WO2014101539A1 - 一种基于感知知识库的视频内容审查系统及方法 - Google Patents

一种基于感知知识库的视频内容审查系统及方法 Download PDF

Info

Publication number
WO2014101539A1
WO2014101539A1 PCT/CN2013/085484 CN2013085484W WO2014101539A1 WO 2014101539 A1 WO2014101539 A1 WO 2014101539A1 CN 2013085484 W CN2013085484 W CN 2013085484W WO 2014101539 A1 WO2014101539 A1 WO 2014101539A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
knowledge
video
sample
pending
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2013/085484
Other languages
English (en)
French (fr)
Inventor
朱定局
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Institute of Advanced Technology of CAS
Original Assignee
Shenzhen Institute of Advanced Technology of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Institute of Advanced Technology of CAS filed Critical Shenzhen Institute of Advanced Technology of CAS
Publication of WO2014101539A1 publication Critical patent/WO2014101539A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/70—Information retrieval; Database structures therefor; File system structures therefor of video data
    • G06F16/78—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
    • G06F16/783—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
    • G06F16/7837—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using objects detected or recognised in the video content

Definitions

  • the present invention relates to multimedia processing technologies, and more particularly to a video content review system and method based on a perceptual knowledge base.
  • video content review is basically a manual viewing method.
  • a video program requires the user to manually read it from start to finish.
  • multiple users need to watch it multiple times, which takes a lot of time and manpower, and the efficiency is low.
  • the prior art also proposes algorithms for identifying specific video content, such as a particular type of nude.
  • the prior art algorithms have poor versatility, and are not identifiable for video content that changes slightly or does not conform to a predefined type, and is less practical.
  • the present invention provides a video content review system based on a perceptual knowledge base, comprising: a video sample database, comprising at least one video sample sub-bank, respectively storing and managing sample videos corresponding to at least one type of specific review categories; sample data
  • the sensing feature extraction module extracts the perceptual feature of the sample video in the video sample sub-database frame by frame, and generates a sample feature description corresponding to each frame; summarizes the sample feature description to obtain sample feature knowledge, and defines the value of the sample feature knowledge.
  • Scope video-aware knowledge base, storing and managing sample feature knowledge and its value range; pending data-aware feature extraction module, the perceived feature is extracted frame by frame, and the pending feature description corresponding to each frame is generated;
  • the feature fuzzy matching module searches the video perception knowledge base for the sample feature knowledge closest to each pending feature description;
  • the video data type fuzzy discriminating module analyzes the found sample feature knowledge and determines the pending feature description corresponding to the pending video frame.
  • Specific review category to which it belongs user view a data type library, storing and managing a pending video frame and a specific review category and a review result comprehensive query module thereof, and querying a pending video frame in the user video data type library and a specific review category to which it belongs according to a user instruction; and/ Or according to the specific type of video alarm setting specified by the user, when the pending video contains a frame that meets the specific type of video alarm setting, the user is alerted.
  • the sample data sensing feature extraction module performs perceptual feature extraction on at least one feature of each frame of the sample video to obtain at least one component of the sample feature description corresponding to each frame.
  • the sample data perceptual feature extraction module aggregates the sample feature descriptions to obtain the sample feature knowledge, which is a weighted average of at least one sub-item of the sample feature description, and obtains at least one sub-item sample feature knowledge; the sample data perceptual feature extraction module respectively
  • the minimum and maximum values of the feature knowledge of each sub-item are taken as the range of values of the sample feature knowledge.
  • the video-aware knowledge base includes a video-aware knowledge sub-library corresponding to at least one type of specific review category, and respectively stores the item sample feature knowledge corresponding to the specific review category and the value range thereof;
  • the to-be-reviewed data-aware feature extraction module performs perceptual feature extraction on at least one feature of each frame of the video to be examined, and generates at least one item of the pending feature description corresponding to each frame.
  • the video data type fuzzy discriminant module compares whether each sub-item of the pending feature description falls into the sample feature.
  • the range of values of the corresponding sub-items of knowledge is to determine that the pending feature description belongs to a specific type corresponding to the sample feature knowledge. Otherwise, the perceptual feature fuzzy matching module searches all the sub-items from the video-perceived knowledge base to correspond to the pending feature description.
  • the video data type fuzzy discriminant module compares whether each sub-item of the pending feature description falls within the range of the corresponding sub-item of the sample feature knowledge, and determines the pending feature. Describe the specific type corresponding to the next-closed sample feature knowledge. Otherwise, the perceptual feature fuzzy matching module continues to search for the next close sample feature knowledge until it determines the specific type to which the pending feature description belongs, or completes all sample features in the video-aware knowledge base. Knowledge search.
  • the video data type The fuzzy discriminant module respectively obtains the total distance of the pending feature description and the first sample feature knowledge, the total distance of the pending feature description and the second sample feature knowledge; the video data type fuzzy discriminating module compares the pending feature description with the first The total distance of the feature knowledge and the total distance between the pending feature description and the second sample feature knowledge, from the sample feature knowledge with a small total distance to the sample feature knowledge with a large total distance, the video data type fuzzy discriminant module is compared pending Each sub-item of the feature description falls within the range of values of the corresponding sub-item of the feature knowledge, thereby determining the specific type of the feature knowledge to which the pending feature description belongs; if the sub-items of the pending feature description are The range of values of the corresponding sub-items
  • the perceptual feature fuzzy matching module searches the video-aware knowledge base for the first sample feature knowledge that the sub-item of the pending feature description is close to the corresponding sub-item, and the other sub-items are corresponding to the corresponding sub-items.
  • the second sample feature knowledge close to the item is repeated, and the total distance is repeatedly determined and compared until the specific type to which the pending feature description belongs is determined, or the search for all sample feature knowledge in the video-aware knowledge base is completed.
  • the video data type fuzzy discriminating module respectively obtains the total distance between the pending feature description and each sample feature knowledge, respectively, that the sub-items of the pending feature description are respectively closest to the corresponding one-item of the feature knowledge in the video-aware knowledge base.
  • the video data type fuzzy discriminating module compares the total distance between the pending feature description and each sample feature knowledge, from the sample feature knowledge with the smallest total distance to the sample feature knowledge with the largest total distance, and the video data type fuzzy discriminant module compares the pending feature descriptions respectively.
  • Each sub-item falls within the range of values of the corresponding sub-items of the feature knowledge, thereby determining which specific type of feature knowledge corresponds to the pending feature description; if the sub-items of the pending feature description are not uniformly Entering the value range of any corresponding item of the feature knowledge, the perceptual feature fuzzy matching module searches the video perception knowledge base for at least the same feature knowledge that the sub-item of the pending feature description is close to the corresponding sub-item, and repeats Find the total distance and compare the judgments until the specific type to which the pending feature description belongs is determined, or complete the pair Find all frequency perception Knowledge Base sample characteristics of knowledge.
