Intelligent identification method and system for illegal content of digital multimedia

Abstract

The application provides a kind of illegal content intelligent identification method and system for digital multimedia, related to digital information processing technical field, obtains to be identified digital multimedia content and its propagation trace information, extracts each transmission node feature mark generation set from content original data and modification record, traces the evolution relationship of adjacent node feature mark set to form trajectory, combines transmission object information to verify whether difference feature mark is illegal, obtains identification result.Based on the identification result, a disposal scheme containing content blocking link identification and content modification trace instruction is generated and sent to each propagation link management terminal.The present application comprehensively considers digital multimedia content and its propagation process, can accurately identify illegal content and effectively control the spread, and maintain the safety of digital information environment.

Description

Technical Field

[0001] This application relates to the field of digital information processing technology, and more specifically, to a method and system for intelligent identification of illegal content in digital multimedia. Background Technology

[0002] In today's rapidly developing digital information age, digital multimedia content (such as images, videos, and audio) is widely disseminated in the online environment. With the increasing frequency of dissemination, the problem of the spread of illegal content within digital multimedia content is becoming increasingly prominent. This illegal content may involve violence, pornography, terrorism, disinformation, etc., seriously endangering social security, public order and morals, and individual rights.

[0003] Currently, the identification of illegal digital multimedia content faces the following main problems. Firstly, traditional identification methods often focus solely on the digital multimedia content itself, directly judging the content through image recognition, text analysis, and other technologies. However, these methods ignore the dynamic changes in content during dissemination, making it difficult to detect illegal features hidden after multiple transmissions and modifications. Secondly, while some existing technologies consider the dissemination process, they merely record the dissemination path, lacking in-depth analysis and exploration of changes in content characteristics during dissemination. This makes it impossible to accurately determine the source and evolution of illegal content, leading to inaccurate identification results and hindering the effective prevention of further dissemination of illegal content. Consequently, these methods fail to meet the current demand for accurate identification and effective control of illegal digital multimedia content. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for intelligent identification of illegal content in digital multimedia.

[0005] According to a first aspect of this application, a method for intelligent identification of illegal content in digital multimedia is provided, the method comprising:

[0006] The digital multimedia content to be identified and its dissemination traceability information are obtained. The digital multimedia content includes the original content data and modification records during the dissemination process. The dissemination traceability information includes the transmission nodes, transmission time and transmission objects of the digital multimedia content at each dissemination stage.

[0007] Based on the transmission nodes and transmission times in the propagation tracing information, the content feature imprints corresponding to each transmission node are extracted from the original content data and modification records of the digital multimedia content to generate a feature imprint set for each transmission node.

[0008] According to the transmission time sequence, the evolutionary relationship between the feature imprint sets of adjacent transmission nodes is tracked, the differences in feature imprints during the evolution process are marked, and the feature imprint evolution trajectory is formed;

[0009] By combining the transmission object information in the propagation tracing information, the differential feature imprints in the feature imprint evolution trajectory are correlated and verified to determine whether the differential feature imprints belong to illegal feature imprints, and the illegal content identification result of the digital multimedia content is obtained.

[0010] Based on the illegal content identification results, a digital multimedia illegal content handling scheme is generated, which includes content blocking link identifiers and content modification tracing instructions, and the digital multimedia illegal content handling scheme is sent to the management terminals of each dissemination link.

[0011] According to a second aspect of this application, an intelligent identification system for illegal content in digital multimedia is provided. The intelligent identification system for illegal content in digital multimedia includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the intelligent identification system for illegal content in digital multimedia implements the aforementioned intelligent identification method for illegal content in digital multimedia.

[0012] According to a third aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, and when the computer-executable instructions are executed, the aforementioned method for intelligent identification of illegal content in digital multimedia is implemented.

[0013] Based on any of the above aspects, the technical effect of this application is as follows:

[0014] By acquiring the digital multimedia content to be identified and its dissemination source information, including the original content data, modification records during the dissemination process, and information on the transmission nodes, transmission times, and recipients at each stage of dissemination, the system extracts content feature imprints corresponding to each transmission node from the original content data and modification records based on the dissemination source information, generating a feature imprint set. This allows for a detailed depiction of the key characteristics of digital multimedia content at different stages of dissemination. By tracing the evolutionary relationship of feature imprint sets between adjacent transmission nodes in chronological order and marking differences, a feature imprint evolution trajectory is formed, clearly presenting the dynamic changes of the content during dissemination. By combining the recipient information with the correlation verification of the difference feature imprints in the feature imprint evolution trajectory, it is possible to accurately determine whether the difference feature imprints belong to illegal feature imprints, thus obtaining reliable illegal content identification results. Based on the illegal content identification results, a digital multimedia illegal content handling scheme is generated, including content blocking stage identifiers and content modification tracing instructions, and sent to the management terminals at each dissemination stage. This achieves effective blocking and source tracing of illegal content dissemination, significantly improving the accuracy of identification and control efficiency of digital multimedia illegal content. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the intelligent identification method for illegal content in digital multimedia provided in an embodiment of this application is shown.

[0016] Figure 2 This paper illustrates a schematic diagram of the component structure of an intelligent identification system for illegal content in digital multimedia provided in an embodiment of this application. Detailed Implementation

[0017] Figure 1 This paper illustrates a flowchart of an intelligent method and system for identifying illegal content in digital multimedia, as provided in an embodiment of this application. The detailed steps include:

[0018] Step S110: Obtain the digital multimedia content to be identified and the dissemination traceability information of the digital multimedia content. The digital multimedia content includes the original content data and the modification records during the content dissemination process. The dissemination traceability information includes the transmission nodes, transmission time and transmission objects of the digital multimedia content in each dissemination stage.

[0019] In this embodiment, offline promotional activities of chain supermarkets are used as the unified application scenario. The digital multimedia content to be identified includes three types of files: promotional videos issued by headquarters, activity poster images, and activity rule texts. The original data of the promotional videos is a video file in a standard encoding format, containing dynamic product display footage, background audio streams, and editable subtitle tracks. Modification records are stored in a structured document format, detailing operations such as regional branches adding local dialect voiceovers and stores adjusting the display position of promotional information. Dissemination traceability information is stored using distributed ledger technology to ensure its immutability. Transmission nodes include the headquarters content management server, regional branch processing terminals, store display terminals, and partner social media platform servers. Transmission time uses a standardized timestamp format to accurately record the transmission time of content between nodes. Transmission object information includes the organizational identification code of each node, the identity of the person in charge of operations, and the compliance rating of historical dissemination content. During the data collection phase, the identity of the responsible person in the information of the transmitted object is desensitized using a one-way hash algorithm. The original data is encrypted end-to-end through a secure transmission protocol during transmission to prevent the data from being illegally obtained or tampered with during transmission and to ensure the security of privacy-sensitive data.

[0020] Step S120: Based on the transmission nodes and transmission time in the propagation tracing information, extract the content feature imprints corresponding to each transmission node from the original content data and modification records of the digital multimedia content, and generate a feature imprint set for each transmission node.

[0021] Step S121: Parse the propagation tracing information, extract the node identifiers of all transmission nodes and the corresponding transmission time of each transmission node, and arrange them in chronological order of transmission time to form a transmission node sequence.

[0022] The structured data of the dissemination and traceability information is analyzed using specialized data parsing tools to extract unique node identifiers for each transmission node. For example, the identifier for the headquarters content management server is "HQ-CMS-2023", the identifier for the regional branch processing terminal is "REG-SH-2023", the identifier for the store display terminal is "STORE-NJ-2023", and the identifier for the partner social media platform server is "SOCIAL-PLAT-2023". Simultaneously, the transmission time corresponding to each transmission node is extracted, and this transmission time exists in the form of a standardized timestamp. The extracted node identifiers are associated with their corresponding transmission times, and then all transmission nodes are sorted according to the chronological order of transmission time to form a transmission node sequence. For example, the sorted sequence would be [HQ-CMS-2023, REG-SH-2023, STORE-NJ-2023, SOCIAL-PLAT-2023].

[0023] Step S122: Separate the original content data and modification records of the digital multimedia content, extract the modification execution time, modification operation content and modified content feature description from the modification records, and sort them according to the modification execution time to form a modification record sequence.

[0024] Professional multimedia processing tools were used to separate the digital multimedia content, separating the original content data from the modification records. For the modification records, document parsing technology was used to extract key information, including the execution time of each modification, the specific content of the modification, and a description of the characteristics of the modified content. The execution time also used a standardized timestamp format. The modification content described in detail the specific changes made to the content, such as "overlaying local dialect narration into the third audio segment of the video." The description of the characteristics of the modified content recorded the new features of the content after the modification, such as "the audio track contains both Mandarin and local dialect audio; the left channel is the original Mandarin narration, and the right channel is the newly added local dialect narration." These extracted modification record information were arranged in chronological order of their execution time to form a modification record sequence.

[0025] Step S123: Compare the transmission time of each transmission node in the transmission node sequence with the modification execution time in the modification record sequence to determine the content status corresponding to each transmission node. If the transmission time of a transmission node is earlier than all modification execution times, then the transmission node corresponds to the original data status of the content. If the transmission time of a transmission node is between two modification execution times, then the transmission node corresponds to the modified content status of the previous modification record. If the transmission time of a transmission node is later than all modification execution times, then the transmission node corresponds to the modified content status of the last modification record.

[0026] Iterate through each transmission node in the transmission node sequence, comparing its transmission time with the execution times of all modifications in the modification record sequence. If a transmission node's transmission time is earlier than any of the modification execution times in the modification record sequence, the content state corresponding to that transmission node is determined to be the original data state. If a transmission node's transmission time falls between two adjacent modification execution times, the content state corresponding to that transmission node is the modified content state formed after the execution of the previous modification record. If a transmission node's transmission time is later than any of the modification execution times in the modification record sequence, the content state corresponding to that transmission node is the modified content state after the execution of the last modification record. Through this comparison method, a unique content state is determined for each transmission node.

