Get the first similarity standard deviation threshold, the second similarity threshold and the third similarity threshold, and combine the first similarity standard deviation, the second similarity and the third similarity to make a multi-level judgment to determine whether the video content is compliant, such as Figure 4 ​As shown, the input data of the consistency judgment algorithm is: the first similarity sequence, the second similarity, the third similarity, the first similarity standard deviation threshold, the first similarity single-point threshold, the first similarity two-point threshold, the first similarity four-point threshold, the second similarity threshold, and the third similarity threshold, and the output is the consistency check result. The specific content is as follows: The first similarity standard deviation threshold, the second similarity threshold, and the third similarity threshold are obtained, combined with the first similarity standard deviation, the second similarity, and the third similarity, to perform multi-level judgment to determine whether the video content is compliant, specifically including: Global anomaly judgment: If the global anomaly judgment condition is met, the video global anomaly is played, and the violation frame is reported to the management platform; the global anomaly judgment condition includes that the first similarity standard deviation is higher than the corresponding first similarity standard deviation threshold, the second similarity is lower than the corresponding second similarity threshold, and the third similarity is lower than the corresponding third similarity threshold; If the global anomaly judgment condition is not met, the video is judged to be non-global anomaly, and local anomaly judgment is performed; Local anomaly judgment: If the local anomaly judgment condition is met, the video local anomaly is judged, and the violation frame is reported to the management platform; the local anomaly judgment condition includes that the first similarity of the first video block image in the first similarity sequence is lower than the corresponding first similarity single-point threshold, the first similarity of the first two video block images in the first similarity sequence is lower than the corresponding first similarity two-point threshold, and the first similarity of the first four video block images in the first similarity sequence is lower than the corresponding first similarity four-point threshold; If the local anomaly judgment condition is not met, the video is non-local anomaly, and the video content is determined to be compliant.

[0078] The image feature processing algorithm, digital watermarking technology, and information hiding technology are used to generate a security mark bound with the original video frame image, the security mark bound with the video frame feature is embedded in the original video, and the security mark is verified in real time to achieve the goal of protecting the video content security.

[0079] The video content protection device based on the multi-channel image feature security mark provided by the application has the advantages that: Figure 5 As shown, it comprises: A first acquisition module is configured to acquire an original video, audit the original video, and obtain a passed video.

[0080] A second acquisition module is configured to acquire video frame images of the passed video; the video frame images have multi-channel high-order features.

[0081] The mark generation module is used for fusing, reducing dimension, compressing and encrypting the multi-channel high-order features of the video frame image through a perceptual hashing algorithm, an interpolation hashing algorithm, a compression coding principle and a SM4 algorithm of the national cryptography.

[0082] The mark embedding module is used for embedding the dynamic security mark into the video frame image according to a discrete cosine transform principle, a relative value quantization mechanism and a discrete inverse cosine transform principle, and encoding to generate an approved video.

[0083] The verification module extracts the dynamic security mark of the played video and verifies the consistency of the feature information extracted from the dynamic security mark and the feature information of the played video, so as to judge whether the video is compliant.

[0084] The video content protection system based on the dynamic security mark in the application comprises a distributed video content security terminal and a video content security management platform, and a technical closed loop of "feature extraction, dynamic security mark generation, hidden embedding and real-time verification" is established.

[0085] Finally, it should be noted that: the above examples are only used to illustrate the technical method of the application, not to limit it, although the application has been described in detail with reference to the preferred embodiment, those skilled in the art should understand that: it can still modify or replace the technical method of the application, and these modifications or equivalent replacements cannot make the modified technical method deviate from the spirit and scope of the technical method of the application.

Claims

1. A video content protection method based on multi-channel image feature security tagging, characterized in that, Includes the following steps: S1. Obtain the original video, review the original video, and obtain the approved video; S2. Obtain video frame images of the approved videos; the video frame images have multi-channel high-order features; S3. By using the perceptual hash algorithm, interpolation hash algorithm, compression coding principle and national cryptographic SM4 algorithm, the multi-channel high-order features of video frame images are fused, reduced in dimension, compressed and encrypted to generate dynamic security tags. S4. Based on the principles of discrete cosine transform, relative value quantization mechanism, multi-voting mechanism, and inverse discrete cosine transform, dynamic security tags are embedded into video frame images and encoded to generate approved videos. S5. Extract the dynamic security marker of the currently playing video and verify the consistency between the feature information extracted from the dynamic security marker and the feature information of the currently playing video to determine whether the video is compliant.