  • the present invention also provides a method for reviewing a video content based on a perceptual knowledge base, comprising: a step that the video sample database includes at least one video sample sub-bank, each pre-storing and managing sample videos corresponding to at least one type of specific review category; sample data The step of extracting, by the perceptual feature extraction module, the pre-stored sample video in the video sample sub-bank, performing perceptual feature extraction according to at least one feature of each frame, and obtaining at least one sub-item of the sample feature description corresponding to each frame; The data-aware feature extraction module weights and averages at least one item of the sample feature description to obtain at least one item sample feature knowledge; respectively, the minimum value and the maximum value of each item sample feature knowledge are taken as the value range of the sample feature knowledge The step of the video-awareness knowledge base storing the item sample feature knowledge corresponding to the specific review category and its value range in the video-aware knowledge sub-library corresponding to at least one type of specific review category; The module separately senses at least one feature of
  • the video data type fuzzy discriminating module analyzes at least the same characteristic knowledge found, and determines the specific review category to which the corresponding pending video frame belongs: the video data type fuzzy discriminating module analyzes the pending feature description The step closest to the corresponding sub-items of the sample feature knowledge in the video-awareness knowledge base; the video data type fuzzy discriminating module is relatively pending for the sub-items of the pending feature description being the closest to the corresponding sub-items of the same feature knowledge.
  • each sub-item of the feature description falls within the range of values of the corresponding sub-item of the sample feature knowledge, thereby determining that the pending feature description belongs to a specific type corresponding to the sample feature knowledge; and each item of the pending feature description is respectively at least The corresponding component of the two-sample feature knowledge is closest, and the video data type fuzzy discriminating module separately obtains the total distance of the pending feature description and each sample feature knowledge; the video data type fuzzy discriminating module compares the pending feature description with each sample.
  • the video data type fuzzy discriminant module compares each of the sub-items of the pending feature description into the range of values of the corresponding sub-items of the feature knowledge, thereby determining where the pending feature description belongs. a specific type of step corresponding to a sample feature knowledge; if the sub-items of the pending feature description fail to fall within the range of values of the corresponding component feature knowledge, the perceptual feature fuzzy matching module is from the video-aware knowledge base.
  • Searching the pending feature description at least the same feature knowledge that each sub-item is close to the corresponding sub-item, repeating the total distance and comparing the judgments until determining the specific type to which the pending feature description belongs, or completing all samples in the video-aware knowledge base The step of searching for feature knowledge.
  • the invention automatically extracts the perceptual knowledge of the sample video data and adopts the database management, and automatically reviews the video content by using the perceptual knowledge, and achieves the purpose of fully automatic review, without manual operation, without writing a specific review algorithm, and has strong versatility.
  • FIG. 1 is a schematic structural diagram of a video content review system based on a perceptual knowledge base according to an embodiment of the present invention
  • FIG. 2 is a schematic flowchart of a video content review method based on a perceptual knowledge base according to an embodiment of the present invention
  • FIG. 3 is a schematic flowchart of determining a specific review category portion to which a pending video frame belongs according to an embodiment of the present invention.
  • FIG. 1 a schematic diagram of a system structure of a video content review system based on a perceptual knowledge base and a schematic diagram of a main process of video content review shown in FIG. 2 , the present invention provides a video content review system based on a perceptual knowledge base. Examples include:
  • a video sample database S comprising at least one video sample sub-library Si, respectively storing and managing sample videos corresponding to at least one type of specific review category;
  • the sample data sensing feature extraction module 1 performs perceptual feature extraction on the sample video in the video sample sub-library Si frame by frame, and generates a sample feature description corresponding to each frame; summarizes the sample feature description to obtain sample feature knowledge, and defines sample features. The range of values of knowledge;
  • the video-aware knowledge base K stores and manages sample feature knowledge and its range of values
  • the data-aware feature extraction module 2 is configured to perform perceptual feature extraction on a frame-by-frame basis, and generate a pending feature description corresponding to each frame;
  • the perceptual feature fuzzy matching module 3 searches for the sample feature knowledge closest to each pending feature description from the video perceptual knowledge base K;
  • the video data type fuzzy discriminating module 4 analyzes the found sample feature knowledge, and determines that the pending feature description corresponds to a specific review category to which the pending video frame belongs;
  • the user video data type library U stores and manages the pending video frame and the specific review category to which it belongs;
  • the review result comprehensive query module 5 queries the pending video frame in the user video data type library U according to the user instruction and the specific review category to which it belongs; and/or according to the specific type of video alarm setting specified by the user, when the pending video includes A frame that meets the specific type of video alert settings will alert the user.
  • the present invention provides an embodiment, in which the sample data sensing feature extraction module 1 performs perceptual feature extraction on at least one feature of each frame of the sample video, and obtains at least one point of the sample feature description corresponding to each frame respectively.
  • the video includes features such as text, image, and/or sound, and the sub-items corresponding to the sample feature description are: text feature description, image feature description, and/or sound feature description.
  • sample data perceptual feature extraction module 1 of the embodiment aggregates the sample feature descriptions to obtain sample feature knowledge, which is a weighted average of at least one sub-item of the sample feature description, and obtains at least one sub-item sample feature knowledge; sample data perceptual feature extraction Module 1 takes the minimum and maximum values for each sub-sample feature knowledge as the value range of the sample feature knowledge.
  • the video-aware knowledge base K includes a video-aware knowledge sub-library Ki corresponding to at least one type of specific review category, and stores the sample-segment feature knowledge corresponding to the specific review category and the value range thereof;
  • the pending data sensing feature extraction module 2 performs perceptual feature extraction on at least one feature of each frame of the video to be examined, and generates at least one item of the pending feature description corresponding to each frame.
  • the present invention provides an embodiment in which the perceptual feature fuzzy matching module 3 and the video data type fuzzy discriminating module 4 cooperate to jointly determine the specific review category to which the pending feature description belongs.