[0027] Step S124: For the content status corresponding to each transmission node, extract content feature imprints according to the presentation format of digital multimedia content.

[0028] Based on the different presentation formats of digital multimedia content, corresponding feature extraction methods are adopted. The presentation formats include video, image, and text. The feature extraction process for each type of presentation format is described in detail below.

[0029] Step S1241: If the presentation format is video, extract the stable image elements that appear in the keyframes of the video under the corresponding content state as keyframe feature marks, extract the frequency combinations that exist continuously in the audio as audio feature marks, and extract the repeated expression segments in the subtitle text as subtitle feature marks.

[0030] Step S1241-1: For the video in the corresponding content state, extract key frames from the video at preset time intervals to form a key frame sequence.

[0031] For video files in the corresponding content state, use video processing tools to extract keyframes at preset time intervals. The preset time interval is set based on the rhythm and speed of change of the video content to ensure accurate capture of important visual information. The extracted keyframes are arranged in chronological order of their appearance in the video, forming a keyframe sequence. For example, for a video showcasing a product's appearance from multiple angles, a larger time interval can be set because the visual changes are relatively smooth; while for video clips containing rapidly changing shots, a smaller time interval is set to avoid missing key moments.

[0032] Step S1241-2: Traverse each keyframe in the keyframe sequence and use image recognition technology to extract the image elements in each keyframe. The image elements include object outlines, color blocks, and text labels.

[0033] A deep learning-based image recognition model is used to process each keyframe in the keyframe sequence. This model, trained on a large amount of labeled image data, can accurately identify and extract image elements such as object outlines, color blocks, and text labels. For object outlines, the model can output their precise coordinate range and approximate shape description in the image; for color blocks, it can identify the main color types and their area proportion and distribution in the image; for text labels, it can identify the text content, font style, font size, and text position information in the image. This extracted image element information is then associated and stored with the corresponding keyframes.

[0034] Step S1241-3: Count the number of keyframes in which each image element appears in the keyframe sequence, calculate the proportion of the number of occurrences to the total number of frames in the keyframe sequence, filter out image elements whose proportion exceeds a preset stable threshold, arrange the filtered image elements whose proportion exceeds the preset stable threshold according to image element type, and form keyframe feature imprints.

[0035] For each extracted image element, the number of keyframes in which it appears in the entire keyframe sequence is counted. Then, the proportion of this occurrence to the total number of frames in the keyframe sequence is calculated, and this proportion is compared with a preset stability threshold. The preset stability threshold is set based on the characteristics of the video content and is used to determine whether image elements appear stably. Image elements whose occurrence proportion exceeds the preset stability threshold are filtered out; for example, the outline of a brand logo in the video appears in most keyframes, and its occurrence proportion exceeds the stability threshold. The filtered image elements are arranged in the order of object outline, color block, and text label type, forming the keyframe feature imprint of this type of digital multimedia content in the corresponding content state.

[0036] Step S1241-4: For the video audio portion of the corresponding content state, segment the audio according to preset time segments to form an audio segment sequence.

[0037] The audio portion of the video file is extracted to obtain an independent audio stream. Audio processing tools are then used to segment the audio stream into segments of preset duration, forming a sequence of audio segments. The preset duration should take into account the characteristics of the audio content. For example, for audio containing continuous narration, shorter duration segments can be set to capture more subtle frequency changes; for audio with a large proportion of background music, longer duration segments can be set.

[0038] Step S1241-5: Perform frequency analysis on each audio segment and extract the frequency combination in each audio segment. The frequency combination includes the main frequency components and the intensity ratio of each component.

[0039] Audio frequency analysis algorithms are used to process each audio segment in an audio clip sequence, analyzing the frequency spectrum distribution of each segment. The main frequency components, which are the primary sound elements constituting the audio segment, are extracted from the frequency spectrum. Simultaneously, the intensity proportion of each main frequency component in the entire frequency spectrum is calculated, expressed as the percentage of each frequency component's energy value to the total energy value. The main frequency components and their corresponding intensity proportions are combined to form the frequency combination for each audio segment.

[0040] Step S1241-6: Count the number of audio segments in the audio segment sequence for each frequency combination, calculate the proportion of the number of occurrences to the total number of segments in the audio segment sequence, filter out frequency combinations whose proportion exceeds a preset duration threshold, sort the filtered frequency combinations whose proportion exceeds the preset duration threshold by frequency range, and form audio feature imprints.

[0041] The frequency combinations of all audio segments are statistically analyzed, recording the frequency of each unique frequency combination in the audio segment sequence. The proportion of each frequency combination's occurrences relative to the total number of audio segments in the sequence is calculated, and this proportion is compared to a preset persistence threshold. The persistence threshold is used to determine whether a frequency combination persists in the audio. Frequency combinations whose occurrence proportion exceeds the persistence threshold are selected; for example, a specific frequency combination in background music may appear in multiple audio segments, exceeding the persistence threshold. The selected frequency combinations are then sorted in ascending order of frequency to form audio feature imprints.

[0042] Step S1241-7: For the video subtitle text under the corresponding content state, segment the subtitle text by sentence to form a subtitle sentence sequence, and extract the expression fragments in each subtitle sentence. The expression fragments include phrases, short sentences and specific expression formats.

[0043] The complete subtitle text is extracted from the video's subtitle track. Using sentence segmentation algorithms from natural language processing, the subtitle text is divided into independent sentences according to punctuation marks and other sentence markers, forming a subtitle sentence sequence. Then, each subtitle sentence is analyzed to extract its descriptive fragments. These fragments can be phrases with specific meanings, such as "limited-time offer," complete short sentences, such as "all items 20% off," or expressions with a fixed format, such as "Event period: XXXX year XX month XX day - XXXX year XX month XX day."

[0044] Step S1241-8: Count the number of sentences in the subtitle sentence sequence for each expression segment, calculate the proportion of the number of occurrences to the total number of sentences in the subtitle sentence sequence, filter out expression segments whose proportion exceeds the preset repetition threshold, arrange the filtered expression segments whose proportion exceeds the preset repetition threshold in sentence order to form subtitle feature marks.

[0045] The number of sentences in the subtitle sequence for each extracted expression fragment is counted, i.e., how many different sentences the expression fragment appears in. The proportion of this occurrence to the total number of sentences in the subtitle sequence is calculated and compared to a preset repetition threshold. The repetition threshold is used to determine whether an expression fragment appears repeatedly in the subtitle text. Expression fragments with an occurrence rate exceeding the repetition threshold are selected; for example, the expression fragment "20% off" appears in multiple subtitle sentences, exceeding the repetition threshold. The selected expression fragments are then arranged according to the order in which they first appear in the subtitle sequence, forming subtitle feature marks.

[0046] Step S1242: If the presentation format is an image, extract the color channel combination with a stable proportion in the pixel distribution of the image under the corresponding content state as the pixel feature mark, extract the recurring pattern structure in the texture as the texture feature mark, and extract the content identifier that is fixed in the region as the region feature mark.

[0047] Step S1242-1: Extract the color channel combination with a stable proportion in the pixel distribution of the image under the corresponding content state as the pixel feature imprint.

[0048] For image-based digital multimedia content, the first step is to analyze the image at the pixel level. The image is decomposed into multiple color channels, such as the common RGB color channels. Professional image analysis software is used to statistically analyze the pixel value distribution of each color channel in the image, calculating the proportion of each color channel's pixel value within the total number of pixels in the image. The distribution of these proportions across different regions of the image is analyzed, and color channel combinations with stable proportions are selected. These combinations, where the proportions of each color channel remain relatively stable across multiple regions of the image with minimal fluctuation, are recorded as pixel feature markers.

[0049] Step S1242-2: Extract the recurring pattern structure in the texture as texture feature imprint.

[0050] Texture analysis algorithms are used to process images and identify regions with texture features. These texture regions are then analyzed in depth to extract recurring pattern structures. By calculating parameters such as the frequency of recurrence, repetition period, and relative positional relationships between patterns within the texture region, representative recurring pattern structures are identified. These patterns can be regular grids, stripes, or specific shaped designs. The extracted recurring pattern structures are recorded as texture feature imprints.

[0051] Step S1242-3: Extract the fixed content identifiers in the region as regional feature imprints.

[0052] Image region segmentation technology is used to divide the image into multiple distinct content regions, such as product display areas, text description areas, and brand logo areas. Each region is analyzed individually to identify fixed content identifiers, such as brand logos, specific icons, and certification marks. These content identifiers have fixed positions and morphological characteristics within the image and do not change with changes in other parts of the image. These fixed content identifiers and their positional information within the image are recorded together to form region feature imprints.

[0053] Step S1243: If the presentation format is text, extract the frequently occurring keywords in the text under the corresponding content state as keyword feature marks, extract the fixed component collocation patterns in the sentence as sentence structure feature marks, and extract the core viewpoint expressions in the paragraph as semantic theme feature marks.

[0054] Step S1243-1: Extract frequently occurring keywords in the text under the corresponding content state as keyword feature marks.

[0055] Step S1243-1-1: For the text in the corresponding content state, perform word segmentation to obtain a valid word sequence.

[0056] Use the word segmentation tool in natural language processing to segment text content, splitting continuous text strings into independent word units. During the word segmentation process, punctuation marks, stop words (such as meaningless words like "de", "shi", "zai", etc.), and some special symbols in the text are removed simultaneously, obtaining a valid vocabulary sequence that only contains meaningful words.

[0057] Step S1243-1-2: Count the frequency of each valid word in the valid vocabulary sequence, calculate the proportion of the frequency of each valid word to the total frequency of all valid words, screen out the valid words whose proportion exceeds the preset high-frequency threshold, and sort the screened valid words whose proportion exceeds the preset high-frequency threshold according to the word frequency, forming a keyword feature imprint.