2. The video content protection method based on multi-channel image feature security tagging according to claim 1, characterized in that, S1 reviews the original videos and obtains approved videos, specifically including: Based on an AI algorithm model using deep learning technology, an intelligent pre-screening is performed on the original video to determine whether it contains sensitive, inappropriate, illegal, or infringing content, yielding the pre-screening results: If the original video does not contain sensitive, inappropriate, illegal, or infringing content in the preliminary review results, then the preliminary review is compliant; otherwise, the preliminary review is non-compliant. Regardless of whether the preliminary review results are compliant or not, a second manual review of the preliminary review results is required. If the review is successful, the video is deemed compliant. If the review fails, the video is deemed non-compliant.

3. The video content protection method based on multi-channel image feature security tagging according to claim 1, characterized in that, S3 uses perceptual hashing, interpolation hashing, compression coding principles, and the national cryptographic SM4 algorithm to fuse, reduce, compress, and encrypt the multi-channel high-order features of video frame images, generating dynamic security tags. Specifically, this includes: The video frame image is binarized and filtered to obtain the processed video frame image, and then the processed video frame image is divided into several block images. Based on the perceptual hash algorithm, perceptual hash features are extracted from several block images and video frame images respectively, and the perceptual hash values ​​of several block images and video frame images are calculated. Based on the interpolation hash algorithm, the interpolation hash features of the video frame image are extracted, and the interpolation hash value of the video frame image is calculated. Based on the principle of compression coding, the perceptual hash value and the interpolated hash value are combined to obtain the first combined value of the perceptual hash and the interpolated hash, and the first combined value is encoded into a QR code; The QR code is encrypted using the national cryptographic algorithm SM4 to obtain a dynamic security tag.

4. The video content protection method based on multi-channel image feature security tagging according to claim 1, characterized in that, S4 embeds dynamic security tags into video frame images based on the principles of discrete cosine transform, relative value quantization, multi-voting mechanism, and inverse discrete cosine transform, and encodes them to generate approved video. Specifically, this includes: Obtain the spatial domain luminance channel of the original video frame image; Based on the principle of discrete cosine transform, the frequency domain transformation is performed on the spatial domain brightness channel, i.e., the Y channel, to obtain the Y channel DCT coefficient matrix; According to the coefficient pair selection rule, coefficient pairs are selected from the Y-channel DCT coefficient matrix, and the coefficient pairs include: coefficient 1 and coefficient 2; Based on the relative value quantization mechanism, the coefficient values ​​of the selected coefficient pair are modified so that when embedding bit 1, the relative difference between coefficient 1 and coefficient 2 of the coefficient pair is greater than or equal to the embedding strength. When embedding bit 0, the relative difference between coefficient 2 and coefficient 1 of the coefficient pair is greater than or equal to the embedding strength. Furthermore, the above operations are used to embed dynamic security tags; A multi-voting mechanism is adopted to achieve redundant embedding of dynamic security tags, namely, selecting three different coefficient pairs to finally obtain the DCT coefficient matrix of embedded dynamic security tags; Based on the principle of inverse discrete cosine transform, the DCT coefficient matrix with embedded dynamic security tags is transformed in the spatial domain to obtain the video frame image with embedded dynamic security tags. Video frames with embedded dynamic security tags are encoded to generate approved videos.

5. The video content protection method based on multi-channel image feature security tagging according to claim 4, characterized in that, The frequency domain expression of the original video frame image luminance channel, i.e., the Y channel, is: ； ； in, For frequency index, The frequency domain coefficient matrix, Image block size, The normalization coefficient is... , These are the pixel position coordinates in the spatial domain. In the spatial domain ( , The pixel value at the location. Using the basis function, the signal is decomposed into a superposition of cosine waves of different frequencies; The spatial domain expression of the Y channel of the video frame image with embedded dynamic security tags is: ； in, For embedding dynamic security tags in video frame image luminance channels ( , The pixel value at the location. The frequency domain coefficient matrix of the luminance channel of the video frame image to embed dynamic security tags. Image block size, The normalization coefficient is... These are basis functions.

6. The video content protection method based on multi-channel image feature security tagging according to claim 4, characterized in that, The coefficient pair selection rules specifically include: selecting mid-frequency coefficient pairs, coefficient pairs with an inter-frequency difference ratio range of 1 to 1.2, and mid-frequency coefficient pairs in the horizontal and vertical directions from the DCT coefficient matrix.