  • the sample feature knowledge found by the perceptual feature fuzzy matching module 3 is divided into the following three cases:
  • the video data type fuzzy discriminating module 4 respectively compares the points of the pending feature description. Whether the items fall within the range of the corresponding sub-items of the sample feature knowledge, and then determine that the pending feature description belongs to a specific type corresponding to the sample feature knowledge; otherwise
  • the perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for the sample feature knowledge that all the sub-items are close to the corresponding sub-items of the pending feature description;
  • the video data type fuzzy discriminating module 4 compares whether each sub-item of the pending feature description falls within the range of the corresponding sub-item of the sub-proportional sample feature knowledge, and determines that the pending feature description belongs to the next close sample feature knowledge. Corresponding to the specific type, otherwise
  • the perceptual feature fuzzy matching module 3 continues to search for the next close sample feature knowledge until it determines the specific type to which the pending feature description belongs, or completes the search for all sample feature knowledge in the video-aware knowledge base K.
  • the sub-item of the pending feature description is closest to the corresponding sub-item of the first sample feature knowledge in the video-aware knowledge base K, and the other sub-items are the most corresponding to the second sample feature knowledge.
  • the video data type fuzzy discriminating module 4 respectively obtains the total distance of the pending feature description and the first sample feature knowledge, the total distance of the pending feature description and the second sample feature knowledge.
  • the video data type fuzzy discriminating module 4 compares the total distance between the pending feature description and the first sample feature knowledge and the total distance between the pending feature description and the second sample feature knowledge, from the sample feature knowledge with a smaller total distance to the total distance.
  • the large sample feature knowledge, the video data type fuzzy discriminating module 4 respectively compares the value ranges of the corresponding sub-items of the feature knowledge in which the sub-items of the pending feature description fall, thereby determining which belongs to the pending feature description.
  • the specific type corresponding to the feature knowledge is the specific type fuzzy discriminating module 4 compared.
  • the perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for the first sample feature knowledge that the sub-item of the pending feature description is close to the corresponding sub-item, and the other sub-items are close to the corresponding sub-items.
  • the two-sample feature knowledge is repeated to find the total distance and compare the judgments until the specific type to which the pending feature description belongs is determined, or the search for all sample feature knowledge in the video-aware knowledge base K is completed.
  • the video data type fuzzy discriminating module 4 respectively obtains the pending feature description and each of the sub-items of the pending feature description that are closest to the corresponding sub-items of the feature knowledge in the video-aware knowledge base K, respectively.
  • the total distance of the sample feature knowledge is the total distance of the sample feature knowledge.
  • the video data type fuzzy discriminating module 4 compares the total distance between the pending feature description and each sample feature knowledge, from the sample feature knowledge with the smallest total distance to the sample feature knowledge with the largest total distance, and the video data type fuzzy discriminating module 4 respectively compares the pending features.
  • Each of the described sub-items falls within the range of values of the corresponding sub-items of the feature knowledge, thereby determining which specific type of feature knowledge the pending feature description belongs to;
  • the perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for at least the same feature knowledge that a sub-item of the pending feature description is close to the corresponding sub-item, and repeatedly obtains the total distance and compares the judgment until the to-be-reviewed feature description is determined.
  • the video sample database S include at least one video sample sub-library Si corresponding to the i-class specific review category.
  • the first video sample sub-library S1 corresponds to the type 1 shooting video
  • the second video sample sub-library S2 corresponds to the two-type sexual intercourse video
  • the third video sample sub-library S3 corresponds to the three-category nude video
  • the fourth video sample sub-repository S4 corresponds to 4 Class bare hip video
  • fifth video sample sub-library S5 corresponds to 5 types of snoring video.
  • the sample data-aware feature extraction module 1 respectively performs text, image and sound characteristics of each frame of the sample video. Performing feature extraction, obtaining text feature description, image feature description and sound feature description corresponding to each frame respectively.
  • the sample data sensing feature extraction module 1 runs the text feature extraction function fw to obtain sample text feature descriptions fw(si1), fw(si2), ..., Fw(sim); running sound feature extraction function fv Obtain sample sound feature descriptions fv(si1), fv(si2), ..., fv(sim); run image feature extraction function fg to obtain sample image feature descriptions fg(si1), fg(si2), ..., fg(sim)) .
  • the sample data perceptual feature extraction module 1 summarizes at least one sub-item of the sample feature description.
  • the sample data perceptual feature extraction module 1 takes the minimum and maximum values of the feature knowledge of each sub-item as the value range of the sample feature knowledge.
  • Text sample feature knowledge minimum value kwi_min min(fw(si1)+fw(si2)+...+fw(sim))
  • text sample feature knowledge maximum value kwi_max max(fw(si1)+fw(si2)+...+ Fw(sim)
  • sound sample feature knowledge minimum value kvi_min min(fv(si1)+fv(si2)+...+fv(sim)
  • sound sample feature knowledge maximum value kvi_max max(fv(si1)+fv (si2)+...+fv(sim))
  • image sample feature knowledge minimum value kgi_min min(fg(si1)+fg(si2)+...+fg(sim))
  • image sample feature knowledge maximum value kgi_max max( Fg(si1)+fg(si2)+...
  • the video-aware knowledge base K includes a plurality of video-aware knowledge sub-libraries Ki, respectively storing sample feature knowledge and its value range: text feature knowledge kwi and its taking The value range (kwi_min, kwi_max), the sound feature knowledge kvi and its range of values (kvi_min, kvi_max), the image feature knowledge kgi and its range of values (kgi_min, kgi_max).
  • FIG. 3 a schematic flowchart of a specific review category portion to which a pending video frame belongs is shown, wherein FIG. 3(a) shows, for the first case, a scheme for determining a specific review category to which the pending video frame vij belongs, including:
  • Step S611 the perceptual feature fuzzy matching module 3 finds a certain sample feature knowledge, all the sub-items of the pending feature description are closest to their corresponding sub-items, proceeding to step S711;
  • the perceptual feature fuzzy matching module 3 finds the first sample feature knowledge ka from the video perceptual knowledge base K, including the first character feature knowledge kwa, the first sound feature knowledge kva, and the first image feature knowledge kga.
  • the pending character feature description fw(vij) is closest to its first character feature knowledge kwa
  • the pending voice feature description fv(vij) is closest to its first sound feature knowledge kva
  • the pending image feature description fg(vij) is first
  • the image feature knowledge kga is the closest.
  • Step S711 the video data type fuzzy discriminating module 4 respectively compares whether each sub-item of the pending feature description falls within the value range of the corresponding sub-item of the sample feature knowledge, and if yes, proceeds to step S712, otherwise proceeds to step S612;
  • the pending character feature description fw(vij) falls within the first character feature knowledge value range (kwa_min, kwa_max), and the pending sound feature description fv(vij) falls within the first sound feature knowledge value range (kva_min, Kva_max) range, And the pending image feature description fg(vij) falls within the first image feature knowledge value range (kga_min, kga_max).