[0058] Count the frequency of occurrence of each word in the valid vocabulary sequence and record the number of times each word appears in the sequence. Then calculate the proportion of the frequency of each word to the total frequency of all valid words. This proportion reflects the importance of the word in the text. Compare the calculated proportion with the preset high-frequency threshold, and screen out the valid words whose proportion exceeds the high-frequency threshold. These words are the high-frequency keywords in the text. Sort the above high-frequency keywords according to their frequencies of occurrence, forming a keyword feature imprint.

[0059] Step S1243-2: Extract the fixed component collocation patterns in the sentences as the sentence structure feature imprint.

[0060] Step S1243-2-1: Conduct grammatical analysis on the sentences in the text to determine the sentence components of each sentence. The sentence components include subject, predicate, object, attributive, adverbial, and complement.

[0061] Use a grammatical analysis tool to conduct a detailed grammatical structure analysis on each sentence in the text, identify and determine the various sentence components in the sentence, such as the subject (indicating the main body described by the sentence), the predicate (indicating the action or state of the main body), the object (indicating the object of the action), the attributive (modifying the nominal component), the adverbial (modifying the verb or adjective), the complement (supplementing and explaining the predicate), etc. Clearly define the specific position and role of each sentence component in the sentence.

[0062] Step S1243-2-2: Extract the collocation patterns of the sentence components in each sentence. The collocation patterns include the combination order of the components and the combination of the词性 of each component.

[0063] Based on the identification of sentence components, the collocation patterns of these components in each sentence are extracted. These patterns include the order of combination between sentence components, such as the "subject + verb + object" order, and the parts of speech corresponding to each component, such as a noun phrase for the subject, a verb for the verb, and a noun phrase for the object. Through the analysis of a large number of sentences, representative collocation patterns are summarized.

[0064] Step S1243-2-3: Count the number of sentences in which each collocation pattern appears in all sentences of the text, calculate the proportion of the number of occurrences to the total number of sentences in the text, filter out the collocation patterns whose proportion exceeds a preset fixed threshold, arrange the filtered collocation patterns whose proportion exceeds the preset fixed threshold according to the grammatical structure type, and form sentence structure feature imprints.

[0065] The number of sentences in the entire text that each extracted collocation pattern appears is counted, i.e., how many sentences use that collocation pattern. The proportion of this occurrence to the total number of sentences in the text is calculated and compared with a preset fixed threshold. The fixed threshold is used to determine whether a collocation pattern consistently appears in the text. Collocation patterns with an occurrence rate exceeding the fixed threshold are selected and then classified and arranged according to the grammatical structure type they represent, forming sentence structure feature imprints.

[0066] Step S1243-3: Extract the core viewpoints in the paragraph as semantic thematic feature imprints.

[0067] Step S1243-3-1: Divide the text into paragraphs to form a paragraph sequence. For each paragraph, extract the core viewpoint statement that expresses the core viewpoint. The core viewpoint statement includes the first sentence of the paragraph, the last sentence of the paragraph, and a statement containing summary words.

[0068] The text is divided into independent paragraphs according to paragraph markers, forming a paragraph sequence. For each paragraph, natural language understanding technology is used to analyze it and extract the statements that express the core viewpoint. These core viewpoint statements usually include the first sentence of the paragraph (which often states the main idea of ​​the paragraph directly), the last sentence of the paragraph (which may summarize the content of the paragraph), and statements containing summarizing words (such as "in conclusion", "therefore", "in short" etc.).

[0069] Step S1243-3-2: Semantically summarize the extracted core viewpoint statements to obtain the core viewpoint expression of each paragraph. Count the number of paragraphs in which each core viewpoint expression appears in the core viewpoint expressions of all paragraphs. Calculate the proportion of the number of occurrences to the total number of paragraphs in the paragraph sequence. Select core viewpoint expressions with a proportion exceeding the preset core threshold. Sort the selected core viewpoint expressions with a proportion exceeding the preset core threshold according to the degree of viewpoint relevance to form semantic theme feature imprints.

[0070] Each extracted core viewpoint statement undergoes semantic summarization to remove redundant information and retain the most essential viewpoint content, resulting in the core viewpoint expression for each paragraph. The number of paragraphs in which each core viewpoint expression appears is counted, i.e., how many different paragraphs mention or emphasize the core viewpoint. The proportion of this occurrence to the total number of paragraphs is calculated and compared to a preset core threshold. The core threshold is used to determine whether the core viewpoint expression occupies an important position in the text. Core viewpoint expressions with an occurrence rate exceeding the core threshold are selected and then sorted according to the inherent logical correlation between these core viewpoint expressions to form semantic thematic feature imprints.

[0071] Step S125: Sort all types of content feature imprints corresponding to each transmission node according to feature imprint type, remove duplicate feature imprints, and form the feature imprint set of that transmission node.

[0072] For each delivery node, all types of content feature imprints extracted from its corresponding content state are categorized and organized according to their type. This includes keyframe feature imprints, audio feature imprints, and subtitle feature imprints for videos; pixel feature imprints, texture feature imprints, and region feature imprints for images; and keyword feature imprints, sentence structure feature imprints, and semantic topic feature imprints for texts. Within the same type of feature imprint, duplicate imprints are removed to ensure that each feature imprint appears only once in the set. The deduplicated feature imprints are then combined to form the complete feature imprint set for that delivery node.

[0073] Step S126: Add a transmission node identifier and a content status identifier to each feature imprint set, and integrate all feature imprint sets of transmission nodes in the order of transmission time to generate feature imprint sets for each transmission node.

[0074] For each transmission node's feature imprint set, add corresponding identification information, including the node's unique identifier and corresponding content status identifier. The content status identifier indicates whether the feature imprint set was generated based on the original data state or the state after a certain modified record. After adding the identifier, integrate all transmission node feature imprint sets according to the chronological order of each node's transmission time to form a complete data structure containing all node feature imprint information, which is the final generated feature imprint set for each transmission node.

[0075] Example Implementation Section:

[0076] Step S130: In chronological order of transmission time, trace the evolutionary relationship between the feature imprint sets of adjacent transmission nodes, mark the differences in feature imprints during the evolution process, and form the feature imprint evolution trajectory.

[0077] Step S131: From the generated feature imprint sets of each transmission node, select two adjacent feature imprint sets in chronological order of transmission time, and label them as the preceding feature imprint set and the following feature imprint set, respectively.

[0078] From the integrated set of feature imprints for each transmission node, feature imprint sets of adjacent nodes are selected sequentially according to their transmission time. The feature imprint set of the earlier transmission node is designated as the preceding feature imprint set, and the feature imprint set of the later transmission node is designated as the following feature imprint set. For example, if the transmission node sequence is [A, B, C, D], then the feature imprint sets of A and B are selected first, with A as the preceding set and B as the following set; then the feature imprint sets of B and C are selected, with B as the preceding set and C as the following set, and so on.

[0079] Step S132: For the previous feature imprint set and the subsequent feature imprint set, compare them one by one according to the feature imprint type.

[0080] For each pair of adjacent sets of preceding and following feature imprints, comparisons are performed sequentially according to the type of feature imprint. That is, keyframe feature imprints are compared first, followed by audio feature imprints, then subtitle feature imprints, then pixel feature imprints, texture feature imprints, and region feature imprints, and finally keyword feature imprints, sentence structure feature imprints, semantic theme feature imprints, etc.

[0081] Step S1321: If the feature imprint type is keyframe feature imprint, compare the keyframe feature imprints in the preceding and following sets, count the proportion of the same screen elements, mark the newly added screen elements and the disappeared screen elements, and record the positions of the newly added and disappeared screen elements in the keyframe.

[0082] The keyframe feature imprints in the preceding feature imprint set are compared one by one with those in the subsequent feature imprint set. First, identical image elements in both sets are identified, their numbers are counted, and the proportion of this number to the total number of keyframe feature imprints in the preceding set is calculated. This proportion reflects the inheritance of keyframe feature imprints. Then, newly added image elements in the subsequent set are marked—that is, image elements that were not present in the preceding set but exist in the subsequent set, and disappeared image elements that existed in the preceding set but are not present in the subsequent set. For newly added and disappeared image elements, their specific positional information within the keyframes, such as their coordinate range within the image, is recorded by comparing the preceding and following keyframe sequences.

[0083] Step S1322: If the feature imprint type is audio feature imprint, compare the audio feature imprints in the preceding and following sets, count the percentage of duration of the same frequency combination, mark the newly added frequency combination and the disappeared frequency combination, and record the time period of the occurrence of the newly added and disappeared frequency combination.

[0084] Compare the audio feature imprints in the preceding and following feature imprint sets to identify common frequency combinations. Calculate the proportion of these common frequency combinations in the total audio duration, i.e., the duration percentage. Mark the frequency combinations that appear in the following set and those that disappear in the preceding set. By analyzing the audio segment sequence, determine the time periods when new frequency combinations begin and end in the following audio, and the time periods when disappearing frequency combinations appear in the preceding audio, and record this time period information.

[0085] Step S1323: If the feature imprint type is subtitle feature imprint, compare the subtitle feature imprints in the preceding and following sets, count the proportion of identical expression segments, mark the newly added expression segments and the disappeared expression segments, and record the sentence positions of the newly added and disappeared expression segments in the subtitle text.

[0086] The subtitle feature marks in the preceding and succeeding sets are compared, the number of identical expression segments is counted, and the proportion of this number to the total number of subtitle feature marks in the preceding set is calculated. New expression segments added in the succeeding set and expression segments that disappeared in the preceding set are marked. By searching the subtitle sentence sequence, the sentence position (e.g., which sentence) of the new expression segment in the subtitle text and the sentence position of the disappeared expression segment in the preceding subtitle text are determined and recorded.