7. The video content protection method based on multi-channel image feature security tagging according to claim 3, characterized in that, S5 extracts the dynamic security tag of the currently playing video and verifies the consistency between the feature information extracted from the dynamic security tag and the feature information of the currently playing video to determine whether the video is compliant. Specifically, this includes: The program acquires the currently playing video frame image and divides it into blocks to obtain several video block images. For each video block image, according to the principle of discrete cosine transform, the brightness channel (Y channel) of the video block image is transformed from the spatial domain to the frequency domain to obtain the Y channel DCT coefficient matrix. Based on the Y-channel DCT coefficient matrix, dynamic security tags are extracted according to the relative difference values ​​of the embedded coefficient pairs. Finally, a referee mechanism is used to vote on multiple redundant security tags to obtain the final security tag information. The extracted dynamic security tag is decrypted to obtain the second combined value; The second combined value includes: a globally aware hash value, an interpolated hash value, and a locally aware hash value; Based on the perceptual hash algorithm, perceptual hash features are extracted from the currently playing video segment images to obtain the perceptual hash value of the currently playing video segment images; Based on the perceptual hash algorithm and the interpolation hash algorithm, the perceptual hash feature extraction and interpolation hash feature extraction are performed on the currently playing video frame image respectively to obtain the perceptual hash value and interpolation hash value of the currently playing video frame image; Based on the local perceptual hash value of the currently playing video frame image extracted from the frame image and the local perceptual hash value extracted from the dynamic security tag, the Hamming distance of the local perceptual hash value is calculated as the first similarity according to the Hamming distance calculation formula. The first similarity includes the Hamming distance of several perceptual hash values; Based on the global perceptual hash value of the currently playing video frame image extracted from the frame image and the global perceptual hash value extracted from the dynamic security tag, the Hamming distance of the global perceptual hash value is calculated as the second similarity according to the Hamming distance calculation formula. Based on the interpolated hash value of the currently playing video frame extracted from the frame image and the interpolated hash value extracted from the dynamic security tag, the Hamming distance of the global interpolated hash value is calculated as the third similarity according to the Hamming distance calculation formula. Based on the consistency judgment mechanism, the played video is verified according to the first similarity, second similarity and third similarity to determine whether the video is compliant.

8. The video content protection method based on multi-channel image feature security tagging according to claim 7, characterized in that, Based on the consistency determination mechanism, the played video is verified according to the first similarity, second similarity, and third similarity to determine whether the video content is compliant. Specifically, this includes: Based on the first similarity dataset, the first similarity standard deviation is calculated, and the first similarity dataset is sorted in ascending order to obtain the first similarity sequence; The first similarity standard deviation threshold, the second similarity threshold, and the third similarity threshold are obtained. The first similarity standard deviation, the second similarity threshold, and the third similarity threshold are combined to perform multi-level judgment to determine whether the video content is compliant.

9. The video content protection method based on multi-channel image feature security tagging according to claim 8, characterized in that, Obtain the first similarity standard deviation threshold, the second similarity threshold, and the third similarity threshold. Combine the first similarity standard deviation, the second similarity threshold, and the third similarity threshold to perform multi-level judgments to determine whether the video content is compliant. Specifically, this includes: Global anomaly check: If the judgment conditions for global anomaly are met, the video playback will show a global anomaly, and the violation frame will be reported to the management platform. The judgment conditions for global anomaly include: the first similarity standard deviation is higher than the corresponding first similarity standard deviation threshold, the second similarity is lower than the corresponding second similarity threshold, and the third similarity is lower than the corresponding third similarity threshold. If the conditions for judging a global exception are not met, then the video playback is judged to be non-global exception, and a local exception judgment is performed. Determining if there is a local anomaly: If the judgment conditions for local anomalies are met, the playback video is judged to be locally abnormal, and the violation frame is reported to the management platform. The judgment conditions for local anomalies include: the first similarity of the playback video block image ranked first in the first similarity sequence is lower than the corresponding first similarity single-point threshold; the first similarity of the video block images ranked first and second in the first similarity sequence is lower than the corresponding first similarity two-point threshold; and the first similarity of the video block images ranked first and fourth in the first similarity sequence is lower than the corresponding first similarity four-point threshold. If the criteria for judging local anomalies are not met, then the video playback is not considered to have local anomalies, and the video content is determined to be compliant.

10. A video content protection device based on multi-channel image feature security tagging, characterized in that, include: The first acquisition module is used to acquire the original video, review the original video, and obtain the video that has passed the review. The second acquisition module is used to acquire video frame images of videos that have passed the review; the video frame images have multi-channel high-order features. The tag generation module is used to fuse, reduce, compress, and encrypt the multi-channel high-order features of video frame images through perceptual hashing algorithm, interpolation hashing algorithm, compression coding principle and national cryptographic SM4 algorithm to generate dynamic security tags. The tag embedding module is used to embed dynamic security tags into video frame images based on the principles of discrete cosine transform, relative value quantization mechanism and inverse discrete cosine transform, and encode them to generate approved video. The verification module extracts the dynamic security markers from the playing video and verifies the consistency between the feature information extracted from the dynamic security markers and the feature information of the playing video to determine whether the video is compliant.

Images