  • step S712 it is determined that the pending video frame vij belongs to the category a specific review category corresponding to the first sample feature knowledge ka, such as a shooting video, and the process ends.
  • Step S612 determining whether the search for all the sample feature knowledge in the video-awareness knowledge base 3 has been completed, if otherwise, proceeding to step S613, if yes, the process ends.
  • Step S613 the perceptual feature fuzzy matching module 3 searches for the sample feature knowledge that all the sub-items are close to the corresponding sub-items of the pending feature description from the video-aware knowledge base K, and proceeds to step 711;
  • the second sample feature knowledge kb close to the pending video frame vij is taken from the video perception knowledge base K.
  • the solution for determining the specific review category to which the pending video frame vij belongs includes:
  • Step S621 The perceptual feature fuzzy matching module 3 finds the first sample feature knowledge ka and the second sample feature knowledge kb, and the corresponding score of a certain item of the pending character feature description fw(vij) and the first sample feature knowledge ka The item is closest, and the other binary items of the pending character feature description fw(vij) are closest to the corresponding binary items of the second sample feature knowledge kn, proceed to step 721;
  • the pending character feature description fw(vij) is closest to the first character feature knowledge kwa
  • the pending voice feature description Fv(vij) is closest to the second sound feature knowledge kvb
  • the pending image feature description fg(vij) is closest to the second image feature knowledge kgb
  • the pending text feature description fw(vij) is closest to the first character feature knowledge kwa
  • the pending voice feature description Fv(vij) is closest to the second sound feature knowledge kvb
  • the pending image feature description fg(vij) is closest to the first image feature knowledge kga
  • the pending text feature description fw(vij) is closest to the second character feature knowledge kwb
  • the pending voice feature description Fv(vij) is closest to the first sound feature knowledge kva
  • the pending image feature description fg(vij) is closest to the first image feature knowledge kga
  • the pending text feature description fw(vij) is closest to the second character feature knowledge kwb
  • the pending voice feature description Fv(vij) is closest to the first sound feature knowledge kva
  • the pending image feature description fg(vij) is closest to the second image feature knowledge kgb
  • the pending text feature description fw(vij) is closest to the second character feature knowledge kwb
  • the pending voice feature description Fv(vij) is closest to the second sound feature knowledge kvb
  • the pending image feature description fg(vij) is closest to the first image feature knowledge kga.
  • Step S721 the video data type fuzzy discriminating module 4 respectively obtains the total distance of the pending feature description and the first sample feature knowledge, the total distance of the pending feature description and the second sample feature knowledge;
  • Step S722 the video data type fuzzy discriminating module 4 compares the total distance da between the pending feature description and the first sample feature knowledge ka and the total distance db of the pending feature description and the second sample feature knowledge kb, if da ⁇ db, Go to step S723, otherwise go to step S727.
  • Step S723 determining whether the pending character feature description fw(vij) falls within the first character feature knowledge value range (kwa_min, kwa_max), and the pending voice feature description fv(vij) falls within the first sound feature knowledge value range ( Kva_min,kva_max), And the pending image feature description fg(vij) falls within the first image feature knowledge value range (kga_min, kga_max); if yes, step S724 is performed, otherwise step S725 is performed.
  • Step S724 The pending video frame vij belongs to the specific category of category a specific review corresponding to the first sample feature knowledge ka, such as a shooting video, and the process ends.
  • Step S725 determining whether the pending character feature description fw(vij) falls within the second character feature knowledge value range (kwb_min, kwb_max), and the pending voice feature description fv(vij) falls within the second sound feature knowledge value range. (kvb_min, kvb_max), And the pending image feature description fg(vij) falls within the second image feature knowledge value range (kgb_min, kgb_max); if yes, step S726 is performed, otherwise step S622 is performed.
  • Step S726 Determine that the pending video frame vij belongs to the category b specific review category corresponding to the second sample feature knowledge kb, such as a robbery video, and the process ends.
  • Step S727 determining whether the pending character feature description fw(vij) falls within the second character feature knowledge value range (kwb_min, kwb_max), and the pending voice feature description fv(vij) falls within the second sound feature knowledge value range. (kvb_min, kvb_max), And the pending image feature description fg(vij) falls within the second image feature knowledge value range (kgb_min, kgb_max); if yes, step S726 is performed, otherwise step S729 is performed.
  • Step S729 determining whether the pending character feature description fw(vij) falls within the first character feature knowledge value range (kwa_min, kwa_max), and the pending sound feature description fv(vij) falls within the first sound feature knowledge value range ( Kva_min,kva_max), And the pending image feature description fg(vij) falls within the first image feature knowledge value range (kga_min, kga_max); if yes, step S724 is performed, otherwise step S622 is performed.
  • Step S622 determining whether the search for all the sample feature knowledge in the video-awareness knowledge base 3 has been completed, if otherwise, proceeding to step S623, if yes, the process ends.
  • Step S623 the perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for the first sample feature knowledge that the corresponding sub-item is close to a sub-item of the pending feature description, and the corresponding sub-items are close to other sub-items.
  • the second sample feature knowledge is then proceeded to step S721.
  • the solution for determining the specific review category to which the pending video frame vij belongs includes:
  • Step S631 the perceptual feature fuzzy matching module 3 finds the first sample feature knowledge ka, the second sample feature knowledge kb, and the third sample feature knowledge kc, each of which has a sub-item that is closest to a sub-item of the pending feature description. , proceeding to step S731;
  • the pending character feature description fw(vij) is closest to the first character feature knowledge kwa
  • the pending sound feature description fv(vij) is closest to the second sound feature knowledge kvb
  • the pending image feature description Fg(vij) is closest to the third image feature knowledge kgc.
  • Step S731 The video data type fuzzy discriminating module 4 respectively obtains a total distance between the pending feature description and each sample feature knowledge
  • the distance dga from the first image feature knowledge kga is obtained as the total distance da of the pending feature description and the first sample feature knowledge ka.
  • the total distance db of the kb of the pending feature description and the second sample feature knowledge and the total distance dc of the pending feature description and the kc of the third sample feature knowledge are respectively obtained.
  • Step S732 The video data type fuzzy discriminating module 4 compares the total distance between the pending feature description and each sample feature knowledge. If da ⁇ db ⁇ dc, step S733 is performed.
  • Step S733 determining whether the pending character feature description fw(vij) falls within the first character feature knowledge value range (kwa_min, kwa_max), and the pending voice feature description fv(vij) falls within the first sound feature knowledge value range ( Kva_min, kva_max) range, And the pending image feature description fg(vij) falls within the first image feature knowledge value range (kga_min, kga_max); if yes, step S734 is performed, otherwise step S735 is performed.