[0087] Step S1324: If the feature imprint type is pixel feature imprint, compare the pixel feature imprints in the preceding and following sets, calculate the difference in the proportion of the same color channel combination, mark the newly added color channel combination and the disappeared color channel combination, and record the area range of the newly added and disappeared color channel combination in the image.

[0088] For pixel feature imprints of image content, color channel combinations in the preceding and following sets are compared one by one. The difference in the proportion of the same color channel combination in the preceding and following sets is calculated. If the difference is less than a preset threshold, it is determined to be a stably inherited color channel combination. Newly added color channel combinations (i.e., combinations not present in the preceding set) and disappeared color channel combinations (i.e., combinations present in the preceding set but not in the following set) in the following set are marked. Through image region analysis technology, the specific region range of the newly added color channel combination in the image (such as the upper left 1 / 4 region, the central region, etc.) and the original region range of the disappeared color channel combination are determined, and the coordinate boundary information of these regions is recorded.

[0089] Step S1325: If the feature imprint type is texture feature imprint, compare the texture feature imprints in the preceding and following sets, count the percentage of repetitions of the same pattern structure, mark the newly added pattern structure and the disappeared pattern structure, and record the texture region of the newly added and disappeared pattern structure in the image.

[0090] The texture feature imprints in the preceding and following sets are compared to extract pattern structure information. The repetition rate of the same pattern structure in both sets is calculated, i.e., the proportion of times the pattern structure appears repeatedly in the texture region relative to the total texture area. New pattern structures (such as mesh-like textures or wavy textures not present in the preceding set) and disappearing pattern structures (such as striped textures present in the preceding set) in the following set are marked. The specific texture regions in the image where new pattern structures appear (such as the surface of product packaging or the background region) are located using a texture segmentation algorithm, as well as the original texture regions where disappearing pattern structures were located, and the position and size information of these regions are recorded.

[0091] Step S1326: If the feature imprint type is a region feature imprint, compare the region feature imprints in the preceding and following sets, count the proportion of the same content identifiers, mark the newly added content identifiers and the disappeared content identifiers, and record the specific locations of the newly added and disappeared content identifiers in the image region.

[0092] Compare the regional feature imprints in the preceding and following sets to identify content identifiers (such as brand logos, certification icons, price tags, etc.). Calculate the percentage of identical content identifiers, i.e., the proportion of content identifiers present in both sets to the total number of regional feature imprints in the preceding set. Mark newly added content identifiers (such as the newly added "Region-Specific" icon) and disappeared content identifiers (such as the existing "Nationwide Warranty" label) in the following set. Using image region localization technology, determine the specific coordinates of the newly added content identifier within the image region (e.g., the range from the lower right corner (X1, Y1) to (X2, Y2)), and the original coordinates of the disappeared content identifier, and record these coordinates.

[0093] Step S1327: If the feature imprint type is keyword feature imprint, compare the keyword feature imprints in the preceding and following sets, count the frequency ratio of the same keyword, mark the newly added keywords and the disappeared keywords, and record the paragraph positions of the newly added and disappeared keywords in the text.

[0094] The keyword feature imprints in the preceding and following sets are compared, and the frequency ratio of the same keyword in the preceding and following sets is calculated, i.e., the sum of the frequencies of the same keyword in the preceding set is proportional to the total frequency of keywords in the preceding set. New keywords (e.g., "limited-time offer") and disappearing keywords (e.g., "discount for purchases over a certain amount") in the following set are marked. Using text paragraph indexing, the paragraph position where the new keyword first appears in the text (e.g., paragraph 3, sentence 2) and the original paragraph position where the disappearing keyword was located (e.g., paragraph 2, sentence 5) are determined, and this positional information is recorded.

[0095] Step S1328: If the feature imprint type is sentence structure feature imprint, compare the sentence structure feature imprints in the preorder and postorder sets, count the percentage of sentences with the same component collocation pattern, mark the newly added component collocation pattern and the disappeared component collocation pattern, and record the sentence position of the newly added and disappeared component collocation pattern in the text.

[0096] By comparing the sentence structure feature imprints in the preceding and following sets, extract the component collocation patterns (such as "subject + predicate + object", "attributive + noun", etc.). Calculate the percentage of sentences with the same component collocation pattern in both sets, i.e., the proportion of sentences using the same collocation pattern out of the total number of sentences with sentence structure feature imprints in the preceding set. Mark newly added component collocation patterns (such as "adverbial + predicate + complement") and disappeared component collocation patterns (such as "subject + linking verb + predicate nominative") in the following set. Locate the sentences with newly added collocation patterns in the text (e.g., sentence 15, sentence 23) and the original sentence positions of disappeared collocation patterns (e.g., sentence 8, sentence 12) using sentence numbering, and record the sequence information of these sentences.

[0097] Step S1329: If the feature imprint type is semantic topic feature imprint, compare the semantic topic feature imprints in the preceding and following sets, count the proportion of the same core viewpoint expressions, mark the newly added core viewpoint expressions and the disappeared core viewpoint expressions, and record the paragraph positions of the newly added and disappeared core viewpoint expressions in the text.

[0098] For example, step S1329-1: Extract all semantic theme feature imprints from the set of preceding feature imprints to obtain a list of preceding core viewpoint statements; extract all semantic theme feature imprints from the set of subsequent feature imprints to obtain a list of subsequent core viewpoint statements.

[0099] All semantic topic feature imprints are extracted from both the preceding and following feature imprint sets. These semantic topic feature imprints are essentially core viewpoint statements. The extracted core viewpoint statements from the preceding set are organized into a list, i.e., the preceding core viewpoint statement list; similarly, the core viewpoint statements from the following set are organized into a following core viewpoint statement list.

[0100] Step S1329-2: Perform a semantic comparison between each core viewpoint statement in the preceding core viewpoint statement list and each core viewpoint statement in the subsequent core viewpoint statement list.

[0101] Using semantic similarity calculation methods, each core viewpoint statement in the preceding core viewpoint statement list is compared one by one with each core viewpoint statement in the subsequent core viewpoint statement list to determine whether they are semantically identical or highly similar. The semantic comparison process considers factors such as the core meaning and key concepts of the statements, rather than just literal matching.

[0102] Step S1329-3: If two core viewpoints are semantically identical or highly similar, they are determined to be identical core viewpoints and recorded as matching statements.

[0103] During the semantic comparison process, if two core viewpoints have the same semantic content, or if they have different forms of expression but highly consistent core meanings, then the two expressions are determined to be the same core viewpoints and are recorded as matching expressions.

[0104] Step S1329-4: Count the number of matching statements and calculate the proportion of the number of matching statements to the total number of preceding core viewpoint statements. This proportion is used as the percentage of the number of statements with the same core viewpoint.

[0105] The number of all statements that are judged to be matching is counted, and then the proportion of this number to the total number of statements in the preceding core viewpoint statement list is calculated. This proportion is the percentage of the number of statements with the same core viewpoint, which is used to measure the degree of inheritance of semantic topic feature imprints in the preceding and following sets.

[0106] Step S1329-5: Traverse the list of subsequent core viewpoint statements, filter out core viewpoint statements that are not in the matching statements, and add the filtered core viewpoint statements as new core viewpoint statements; traverse the list of preceding core viewpoint statements, filter out core viewpoint statements that are not in the matching statements, and add the filtered core viewpoint statements as disappeared core viewpoint statements.

[0107] The process iterates through the list of subsequent core viewpoint statements, filtering out those not marked as matching statements. These statements represent the core viewpoint statements added during the evolution process. Similarly, the process iterates through the list of preceding core viewpoint statements, filtering out those not marked as matching statements. These statements represent the core viewpoint statements that disappeared during the evolution process.

[0108] Step S1329-6: For each newly added core viewpoint statement, retrieve the text content associated with the subsequent feature imprint set corresponding to the core viewpoint statement, locate the paragraph position of the core viewpoint statement in the text, and record the paragraph number and the sentence position within the paragraph.

[0109] For each newly added core viewpoint statement, the original text content associated with its subsequent feature imprint set is retrieved. Text retrieval technology is used to locate the specific paragraph containing the new core viewpoint statement within the original text, the paragraph number is recorded, and the specific sentence position of the statement within the paragraph is determined, such as which sentence within the paragraph.

[0110] Step S1329-7: For each missing core viewpoint statement, retrieve the text content associated with the preceding feature imprint set corresponding to the core viewpoint statement, locate the paragraph position of the core viewpoint statement in the text, and record the paragraph number and the sentence position within the paragraph.

[0111] For each missing core viewpoint statement, the original text content associated with its pre-existing feature set is retrieved. Text retrieval technology is used to locate the specific paragraph containing the missing core viewpoint statement within the original text, and the paragraph number and the statement's position within that paragraph are recorded.

[0112] Step S1329-8: Integrate the percentage of identical core viewpoint expressions, the list of newly added core viewpoint expressions and their corresponding paragraph positions, and the list of disappeared core viewpoint expressions and their corresponding paragraph positions to form a semantic theme feature imprint comparison result.

[0113] The percentage of identical core viewpoints obtained from the above statistics, the list of newly added core viewpoints and their corresponding paragraph positions, and the list of disappeared core viewpoints and their corresponding paragraph positions are integrated to form a complete semantic theme feature imprint comparison result. This semantic theme feature imprint comparison result reflects the evolution of semantic theme feature imprints between adjacent transmission nodes.

[0114] Step S133: Classify and record the differences in the feature imprint comparison results of adjacent transmission nodes according to the feature imprint type, and form feature evolution units of adjacent nodes.

[0115] All differences obtained by comparing the feature imprint sets of adjacent transmission nodes are categorized and organized according to the type of feature imprint. For example, differences in keyframe feature imprints (added and disappeared image elements and their positions), audio feature imprints (frequency combinations and time periods of addition and disappearance), and subtitle feature imprints are recorded separately. The differences recorded in the above categories are combined to form an adjacent node feature evolution unit, which completely records the evolutionary differences in feature imprints between two adjacent transmission nodes.