  • step S734 the pending video frame vij belongs to the category a specific review category corresponding to the first sample feature knowledge ka, such as a shooting video, and the process ends.
  • Step S735 determining whether the pending character feature description fw(vij) falls within the second character feature knowledge value range (kwb_min, kwb_max), and the pending sound feature description fv(vij) falls within the second sound feature knowledge value range. (kvb_min, kvb_max) range, And the pending image feature description fg(vij) falls within the second image feature knowledge value range (kgb_min, kgb_max); if yes, step S736 is performed, otherwise step S737 is performed.
  • step S736 it is determined that the pending video frame vij belongs to the category b specific review category corresponding to the second sample feature knowledge kb, such as a robbery video, and the process ends.
  • Step S737 determining whether the pending character feature description fw(vij) falls within the third character feature knowledge value range (kwc_min, kwc_max), and the pending voice feature description fv(vij) falls within the third sound feature knowledge value range. (kvc_min, kvc_max) range, And the pending image feature description fg(vij) falls within the third image feature knowledge value range (kgc_min, kgc_max); if yes, step S738 is performed, otherwise step S632 is performed.
  • Step S738 Determine that the pending video frame vij belongs to a category c specific review category corresponding to the third sample feature knowledge kc, such as a bare hip video.
  • Step S632 determining whether the search for all the sample feature knowledge in the video-awareness knowledge base 3 has been completed, if otherwise, proceeding to step S633, if yes, the process ends.
  • Step S633 The perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for the first sample feature knowledge ka, the second sample feature knowledge kb, and the third corresponding to each sub-item of the pending feature description. The sample feature knowledge kc is then proceeded to step S731.
  • the comparison order is adjusted according to the order of the sample feature knowledge with the smallest total distance to the maximum total distance, and the steps similar to steps S733 to S738 are used to compare sequentially.
  • the sub-items of the pending feature description fall into the range of values of the corresponding sub-items of the feature knowledge, thereby determining which specific type of feature knowledge belongs to the pending feature description, and the specific process is not described again. If the three sample feature knowledge is compared, the sub-items of the pending feature description do not fall within the value range of the same sample feature knowledge, and then step S632 is performed.
  • the data processing such as feature extraction, summary, distance finding, total distance, and range of value obtained by the present invention, is not limited to the exemplary solutions given in the above embodiments, and may be implemented by other schemes.
  • the present invention also provides a video content review method based on a perceptual knowledge base, including:
  • Step S1 The video sample database S includes at least one video sample sub-library Si, and pre-stores and manages sample videos corresponding to at least one type of specific review categories, respectively;
  • Step S2 The sample data perceptual feature extraction module 1 performs perceptual feature extraction on the sample video pre-stored in the video sample sub-library Si according to at least one feature of each frame, and obtains at least a sample feature description corresponding to each frame respectively.
  • Step S3 The sample data perceptual feature extraction module 2 weights and averages at least one sub-item of the sample feature description to obtain at least one sub-item sample feature knowledge; respectively, the minimum and maximum values of each sub-sample feature knowledge are taken as sample features. The range of values of knowledge;
  • Step S4 The video-aware knowledge base K stores the component sample feature knowledge corresponding to the specific review category and the value range thereof in the video-aware knowledge sub-library Ki corresponding to at least one type of specific review category;
  • step S5 the data-aware feature extraction module 2 performs a perceptual feature extraction on at least one feature of each frame of the video to be examined, and generates at least one item corresponding to each feature of the pending feature description corresponding to each frame;
  • Step S6 The perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for the sample feature knowledge that is closest to the at least one sub-item of the pending feature description.
  • Step S7 The video data type fuzzy discriminating module 4 analyzes at least the same characteristic knowledge found, and determines a specific review category to which the corresponding pending video frame belongs;
  • Step S8 The user video data type library U stores and manages the pending video frame and the specific review category to which it belongs;
  • Step S9 the review result comprehensive query module 5 queries the pending video frame in the user video data type library U and the specific review category to which it belongs according to the user instruction; and/or according to the specific type of video alarm setting specified by the user, when the video is pending A frame containing a specific type of video alert setting is sent to alert the user.
  • step S6 and step S7 specifically include:
  • Step S701 the video data type fuzzy discriminating module 4 analyzes that each sub-item of the pending feature description is closest to the corresponding sub-items of the sample feature knowledge in the video-aware knowledge base K;
  • Step S702 The video data type fuzzy discriminating module 4 compares each sub-item of the pending feature description with the corresponding sub-item of the same feature knowledge, and compares whether each sub-item of the pending feature description falls into the corresponding score of the sample feature knowledge. The range of values of the item, thereby determining that the pending feature description belongs to a specific type corresponding to the sample feature knowledge;
  • Step S703 The video data type fuzzy discriminating module 4 obtains the total distance between the pending feature description and each sample feature knowledge, respectively, that the sub-items of the pending feature description are respectively closest to the corresponding sub-items of the at least two sample feature knowledge;
  • Step S704 the video data type fuzzy discriminating module 4 compares the total distance between the pending feature description and each sample feature knowledge, from the sample feature knowledge with the smallest total distance to the sample feature knowledge with the largest total distance, and the video data type fuzzy discriminating module 4 respectively compares The sub-items of the pending feature description fall into the range of values of the corresponding sub-items of the feature knowledge, thereby determining which specific feature type the corresponding feature knowledge belongs to; If the item does not fall within the range of values of the corresponding item of the sample feature knowledge found, then
  • Step S601 The perceptual feature fuzzy matching module 3 searches the video perceptual knowledge base K for at least the same feature knowledge that each sub-item of the pending feature description is close to the corresponding sub-item, and repeatedly obtains the total distance and compares the judgment until the video perception is completed. Search for all sample feature knowledge in knowledge base K.

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Library & Information Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Image Analysis (AREA)

Abstract

本发明提出一种基于感知知识库的视频内容审査系统，包括视频样本数据库、样本数据感知特征抽取模块、视频感知知识库、感知特征模糊匹配模块、视频数据类型模糊判别模块、用户视频数据类型库和审査结果综合査询模块，根据用户指令査询用户视频数据类型库中的待审视频帧及其所属的特定审査类别；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警。本发明还提出一种基于感知知识库的视频内容审査方法。本发明对样本视频数据自动抽取感知知识并釆用数据库管理，利用感知知识自动地审査视频内容，达到了全自动审査的目的，无需人工操作、无需编写特定的审査算法，通用性强。

Description

一种基于感知知识库的视频内容审查系统及方法
【技术领域】
本发明涉及多媒体处理技术，更具体地说，涉及一种基于感知知识库的视频内容审查系统及方法。
【背景技术】
目前视频内容审查基本上都是采用人工观看的方式，一个视频节目需要用户人工从头到尾看一遍，有时还需要多个用户分别多次观看，耗费大量时间和人力，效率较低。现有技术也提出一些算法，用于识别特定的视频内容，如特定类型的裸体。但现有技术的算法通用性差，对于稍有变化或不符合预先定义类型的视频内容就无法识别，实用性较低。
【发明内容】
基于此，本发明提出一种基于感知知识库的视频内容审查系统，包括：视频样本数据库，包括至少一视频样本子库，分别存储并管理与至少一类特定审查类别对应的样本视频；样本数据感知特征抽取模块，对视频样本子库中的样本视频逐帧进行感知特征抽取，产生与其每一帧对应的样本特征描述；对样本特征描述进行汇总得到样本特征知识，定义样本特征知识的取值范围；视频感知知识库，存储并管理样本特征知识及其取值范围；待审数据感知特征抽取模块，对待审视频逐帧进行感知特征抽取，产生与其每一帧对应的待审特征描述；感知特征模糊匹配模块，分别从视频感知知识库中搜索与各待审特征描述最接近的样本特征知识；视频数据类型模糊判别模块，分析找到的样本特征知识，确定待审特征描述对应待审视频帧所属的特定审查类别；用户视频数据类型库，存储并管理待审视频帧及其所属的特定审查类别和审查结果综合查询模块，根据用户指令查询用户视频数据类型库中的待审视频帧及其所属的特定审查类别；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警。
其中，上述样本数据感知特征抽取模块分别对样本视频每一帧的至少一特征进行感知特征抽取，获得与每一帧分别对应的样本特征描述的至少一分项。
其中，上述样本数据感知特征抽取模块对样本特征描述进行汇总得到样本特征知识，是对样本特征描述的至少一分项分别加权平均，得到至少一分项样本特征知识；样本数据感知特征抽取模块分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围。
其中，视频感知知识库包括与至少一类特定审查类别对应的视频感知知识子库，分别存储与特定审查类别对应的分项样本特征知识及其取值范围；
其中，上述待审数据感知特征抽取模块分别对待审视频每一帧的至少一特征进行感知特征抽取，产生与其每一帧分别对应的待审特征描述的至少一分项。
对于待审特征描述的所有分项均与视频感知知识库中某一样本特征知识相应分项最接近的，视频数据类型模糊判别模块分别比较待审特征描述的各分项是否均落入样本特征知识的相应分项的取值范围，是则确定待审特征描述属于样本特征知识对应的特定类型,否则感知特征模糊匹配模块从视频感知知识库中搜索所有分项均与待审特征描述的相应分项次接近的样本特征知识；视频数据类型模糊判别模块分别比较待审特征描述的各分项是否均落入次接近的样本特征知识的相应分项的取值范围，是则确定待审特征描述属于次接近的样本特征知识对应的特定类型,否则感知特征模糊匹配模块继续查找次接近的样本特征知识，直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
对于待审特征描述某一分项与视频感知知识库中第一样本特征知识的对应分项最接近，且其他分项均与第二样本特征知识的相应分项最接近的，视频数据类型模糊判别模块分别求出待审特征描述与第一样本特征知识的总距离、待审特征描述与第二样本特征知识的总距离；视频数据类型模糊判别模块比较待审特征描述与第一样本特征知识的总距离和待审特征描述与第二样本特征知识的总距离，从总距离较小的样本特征知识到总距离较大的样本特征知识，视频数据类型模糊判别模块分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；如果待审特征描述的各分项既未均落入总距离较小的样本特征知识的相应分项的取值范围，也未均落入总距离较大的样本特征知识相应分项的取值范围，则感知特征模糊匹配模块从视频感知知识库中搜索待审特征描述某一分项与对应分项次接近的第一样本特征知识，且其他分项均与相应分项次接近的第二样本特征知识，重复求出总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
对于待审特征描述各分项分别与视频感知知识库中至少一样本特征知识的对应一分项最接近的，视频数据类型模糊判别模块分别求出待审特征描述与各样本特征知识的总距离；视频数据类型模糊判别模块比较待审特征描述与各样本特征知识的总距离，从总距离最小的样本特征知识到总距离最大的样本特征知识，视频数据类型模糊判别模块分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；如果待审特征描述的各分项未均落入任何一样本特征知识相应分项的取值范围，，则感知特征模糊匹配模块从视频感知知识库中搜索待审特征描述某一分项与对应分项次接近的至少一样本特征知识，重复求出总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
本发明还提出一种基于感知知识库的视频内容审查方法，包括：视频样本数据库包括至少一视频样本子库，分别预先存储并管理与至少一类特定审查类别对应的样本视频的步骤；样本数据感知特征抽取模块对视频样本子库中预先存储的样本视频，分别根据每一帧的至少一特征进行感知特征抽取，获得与每一帧分别对应的样本特征描述的至少一分项的步骤；样本数据感知特征抽取模块对样本特征描述的至少一分项分别加权平均，得到至少一分项样本特征知识；分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围的步骤；视频感知知识库将与特定审查类别对应的分项样本特征知识及其取值范围存入与至少一类特定审查类别对应的视频感知知识子库中的步骤；待审数据感知特征抽取模块分别对待审视频每一帧的至少一特征进行感知特征抽取，产生与其每一帧分别对应的待审特征描述的至少一分项的步骤；感知特征模糊匹配模块分别从视频感知知识库中搜索相应分项与待审特征描述的至少一分项最接近的样本特征知识的步骤；视频数据类型模糊判别模块分析找到的至少一样本特征知识，确定与其对应的待审视频帧所属的特定审查类别的步骤；用户视频数据类型库存储并管理待审视频帧及其所属的特定审查类别的步骤；和审查结果综合查询模块根据用户指令查询用户视频数据类型库中的待审视频帧及其所属的特定审查类别的步骤；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警的步骤。