[0116] Step S134: Traverse all adjacent feature imprint sets of the transmission nodes, generate multiple adjacent node feature evolution units, integrate all adjacent node feature evolution units in the order of transmission time, add the evolution time span of each unit, and form a feature imprint evolution trajectory.

[0117] Following the sequence of transmission nodes, the feature imprint sets of all adjacent transmission nodes are compared sequentially to generate corresponding adjacent node feature evolution units. Each evolution unit is supplemented with corresponding evolution time span information, which is the difference between the transmission time of the subsequent transmission node and the transmission time of the preceding transmission node. All generated adjacent node feature evolution units are integrated according to their transmission time order to form a complete feature imprint evolution trajectory. This trajectory demonstrates the evolution of the feature imprints of digital multimedia content throughout the entire dissemination process.

[0118] Step S140: Combining the transmission object information in the propagation tracing information, perform correlation verification on the differential feature imprints in the feature imprint evolution trajectory, determine whether the differential feature imprints belong to illegal feature imprints, and obtain the illegal content identification result of the digital multimedia content.

[0119] Step S141: Extract the transmission object information corresponding to each transmission node from the transmission tracing information. The transmission object information includes the identity identifier of the transmission object and historical transmission content records.

[0120] From the dissemination tracing information, for each transmission node, the corresponding transmission object information is extracted. The identity of the transmission object is a unique identifier of the organization or individual to which the transmission node belongs, and the historical dissemination content record contains relevant information about the digital multimedia content disseminated by the transmission object in the past period, such as the title, type, dissemination time, and modification record summary of the disseminated content.

[0121] Step S142: Extract all differential feature imprints from the feature imprint evolution trajectory, and classify and arrange them according to the feature imprint type and the corresponding adjacent node feature evolution units.

[0122] The feature evolution trajectories of all adjacent nodes are analyzed to extract all differential feature imprints, including newly added and disappeared feature imprints. These differential feature imprints are then categorized and arranged according to their feature imprint types (e.g., keyframes, audio, subtitles) and corresponding adjacent node feature evolution units to form a structured list of differential feature imprints for subsequent verification.

[0123] Step S143: For each differential feature imprint, retrieve the transmission node identifier in the adjacent node feature evolution unit corresponding to the differential feature imprint, and find the transmission object information corresponding to the transmission node identifier.

[0124] For each differential feature imprint, its adjacent node feature evolution unit is determined, and the corresponding transmission node identifier is obtained from this evolution unit. This transmission node identifier is the identifier of the transmission node to which the subsequent feature imprint set belongs. Based on this transmission node identifier, the corresponding transmission object information is searched and obtained from the propagation tracing information.

[0125] Step S144: Analyze the historical dissemination content records in the information of the target audience, extract the feature imprints contained in the historical dissemination content, compare them with the current difference feature imprints, and calculate the similarity between the two.

[0126] Step S1441: Extract all digital multimedia content that the target has previously disseminated from the historical dissemination content record of the target information, and arrange them in chronological order of dissemination time to form a historical dissemination content sequence.

[0127] Based on the historical dissemination content records in the recipient's information, extract all digital multimedia content that the recipient has disseminated in the past. Arrange the aforementioned historical dissemination content in chronological order of its dissemination to form a historical dissemination content sequence for systematic analysis.

[0128] Step S1442: For each historical content in the historical dissemination content sequence, extract the corresponding feature imprint according to its presentation format. The extraction method is consistent with the feature imprint extraction method of the current digital multimedia content.

[0129] For each historical content in the historical content sequence, the same feature extraction method as the current digital multimedia content to be identified is used to extract the corresponding feature imprints according to its presentation form (video, image, text), such as keyframe feature imprints, audio feature imprints, subtitle feature imprints, etc.

[0130] Step S1443: Classify the feature imprints of each historical dissemination content by type to form a set of historical feature imprints, and associate each set of historical feature imprints with the corresponding historical dissemination time.

[0131] The feature imprints extracted from each piece of historical dissemination content will be categorized and organized according to feature imprint type, forming a set of historical feature imprints for that piece of historical dissemination content. Simultaneously, each set of historical feature imprints will be associated with its corresponding historical dissemination time to facilitate subsequent analysis of how the feature imprints change over time.

[0132] Step S1444: For the current differential feature imprint, determine its feature imprint type and specific content description.

[0133] Identify the type of feature imprint that needs to be verified, such as whether it belongs to keyframe feature imprint or audio feature imprint, and describe the specific content of the feature imprint in detail so as to make accurate comparison with historical feature imprints.

[0134] Step S1445: Extract historical feature imprints of the same type as the current differential feature imprints from each historical feature imprint set.

[0135] Traverse all historical feature imprint sets and extract historical feature imprints that are the same type as the current difference feature imprints in order to perform targeted similarity comparisons.

[0136] Step S1446: Compare the specific content description of the current difference feature imprint with the specific content description of the historical feature imprint, and determine the comparison dimension according to the feature imprint type.

[0137] Based on the feature imprint type of the current differential feature imprint, the corresponding comparison dimensions are determined. For example, for keyframe feature imprints, the comparison dimensions may include the shape, color, and positional distribution of image elements; for audio feature imprints, the comparison dimensions may include the composition of frequency combinations and the intensity ratio of each frequency component; for text-based feature imprints, the comparison dimensions may include the semantics of keywords and the collocation patterns of sentence components.

[0138] For keyframe feature imprints, the comparison dimensions include the shape, color, and positional distribution of image elements; for audio feature imprints, the comparison dimensions include the numerical range and persistence pattern of frequency combinations; for subtitle feature imprints, the comparison dimensions include the textual composition and semantic tendency of the descriptive segment; for pixel feature imprints, the comparison dimensions include the proportion and distribution area of ​​color channel combinations; for texture feature imprints, the comparison dimensions include the line direction and repetition density of the pattern structure; for region feature imprints, the comparison dimensions include the shape of the content identifier and the size of the area it occupies; for keyword feature imprints, the comparison dimensions include the textual composition of the keyword and its context of occurrence; for sentence structure feature imprints, the comparison dimensions include the grammatical relationship of component combinations and sentence length; for semantic theme feature imprints, the comparison dimensions include the expression method of the core viewpoint and the field involved.

[0139] Step S1447: Calculate the similarity score between the current difference feature imprint and each historical feature imprint according to the comparison dimension, and count the number of historical feature imprints whose similarity score exceeds the preset similarity threshold among all historical feature imprints.

[0140] For each defined comparison dimension, a corresponding scoring criterion is set. The current differential feature imprint is compared with each historical feature imprint along each dimension, and a corresponding similarity score is assigned to each dimension according to the scoring criterion. The similarity scores of each dimension are added together to obtain the total similarity score between the current differential feature imprint and the historical feature imprint. A preset similarity threshold is set, and the number of historical feature imprints whose total similarity score exceeds the threshold is counted.

[0141] Step S1448: Calculate the proportion of the number of historical feature imprints exceeding the preset similarity threshold to the total number of historical feature imprints of this type, and use this proportion as the similarity between the current difference feature imprint and the historical dissemination content feature imprint of the transmission object.

[0142] The proportion of the number of historical feature imprints exceeding the preset similarity threshold obtained in calculation step S1447 to the total number of historical feature imprints of this type is the degree of similarity between the current difference feature imprint and the historical dissemination content feature imprint of the transmission object. The higher the value, the more similar the current difference feature imprint is to the historical dissemination feature imprint of the transmission object.

[0143] Step S145: If the similarity meets the preset conditions, further check whether there is any content marked as illegal in the historical dissemination content of the transmission object. If so, record the association between the feature imprint corresponding to the illegal content and the current difference feature imprint.

[0144] The calculated similarity is compared with preset conditions. If the similarity meets the preset conditions (e.g., greater than or equal to a preset ratio), the historical dissemination records of the target are further reviewed to check for any content marked as illegal by relevant regulatory agencies or systems. If illegal content is found, the characteristic imprints corresponding to the illegal content are extracted, and their correlation with the current difference characteristic imprints is analyzed, such as the similarity in features, whether they belong to the same type of illegality, etc., and the above correlation is recorded in detail.

[0145] Step S146: Retrieve the pre-stored illegal feature imprint library, compare the current differential feature imprint with the illegal feature imprint in the illegal feature imprint library, and calculate the matching degree.

[0146] Step S1461: Retrieve the pre-stored illegal feature imprint library. The illegal feature imprint library stores illegal feature imprints according to feature imprint type. Each illegal feature imprint contains a specific content description and a corresponding illegal category label.

[0147] The system retrieves a pre-stored illegal signature database from its database. This database contains a large number of known illegal signatures, which are categorized and stored according to signature type. Each illegal signature includes a detailed description and a corresponding illegal category label, such as "violent content," "false advertising," or "vulgar information."

[0148] Step S1462: Determine the feature imprint type of the current differential feature imprint, and retrieve all illegal feature imprints of that type from the illegal feature imprint library.

[0149] Based on the current differential feature imprint type, find the corresponding category in the illegal feature imprint database and retrieve all illegal feature imprints stored under that category.

[0150] Step S1463: For each retrieved illegal feature imprint, compare it with the current differential feature imprint in a dimension-by-dimensional manner. The comparison dimensions are consistent with the comparison dimensions of historical feature imprints.

[0151] For each illegal signature retrieved from the illegal signature database, a detailed comparison is performed with the current differential signature, using the same comparison dimensions as historical signatures.

[0152] Step S1464: Calculate the matching score for each dimension based on the comparison results. If the dimensions are completely identical, the matching score is full. If they are partially identical, the score is calculated according to the similarity ratio. If they are completely inconsistent, the matching score is zero.