其中，上述视频数据类型模糊判别模块分析找到的至少一样本特征知识，确定与其对应的待审视频帧所属的特定审查类别的步骤包括：视频数据类型模糊判别模块分析待审特征描述各分项分别与视频感知知识库中哪些样本特征知识的对应分项最接近的步骤；对于待审特征描述各分项均与一样本特征知识的对应分项最接近的，视频数据类型模糊判别模块比较待审特征描述的各分项是否落入样本特征知识的相应分项的取值范围，由此确定待审特征描述属于样本特征知识对应的特定类型的步骤；对于待审特征描述各分项分别与至少二样本特征知识的对应分项最接近的，视频数据类型模糊判别模块分别求出待审特征描述与各样本特征知识的总距离的步骤；视频数据类型模糊判别模块比较待审特征描述与各样本特征知识的总距离，从总距离最小的样本特征知识到总距离最大的样本特征知识，视频数据类型模糊判别模块分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型的步骤；如果待审特征描述的各分项未能均落入找到的样本特征知识相应分项的取值范围，则感知特征模糊匹配模块从视频感知知识库中搜索待审特征描述各分项与对应分项次接近的至少一样本特征知识，重复求出总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找的步骤。
本发明对样本视频数据自动抽取感知知识并采用数据库管理，利用感知知识自动地审查视频内容，达到了全自动审查的目的，无需人工操作、无需编写特定的审查算法，通用性强。
【附图说明】
图1为本发明一实施例的基于感知知识库的视频内容审查系统系统结构示意图；
图2为本发明一实施例的基于感知知识库的视频内容审查方法流程示意图；
图3为本发明一实施例的确定待审视频帧所属的特定审查类别部分流程示意图。
【具体实施方式】
为了使本发明的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本发明进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本发明，并不用于限定本发明。
本发明第一实施例参照图1示出的基于感知知识库的视频内容审查系统系统结构示意图和图2示出的视频内容审查主要流程示意图，本发明提出基于感知知识库的视频内容审查系统一实施例包括：
视频样本数据库S，包括至少一视频样本子库Si，分别存储并管理与至少一类特定审查类别对应的样本视频；
样本数据感知特征抽取模块1，对视频样本子库Si中的样本视频逐帧进行感知特征抽取，产生与其每一帧对应的样本特征描述；对样本特征描述进行汇总得到样本特征知识，定义样本特征知识的取值范围；
视频感知知识库K，存储并管理样本特征知识及其取值范围；
待审数据感知特征抽取模块2，对待审视频逐帧进行感知特征抽取，产生与其每一帧对应的待审特征描述；
感知特征模糊匹配模块3，分别从视频感知知识库K中搜索与各待审特征描述最接近的样本特征知识；
视频数据类型模糊判别模块4，分析找到的样本特征知识，确定待审特征描述对应待审视频帧所属的特定审查类别；
用户视频数据类型库U，存储并管理待审视频帧及其所属的特定审查类别；
审查结果综合查询模块5，根据用户指令查询用户视频数据类型库U中的待审视频帧及其所属的特定审查类别；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警。
基于上述实施例，本发明提出一实施例，样本数据感知特征抽取模块1分别对样本视频每一帧的至少一特征进行感知特征抽取，获得与每一帧分别对应的样本特征描述的至少一分项。视频包括文字、图像和/或声音等特征，则对应获得样本特征描述的分项为：文字特征描述、图像特征描述和/或声音特征描述。
本实施例的样本数据感知特征抽取模块1对样本特征描述进行汇总得到样本特征知识，是对样本特征描述的至少一分项分别加权平均，得到至少一分项样本特征知识；样本数据感知特征抽取模块1分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围。
视频感知知识库K包括与至少一类特定审查类别对应的视频感知知识子库Ki，分别存储与特定审查类别对应的分项样本特征知识及其取值范围；
待审数据感知特征抽取模块2分别对待审视频每一帧的至少一特征进行感知特征抽取，产生与其每一帧分别对应的待审特征描述的至少一分项。
基于上述实施例，本发明提出一实施例，其中感知特征模糊匹配模块3和视频数据类型模糊判别模块4配合，共同确定待审特征描述对应待审视频帧所属的特定审查类别。
本实施例将感知特征模糊匹配模块3找到的样本特征知识分为以下三种情况：
第一种情况，对于待审特征描述的所有分项均与视频感知知识库K中某一样本特征知识相应分项最接近的，视频数据类型模糊判别模块4分别比较待审特征描述的各分项是否均落入样本特征知识的相应分项的取值范围，是则确定待审特征描述属于样本特征知识对应的特定类型,否则
感知特征模糊匹配模块3从视频感知知识库K中搜索所有分项均与待审特征描述的相应分项次接近的样本特征知识；
视频数据类型模糊判别模块4分别比较待审特征描述的各分项是否均落入次接近的样本特征知识的相应分项的取值范围，是则确定待审特征描述属于次接近的样本特征知识对应的特定类型,否则
感知特征模糊匹配模块3继续查找次接近的样本特征知识，直到确定待审特征描述所属的特定类型，或完成对视频感知知识库K中所有样本特征知识的查找。
第二种情况，对于待审特征描述某一分项与视频感知知识库K中第一样本特征知识的对应分项最接近，且其他分项均与第二样本特征知识的相应分项最接近的，视频数据类型模糊判别模块4分别求出待审特征描述与第一样本特征知识的总距离、待审特征描述与第二样本特征知识的总距离。
视频数据类型模糊判别模块4比较待审特征描述与第一样本特征知识的总距离和待审特征描述与第二样本特征知识的总距离，从总距离较小的样本特征知识到总距离较大的样本特征知识，视频数据类型模糊判别模块4分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型。
如果待审特征描述的各分项既未均落入总距离较小的样本特征知识的相应分项的取值范围，也未均落入总距离较大的样本特征知识相应分项的取值范围，则
感知特征模糊匹配模块3从视频感知知识库K中搜索待审特征描述某一分项与对应一分项次接近的第一样本特征知识，且其他分项均与相应分项次接近的第二样本特征知识，重复求出总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库K中所有样本特征知识的查找。
第三种情况，对于待审特征描述各分项分别与视频感知知识库K中至少一样本特征知识的对应分项最接近的，视频数据类型模糊判别模块4分别求出待审特征描述与各样本特征知识的总距离。
视频数据类型模糊判别模块4比较待审特征描述与各样本特征知识的总距离，从总距离最小的样本特征知识到总距离最大的样本特征知识，视频数据类型模糊判别模块4分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；