[0153] A full score is set for each comparison dimension. If the current difference feature imprint and the illegal feature imprint are completely identical in a certain dimension, the matching score for that dimension is full. If some content is identical, the corresponding score is calculated based on the proportion of the identical part. If they are completely different, the matching score for that dimension is zero.

[0154] Step S1465: Sum the matching scores of all dimensions to obtain the total matching score between the current differential feature imprint and the illegal feature imprint.

[0155] The matching scores of the current differential feature imprint and an illegal feature imprint across all comparison dimensions are added together to obtain the total matching score, which reflects the overall matching degree between the current differential feature imprint and the illegal feature imprint.

[0156] Step S1466: Count the number of illegal feature imprints whose total matching score exceeds the preset matching threshold among all illegal feature imprints, and calculate the proportion of the number of illegal feature imprints exceeding the preset matching threshold to the total number of illegal feature imprints of this type.

[0157] A preset matching threshold is set, and the number of illegal feature imprints whose total matching score exceeds this threshold is counted. The proportion of this number to the total number of illegal feature imprints of this type is calculated. This proportion reflects the overall matching situation between the current differential feature imprints and illegal feature imprints of this type in the illegal feature imprint database.

[0158] Step S1467: Simultaneously, calculate the average total matching score between the current differential feature imprint and all illegal feature imprints exceeding the threshold. Combine the ratio with the average value and calculate the matching degree between the current differential feature imprint and the illegal feature imprints in the illegal feature imprint library according to the preset weight.

[0159] Calculate the average of the total matching scores of the current differential feature imprint and all illegal feature imprints whose total matching scores exceed a preset matching threshold. Then, weight the ratio calculated in step S1466 with this average value according to preset weights to obtain the final matching degree between the current differential feature imprint and the illegal feature imprints in the illegal feature imprint library.

[0160] Step S147: If the matching degree exceeds the preset threshold, mark the difference feature mark as a suspected illegal feature mark and record the type of illegal feature mark matched; if the matching degree does not exceed the preset threshold, combine the historical dissemination content association of the transmission object, if there is illegal content association, mark it as a suspected illegal feature mark, if not, mark it as a normal difference feature mark.

[0161] The calculated matching degree is compared with a preset threshold. If the matching degree exceeds the preset threshold, the difference feature imprint is directly marked as a suspected illegal feature imprint, and the type of illegal feature imprint it matches is recorded. If the matching degree does not exceed the preset threshold, the judgment is made in conjunction with the historical dissemination content association of the transmission object obtained in step S145. If there is illegal content association, the difference feature imprint is also marked as a suspected illegal feature imprint; if there is no illegal content association, it is marked as a normal difference feature imprint.

[0162] Step S148: After marking all differential feature marks, count the number of suspected illegal feature marks and their corresponding feature mark types.

[0163] After marking all the differential feature imprints, count the total number of suspected illegal feature imprints, and record the feature imprint type corresponding to each suspected illegal feature imprint, such as how many keyframe-type suspected illegal feature imprints, how many audio-type, etc.

[0164] Step S149: Analyze the distribution of suspected illegal feature imprints in the feature imprint evolution trajectory, and determine the transmission node where the suspected illegal feature imprint first appears and the transmission node for subsequent diffusion.

[0165] By examining the evolutionary trajectory of signatures, the distribution of all suspected illegal signatures within the feature evolution units of each adjacent node is analyzed. The adjacent node feature evolution unit where the suspected illegal signature first appears is identified, thus determining the initial transmission node (i.e., the subsequent transmission node of that evolution unit). Simultaneously, the existence of the suspected illegal signature in subsequent evolution units is tracked to determine which transmission nodes it subsequently spread to.

[0166] Step S1410: Based on the type, quantity and distribution of suspected illegal feature marks, determine the illegal content category of the digital multimedia content. The illegal content category is determined according to the type of illegal feature mark corresponding to the suspected illegal feature mark.

[0167] Taking into account the type (e.g., violence, false advertising), quantity (total quantity and quantity of each category) of suspected illegal signatures, as well as their distribution across dissemination nodes (first appearance node and diffusion node), the illegal content category of the digital multimedia content is determined. The determination of the illegal content category primarily relies on the type of illegal signature matched with the suspected illegal signature. If multiple types of suspected illegal signatures exist, the type with the highest proportion or the most serious violation is taken as the primary illegal content category.

[0168] Step S1411: Record the illegal content category, the specific content of the suspected illegal feature imprint, the first occurrence node and the diffusion node information to form the illegal content identification result of the digital multimedia content.

[0169] The illegal content categories identified, the specific descriptions of all suspected illegal features, the transmission node identifiers of the first appearance of suspected illegal features, and the transmission node identifiers of subsequent spread are recorded in detail and integrated to form an illegal content identification result report for the digital multimedia content.

[0170] Step S150: Based on the illegal content identification results, generate a digital multimedia illegal content handling scheme that includes content blocking link identifiers and content modification tracing instructions, and send the digital multimedia illegal content handling scheme to the management terminals of each dissemination link.

[0171] Step S151: Analyze the illegal content identification results and extract the illegal content category, the transmission node identifier of the first appearance of the suspected illegal feature imprint, and the transmission node identifier of subsequent diffusion.

[0172] The illegal content identification report is analyzed to extract key information, including the illegal content category of the digital multimedia content, the unique identifier of the transmission node where the suspected illegal feature first appears, and the identifiers of each transmission node to which the suspected illegal feature subsequently spreads.

[0173] Step S152: The transmission node identifier where the suspected illegal feature imprint first appears is identified as the illegal feature origin node identifier, and the subsequent transmission node identifiers are identified as illegal feature diffusion node identifiers. The origin node identifier and the diffusion node identifier are integrated to form the content blocking node identifier.

[0174] The node at which suspected illicit signatures first appear is identified as the origin of the illicit signature, representing the source of the signature. Subsequent nodes are identified as the diffusion stages, serving as channels for further dissemination of illicit content. The origin and diffusion stages are then combined to form a content blocking stage identifier, clearly defining the specific stages where content blocking is necessary.

[0175] Step S153: Based on the category of illegal content, retrieve the preset content modification tracing rules, which include the modification operation type to be traced and the corresponding tracing method.

[0176] Based on the identified category of illegal content, the system retrieves pre-defined content modification tracing rules from its rule base. These rules specify the types of modifications that need to be traced for different categories of illegal content, such as "subtitle text modification" and "audio replacement," as well as the corresponding tracing methods, such as "finding modification records within a specific time period" and "locating the person who performed the modification and the approval process."

[0177] Step S154: Based on the modification records of the digital multimedia content, determine the modification operations related to the suspected illegal feature imprints, and extract the execution time, execution object, and modification content details of the modification operations.

[0178] This study analyzes the correlation between modification records of digital multimedia content and suspected illegal signatures to identify modification operations associated with these signatures. Detailed information about these modification operations is extracted, including the execution time, the object performing the operation (e.g., the specific operator or system), and specific details of the modified content, such as a comparison of the content before and after the modification.

[0179] Step S155: Generate a content modification traceability instruction according to the traceability method in the traceability rules. The instruction includes the identifier of the modification operation to be traced, the traceability information type, and the traceability result feedback requirements.

[0180] Based on the tracing methods specified in the tracing rules, a content modification tracing instruction is generated for the extracted modification operations related to suspected illegal features. This content modification tracing instruction includes a unique identifier of the modification operation to be traced, the type of information to be traced (such as modifying the executor's identity, modifying approval records, modifying the original materials on which the modification is based, etc.), and feedback requirements for the tracing results, such as the time limit for feedback and the format of the feedback information.

[0181] Step S156: Integrate the content blocking link identifier with the content modification traceability instruction, and add the generation time and solution number of the disposal plan. The solution number is generated by combining the generation time and the unique identifier of the digital multimedia content.

[0182] By integrating content blocking step identifiers with content modification tracing instructions, a preliminary solution for handling illegal digital multimedia content is formed. A generation time, accurate to the second, is added to this solution to record the moment of its creation. Simultaneously, a solution number is generated, composed of the solution generation time and a unique identifier for the digital multimedia content (such as the content's hash value or file ID), ensuring the uniqueness of the solution number.

[0183] Step S157: Check the clarity and unambiguity of each information item in the integrated digital multimedia illegal content handling plan, and adjust the information items that are vague or ambiguous.

[0184] The integrated plan for handling illegal digital multimedia content undergoes content review, with a focus on checking the clarity and accuracy of each information item and identifying any ambiguities or potential for misunderstanding. Any vague or ambiguous information items are promptly adjusted and corrected to ensure the accuracy and comprehensibility of the plan's content.

[0185] Step S158: Determine the address of the propagation link management terminal to which each transmission node belongs, corresponding to the content blocking link identifier; encapsulate the adjusted illegal digital multimedia content handling plan according to the preset format; and send the encapsulated illegal digital multimedia content handling plan to the corresponding propagation link management terminal.

[0186] Based on the identifiers of each transmission node in the content blocking stage identifier, locate and determine the network address of the propagation stage management terminal to which each transmission node belongs in the system configuration information. Encapsulate the adjusted illegal digital multimedia content handling plan according to a preset format (such as encrypted XML or JSON format) to ensure the security and integrity of the plan during transmission. Send the encapsulated handling plan to the corresponding propagation stage management terminal via a secure communication protocol.

[0187] Step S159: Record the scheme reception time and reception confirmation information of each management terminal to form a processing scheme transmission record, and associate it with the illegal content identification result of the digital multimedia content.

[0188] After sending the handling plan to each management terminal, record the time when each terminal successfully receives the plan and the returned confirmation information, such as a confirmation code or confirmation message. Associate these sending records with the illegal content identification results of the digital multimedia content for subsequent tracking and auditing of the handling plan's execution.

[0189] Figure 2 This application illustrates an intelligent identification system 100 for illegal content in digital multimedia, comprising a processor 1001, a memory 1003, and program code stored in the memory 1003. The processor 1001 executes the program code to implement the steps of the intelligent identification method for illegal content in digital multimedia. The processor 1001 and the memory 1003 are connected, for example, via a bus 1002. Optionally, the intelligent identification system 100 may further include a transceiver 1004, which can be used for data interaction between this intelligent identification system and other intelligent identification systems for illegal content in digital multimedia, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of this intelligent identification system 100 for illegal content in digital multimedia does not constitute a limitation on the embodiments of this application.

[0190] The memory 1003 is used to store program code for executing the embodiments of this application, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0191] This application provides a computer-readable storage medium storing program code, which, when executed by a processor, can implement the steps and corresponding content of the aforementioned method embodiments.

[0192] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent identification of illegal content in digital multimedia, characterized in that, The method includes: The digital multimedia content to be identified and its dissemination traceability information are obtained. The digital multimedia content includes the original content data and modification records during the dissemination process. The dissemination traceability information includes the transmission nodes, transmission time and transmission objects of the digital multimedia content at each dissemination stage. Based on the transmission nodes and transmission times in the propagation tracing information, the content feature imprints corresponding to each transmission node are extracted from the original content data and modification records of the digital multimedia content to generate a feature imprint set for each transmission node. According to the transmission time sequence, the evolutionary relationship between the feature imprint sets of adjacent transmission nodes is tracked, the differences in feature imprints during the evolution process are marked, and the feature imprint evolution trajectory is formed; By combining the transmission object information in the propagation tracing information, the differential feature imprints in the feature imprint evolution trajectory are correlated and verified to determine whether the differential feature imprints belong to illegal feature imprints, and the illegal content identification result of the digital multimedia content is obtained. Based on the illegal content identification results, a digital multimedia illegal content handling scheme is generated, which includes content blocking link identifiers and content modification tracing instructions, and the digital multimedia illegal content handling scheme is sent to the management terminals of each dissemination link.

2. The intelligent identification method for illegal content in digital multimedia according to claim 1, characterized in that, The step involves extracting content feature imprints corresponding to each transmission node from the original content data and modification records of the digital multimedia content based on the transmission nodes and transmission times in the propagation tracing information, and generating a feature imprint set for each transmission node, including: The propagation and tracing information is analyzed, and the node identifiers of all transmission nodes and the corresponding transmission times of each transmission node are extracted. The transmission nodes are then arranged in chronological order to form a sequence of transmission nodes. Separate the original content data and modification records of the digital multimedia content, extract the modification execution time, modification operation content and post-modification content feature description from the modification records, and sort them according to the modification execution time to form a modification record sequence; The transmission time of each transmission node in the transmission node sequence is compared with the modification execution time in the modification record sequence to determine the content status corresponding to each transmission node. If the transmission time of the transmission node is earlier than all modification execution times, then the transmission node corresponds to the original data status of the content. If the transmission time of the transmission node is between two modification execution times, then the transmission node corresponds to the modified content status of the previous modification record. If the transmission time of the transmission node is later than all modification execution times, then the transmission node corresponds to the modified content status of the last modification record. For each transmission node, extract content feature imprints based on the presentation format of the digital multimedia content, according to the content status corresponding to each transmission node. If the presentation format is video, extract the stable image elements that appear in the keyframes of the video under the corresponding content state as keyframe feature marks, extract the frequency combinations that exist continuously in the audio as audio feature marks, and extract the repeated expression segments in the subtitle text as subtitle feature marks. If the presentation format is an image, extract the color channel combination with a stable proportion in the pixel distribution of the image under the corresponding content state as the pixel feature imprint, extract the recurring pattern structure in the texture as the texture feature imprint, and extract the content identifier that is fixed in the region as the region feature imprint. If the presentation format is text, extract the frequently occurring keywords in the text under the corresponding content state as keyword feature marks, extract the fixed component collocation patterns in the sentence as sentence structure feature marks, and extract the core viewpoint expressions in the paragraph as semantic theme feature marks. All content feature imprints corresponding to each transmission node are sorted and arranged according to feature imprint type, and duplicate feature imprints are removed to form the feature imprint set of that transmission node. Add a transmission node identifier and a content status identifier to each feature imprint set, and integrate the feature imprint sets of all transmission nodes in the order of transmission time to generate the feature imprint set of each transmission node.

3. The intelligent identification method for illegal content in digital multimedia according to claim 1, characterized in that, The process of tracing the evolutionary relationships between the feature imprint sets of adjacent transmission nodes according to the transmission time sequence, marking the differences in feature imprints during the evolution process, and forming a feature imprint evolution trajectory includes: From the feature imprint sets generated for each transmission node, select two adjacent feature imprint sets in chronological order of transmission time, and label them as the preceding feature imprint set and the following feature imprint set, respectively. For the preceding feature imprint set and the subsequent feature imprint set, compare them one by one according to the feature imprint type; If the feature imprint type is keyframe feature imprint, compare the keyframe feature imprints in the preceding and following sets, count the proportion of the same screen elements, mark the newly added screen elements and the disappeared screen elements, and record the positions of the newly added and disappeared screen elements in the keyframe. If the feature imprint type is audio feature imprint, compare the audio feature imprints in the preceding and following sets, count the percentage of duration of the same frequency combination, mark the newly added frequency combination and the disappeared frequency combination, and record the time period of the occurrence of the newly added and disappeared frequency combination. If the feature imprint type is subtitle feature imprint, compare the subtitle feature imprints in the preceding and following sets, count the proportion of the same expression segments, mark the newly added expression segments and the disappeared expression segments, and record the sentence positions of the newly added and disappeared expression segments in the subtitle text. If the feature imprint type is pixel feature imprint, compare the pixel feature imprints in the preceding and following sets, calculate the difference in the proportion of the same color channel combination, mark the newly added color channel combination and the disappeared color channel combination, and record the area range of the newly added and disappeared color channel combination in the image. If the feature imprint type is texture feature imprint, compare the texture feature imprints in the preceding and following sets, count the percentage of repetitions of the same pattern structure, mark the newly added pattern structure and the disappeared pattern structure, and record the texture region of the newly added and disappeared pattern structure in the image. If the feature imprint type is a region feature imprint, compare the region feature imprints in the preceding and following sets, count the proportion of the same content identifiers, mark the newly added content identifiers and the disappeared content identifiers, and record the specific locations of the newly added and disappeared content identifiers in the image region. If the feature imprint type is keyword feature imprint, compare the keyword feature imprints in the preceding and following sets, count the percentage of occurrence of the same keyword, mark the newly added and disappeared keywords, and record the paragraph positions of the newly added and disappeared keywords in the text. If the feature imprint type is sentence structure feature imprint, compare the sentence structure feature imprints in the preorder and postorder sets, count the percentage of sentences with the same component collocation pattern, mark the newly added component collocation pattern and the disappeared component collocation pattern, and record the sentence position of the newly added and disappeared component collocation pattern in the text. If the feature imprint type is semantic theme feature imprint, compare the semantic theme feature imprints in the preceding and following sets, count the proportion of the same core viewpoint expressions, mark the newly added core viewpoint expressions and the disappeared core viewpoint expressions, and record the paragraph positions of the newly added and disappeared core viewpoint expressions in the text. The differences in the feature imprint comparison results of adjacent transmission nodes are recorded according to the feature imprint type, forming feature evolution units of adjacent nodes; Traverse all adjacent feature imprint sets of the transmission nodes to generate multiple adjacent node feature evolution units. Integrate all adjacent node feature evolution units in the order of transmission time, add the evolution time span of each unit, and form the feature imprint evolution trajectory.

4. The intelligent identification method for illegal content in digital multimedia according to claim 1, characterized in that, The step of combining the transmission object information in the propagation tracing information to perform correlation verification on the differential feature imprints in the feature imprint evolution trajectory, determining whether the differential feature imprints belong to illegal feature imprints, and obtaining the illegal content identification result of the digital multimedia content includes: Extract the transmission object information corresponding to each transmission node from the propagation tracing information. The transmission object information includes the identity of the transmission object and historical propagation content records. Extract all differential feature imprints from the feature imprint evolution trajectory, and classify and arrange them according to feature imprint type and corresponding adjacent node feature evolution units; For each differential feature imprint, retrieve the transmission node identifier from the adjacent node feature evolution unit corresponding to the differential feature imprint, and find the transmission object information corresponding to the transmission node identifier. Analyze the historical dissemination content records in the information of the target audience, extract the feature imprints contained in the historical dissemination content, compare them with the current difference feature imprints, and calculate the similarity between the two. If the similarity meets the preset conditions, further check whether there is any content marked as illegal in the historical dissemination content of the transmission object. If so, record the association between the feature imprint corresponding to the illegal content and the current difference feature imprint. Retrieve the pre-stored illegal feature imprint database, compare the current differential feature imprint with the illegal feature imprints in the illegal feature imprint database, and calculate the degree of matching. If the matching degree exceeds the preset threshold, the difference feature mark is marked as a suspected illegal feature mark, and the type of illegal feature mark matched is recorded; if the matching degree does not exceed the preset threshold, the historical dissemination content association of the transmission object is considered. If there is an illegal content association, it is also marked as a suspected illegal feature mark; if not, it is marked as a normal difference feature mark. After marking all the differential feature marks, count the number of suspected illegal feature marks and their corresponding feature mark types; Analyze the distribution of suspected illegal signatures in the signature evolution trajectory to determine the transmission node of the first appearance of the suspected illegal signature and the transmission node of subsequent diffusion; Based on the type, quantity, and distribution of suspected illegal signatures, the category of illegal content in the digital multimedia content is determined, wherein the category of illegal content is determined according to the type of illegal signature corresponding to the suspected illegal signature. Record the categories of illegal content, the specific content of suspected illegal features, the information of the first occurrence node and the diffusion node, and form the illegal content identification result of the digital multimedia content.

5. The intelligent identification method for illegal content in digital multimedia according to claim 4, characterized in that, The analysis involves recording historical dissemination content in the target information, extracting feature imprints from the historical dissemination content, comparing them with current difference feature imprints, and calculating the similarity between the two, including: Extract all digital multimedia content that the target has previously disseminated from the historical dissemination content records of the target information, and arrange them in chronological order of dissemination time to form a historical dissemination content sequence. For each piece of historical dissemination content in the historical dissemination content sequence, the corresponding feature imprint is extracted according to its presentation format. The extraction method is consistent with the feature imprint extraction method of current digital multimedia content. The characteristic imprints of each historical dissemination content are classified by type to form a set of historical characteristic imprints, and each set of historical characteristic imprints is associated with the corresponding historical dissemination time. For the current differential feature imprints, determine their feature imprint types and specific content descriptions; Extract historical feature imprints of the same type as the current differential feature imprint from each historical feature imprint set; Compare the specific content descriptions of current differential feature imprints with the specific content descriptions of historical feature imprints, and determine the comparison dimensions according to the feature imprint type; For keyframe feature imprints, the comparison dimensions include the shape, color, and positional distribution of image elements; for audio feature imprints, the comparison dimensions include the numerical range and persistence pattern of frequency combinations; for subtitle feature imprints, the comparison dimensions include the textual composition and semantic tendency of the descriptive segment; for pixel feature imprints, the comparison dimensions include the proportion and distribution area of ​​color channel combinations; for texture feature imprints, the comparison dimensions include the line direction and repetition density of the pattern structure; for region feature imprints, the comparison dimensions include the shape of the content identifier and the size of the area it occupies; for keyword feature imprints, the comparison dimensions include the textual composition of the keyword and its context of occurrence; for sentence structure feature imprints, the comparison dimensions include the grammatical relationship of component combinations and sentence length; for semantic theme feature imprints, the comparison dimensions include the expression method of the core viewpoint and the field involved. Based on the comparison dimensions, calculate the similarity score between the current differential feature imprint and each historical feature imprint, and count the number of historical feature imprints whose similarity scores exceed the preset similarity threshold among all historical feature imprints. Calculate the proportion of historical feature imprints exceeding a preset similarity threshold to the total number of historical feature imprints of that type, and use this proportion as the similarity between the current difference feature imprint and the historical dissemination content feature imprint of the transmission object.

6. The intelligent identification method for illegal content in digital multimedia according to claim 4, characterized in that, The process of retrieving the pre-stored illegal feature imprint database, comparing the current differential feature imprint with the illegal feature imprints in the database, and calculating the degree of matching includes: Retrieve a pre-stored illegal feature imprint library, which stores illegal feature imprints according to feature imprint type. Each illegal feature imprint contains a specific content description and a corresponding illegal category label. Determine the feature imprint type of the current differential feature imprint, and retrieve all illegal feature imprints of that type from the illegal feature imprint database; For each retrieved illegal feature imprint, it is compared with the current differential feature imprint in a dimension-by-dimensional manner, and the comparison dimensions are consistent with the comparison dimensions of historical feature imprints. Based on the comparison results of each dimension, calculate the matching score for that dimension. If the content of the dimension is completely consistent, the matching score is full. If it is partially consistent, the score is calculated according to the consistency ratio. If it is completely inconsistent, the matching score is zero. Add up the matching scores of all dimensions to get the total matching score between the current differential feature imprint and the illegal feature imprint; Count the number of illegal feature marks whose total matching score exceeds the preset matching threshold among all illegal feature marks, and calculate the proportion of the number of illegal feature marks exceeding the preset matching threshold to the total number of illegal feature marks of this type. Simultaneously, the average total matching score of the current differential feature imprint and all illegal feature imprints exceeding the threshold is calculated. The ratio and the average value are combined and the matching degree between the current differential feature imprint and the illegal feature imprints in the illegal feature imprint library is calculated according to the preset weight.

7. The intelligent identification method for illegal content in digital multimedia according to claim 1, characterized in that, Based on the illegal content identification results, a digital multimedia illegal content handling scheme is generated, which includes content blocking link identifiers and content modification tracing instructions. This scheme is then sent to the management terminals at each dissemination link, including: Analyze the illegal content identification results and extract the illegal content category, the transmission node identifier of the first appearance of the suspected illegal feature imprint, and the transmission node identifier of subsequent diffusion; The transmission node where the first suspected illegal feature appears is identified as the illegal feature origin node identifier, and the transmission node identifiers of subsequent diffusion are identified as illegal feature diffusion node identifiers. The origin node identifier and the diffusion node identifier are integrated to form the content blocking node identifier. Based on the category of illegal content, the preset content modification tracing rules are retrieved. The tracing rules include the types of modification operations to be traced and the corresponding tracing methods. Based on the modification records of the digital multimedia content, the modification operations related to the suspected illegal signatures are identified, and the execution time, execution object, and modification content details of the modification operations are extracted. Based on the tracing method in the tracing rules, a content modification tracing instruction is generated. The instruction includes the identifier of the modification operation to be traced, the type of tracing information, and the requirements for tracing result feedback. Integrate the content blocking process identifier with the content modification traceability instruction, and add the generation time and plan number of the disposal plan. The plan number is generated by combining the generation time and the unique identifier of the digital multimedia content. Check the clarity and unambiguity of the descriptions of each information item in the integrated digital multimedia illegal content handling plan, and adjust the descriptions of information items that are vague or ambiguous. Determine the address of the propagation link management terminal to which each transmission node belongs, corresponding to the content blocking link identifier; encapsulate the adjusted illegal digital multimedia content handling plan according to the preset format; and send the encapsulated illegal digital multimedia content handling plan to the corresponding propagation link management terminal. Record the time of receiving the plan and the confirmation information of each management terminal to form a record of the handling plan sending, and link it to the illegal content identification result of the digital multimedia content.

8. The intelligent identification method for illegal content in digital multimedia according to claim 2, characterized in that, If the presentation format is video, extract consistently occurring image elements from the keyframes of the video in the corresponding content state as keyframe feature marks, extract persistent frequency combinations from the audio as audio feature marks, and extract recurring descriptive segments from the subtitle text as subtitle feature marks, including: For videos with corresponding content states, keyframes are extracted from the video at preset time intervals to form a keyframe sequence; Traverse each keyframe in the keyframe sequence and use image recognition technology to extract the image elements in each keyframe, including object outlines, color blocks and text labels. The number of keyframes in which each image element appears in the keyframe sequence is counted, the proportion of the number of occurrences to the total number of frames in the keyframe sequence is calculated, and image elements with a proportion exceeding a preset stable threshold are selected. The selected image elements with a proportion exceeding the preset stable threshold are arranged according to image element type to form keyframe feature imprints. For the video and audio portions of the corresponding content state, the audio is segmented according to preset time segments to form an audio segment sequence; Frequency analysis is performed on each audio segment to extract frequency combinations from each audio segment. The frequency combinations include frequency components and the intensity ratio of each component. The number of audio segments in which each frequency combination appears in the audio segment sequence is counted, the proportion of the number of occurrences to the total number of segments in the audio segment sequence is calculated, frequency combinations with a proportion exceeding a preset duration threshold are selected, and the selected frequency combinations with a proportion exceeding the preset duration threshold are sorted by frequency range to form audio feature imprints. For the video subtitle text under the corresponding content state, the subtitle text is segmented by sentence to form a subtitle sentence sequence, and the expression fragments in each subtitle sentence are extracted. The expression fragments include phrases, short sentences and expression formats. The number of sentences in the subtitle sequence for each expression segment is counted, and the proportion of the number of occurrences to the total number of sentences in the subtitle sequence is calculated. Expression segments with a proportion exceeding a preset repetition threshold are selected, and the selected expression segments with a proportion exceeding the preset repetition threshold are arranged in sentence order to form subtitle feature marks.

9. The intelligent identification method for illegal content in digital multimedia according to claim 2, characterized in that, If the presentation format is text-based, the high-frequency keywords in the text under the corresponding content state are extracted as keyword feature marks, fixed component collocation patterns in sentences are extracted as sentence structure feature marks, and core viewpoint expressions in paragraphs are extracted as semantic theme feature marks, including: For the text in the corresponding content state, perform word segmentation to obtain an effective word sequence; The frequency of each effective word in the effective word sequence is counted, the proportion of each effective word to the total frequency of all effective words is calculated, and effective words with a proportion exceeding the preset high-frequency threshold are selected. The selected effective words with a proportion exceeding the preset high-frequency threshold are sorted by word frequency to form keyword feature imprints. The sentences in the text are subjected to grammatical analysis to determine the sentence components of each sentence, which include subject, predicate, object, attributive, adverbial and complement. Extract the collocation pattern of sentence components in each sentence, the collocation pattern including the combination order of components and the part of speech combination of each component; Count the number of sentences in which each collocation pattern appears in the text, and calculate the proportion of the number of occurrences to the total number of sentences in the text. The collocation patterns with a proportion exceeding a preset fixed threshold are selected and arranged according to grammatical structure type to form sentence structure feature imprints. The text is divided into paragraphs to form a paragraph sequence. For each paragraph, the core viewpoint statement that expresses the core viewpoint is extracted. The core viewpoint statement includes the first sentence of the paragraph, the last sentence of the paragraph, and a statement containing summarizing words. Semantically summarize the extracted core viewpoint statements to obtain the core viewpoint expression of each paragraph. Count the number of paragraphs in which each core viewpoint expression appears in the core viewpoint expressions of all paragraphs, and calculate the proportion of the number of occurrences to the total number of paragraphs in the paragraph sequence. Core viewpoints exceeding a preset core threshold are selected and sorted by their relevance to form semantic theme feature imprints.

10. An intelligent identification system for illegal content in digital multimedia, characterized in that, The method includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the intelligent identification method for illegal content in digital multimedia as described in any one of claims 1-9.

Images