如果待审特征描述的各分项未均落入任何一样本特征知识相应分项的取值范围，则
感知特征模糊匹配模块3从视频感知知识库K中搜索待审特征描述某一分项与对应分项次接近的至少一样本特征知识，重复求出总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库K中所有样本特征知识的查找。
以下给出本发明一具体实施例说明本发明方案。
设视频样本数据库S包括至少一视频样本子库Si，分别对应i类类特定审查类别。如第一视频样本子库S1对应1类枪击视频；第二视频样本子库S2对应2类性交视频；第三视频样本子库S3对应3类裸乳视频；第四视频样本子库S4对应4类裸臀视频；第五视频样本子库S5对应5类打砸视频等。
设第i类视频字库Si中包含样本视频的m个视频帧，Si={si1,si2,…,sim},样本数据感知特征抽取模块1分别对样本视频每一帧的文字、图像和声音特征进行感知特征抽取，获得与每一帧分别对应的文字特征描述、图像特征描述和声音特征描述。样本数据感知特征抽取模块1运行文字特征抽取函数fw得到样本文字特征描述fw（si1）、fw（si2）、…、 fw（sim）；运行声音特征抽取函数fv 得到样本声音特征描述fv（si1）、fv（si2）、…、fv（sim）；运行图像特征抽取函数fg得到样本图像特征描述fg（si1）、fg（si2）、…、fg（sim））。
样本数据感知特征抽取模块1对样本特征描述的至少一分项进行汇总，本实施例是采用加权平均，得到至少一分项样本特征知识，包括：文字样本特征知识kwi=（fw（si1）+fw（si2）+…+fw（sim））/m；声音样本特征知识kvi=（fv（si1）+fv（si2）+…+fv（sim））/m；图像样本特征知识kgi=（fg（si1）+fg（si2）+…+fg（sim））/m。
样本数据感知特征抽取模块1分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围。文字样本特征知识最小值kwi_min=min（fw（si1）+fw（si2）+…+fw（sim）），文字样本特征知识最大值kwi_max=max（fw（si1）+fw（si2）+…+fw（sim））；声音样本特征知识最小值kvi_min=min（fv（si1）+fv（si2）+…+fv（sim）），声音样本特征知识最大值kvi_max=max（fv（si1）+fv（si2）+…+fv（sim））；图像样本特征知识最小值kgi_min=min（fg（si1）+fg（si2）+…+fg（sim）），图像样本特征知识最大值kgi_max=max（fg（si1）+fg（si2）+…+fg（sim））） 。
与视频样本库S中的视频样本子库Si一一对应地，视频感知知识库K包括多个视频感知知识子库Ki，分别存储样本特征知识及其取值范围：文字特征知识kwi及其取值范围(kwi_min,kwi_max)、声音特征知识kvi及其取值范围(kvi_min,kvi_max)、图像特征知识kgi及其取值范围(kgi_min,kgi_max)。
设用户提交的待审视频Vi={vi1,vi2,…,vin}，包含待审视频的n个视频帧。待审数据感知特征抽取模块2对其每一帧的文字、图像和声音特征进行感知特征抽取，运行文字特征抽取函数fw得到待审文字特征描述fw（vij）={fw（vi1），fw（vi2），…,fw（vin）}、运行声音特征抽取fv得到待审声音特征描述fv（vij）={fv（vi1），fv（vi2），…,fv（vin）}；运行图像特征抽取fg得到得到待审图像特征描述fg（vij）={ fg（vi1），fg（vi2），…,fg（vin）}。
参照图3示出的确定待审视频帧所属的特定审查类别部分流程示意图，其中图3（a）示出对于第一种情况，确定待审视频帧vij所属的特定审查类别的方案，包括：
步骤S611、感知特征模糊匹配模块3找到某一样本特征知识，待审特征描述的所有分项均与其相应分项最接近，进行步骤S711；
即感知特征模糊匹配模块3从视频感知知识库K中找到第一样本特征知识ka，包括第一文字特征知识kwa、第一声音特征知识kva和第一图像特征知识kga。待审文字特征描述fw（vij）与其第一文字特征知识kwa最接近，且待审声音特征描述fv（vij）与其第一声音特征知识kva最接近且待审图像特征描述fg（vij）与其第一图像特征知识kga最接近。
步骤S711、视频数据类型模糊判别模块4分别比较待审特征描述的各分项是否均落入样本特征知识的相应分项的取值范围，如果是则进行步骤S712，否则进行步骤S612；
即判断是否待审文字特征描述fw（vij）落入第一文字特征知识取值范围（kwa_min,kwa_max）,且待审声音特征描述fv（vij）落入第一声音特征知识取值范围（kva_min,kva_max）范围, 且待审图像特征描述fg（vij）落入第一图像特征知识取值范围（kga_min,kga_max）。
步骤S712、确定待审视频帧vij属于与第一样本特征知识ka对应的第a类特定审查类别，如枪击类视频，本流程结束。
步骤S612、判断是否已经完成对视频感知知识库3中所有样本特征知识的查找，如果否则进行步骤S613，如果是则本流程结束。
步骤S613、感知特征模糊匹配模块3从视频感知知识库K中搜索所有分项均与待审特征描述的相应分项次接近的样本特征知识，进行步骤711；
即从视频感知知识库K中取与待审视频帧vij次接近的第二样本特征知识kb。
本发明又提出一实施例，参照图3（b）所示的对于第二种情况，确定待审视频帧vij所属的特定审查类别的方案包括：
步骤S621、感知特征模糊匹配模块3找到第一样本特征知识ka和第二样本特征知识kb，待审文字特征描述fw（vij）的某一分项与第一样本特征知识ka的对应分项最接近，待审文字特征描述fw（vij）的其他二分项均与第二样本特征知识kn的相应二分项最接近，则进行步骤721；
例如待审文字特征描述fw（vij）与第一文字特征知识kwa最接近，待审声音特征描述 fv（vij）与第二声音特征知识kvb最接近，且待审图像特征描述fg（vij）与第二图像特征知识kgb最接近，或
如果待审文字特征描述fw（vij）与第一文字特征知识kwa最接近，待审声音特征描述 fv（vij）与第二声音特征知识kvb最接近，且待审图像特征描述 fg（vij）与第一图像特征知识kga最接近，或
如果待审文字特征描述fw（vij）与第二文字特征知识kwb最接近，待审声音特征描述 fv（vij）与第一声音特征知识kva最接近，且待审图像特征描述 fg（vij）与第一图像特征知识kga最接近，或
如果待审文字特征描述fw（vij）与第二文字特征知识kwb最接近，待审声音特征描述 fv（vij）与第一声音特征知识kva最接近，且待审图像特征描述 fg（vij）与第二图像特征知识kgb最接近，或
如果待审文字特征描述fw（vij）与第二文字特征知识kwb最接近，待审声音特征描述 fv（vij）与第二声音特征知识kvb最接近，且待审图像特征描述fg（vij）与第一图像特征知识kga最接近……等多种情况。
步骤S721、视频数据类型模糊判别模块4分别求出待审特征描述与第一样本特征知识的总距离、待审特征描述与第二样本特征知识的总距离；
例如分别求出待审文字特征描述fw（vij）与第一文字特征知识kwa的距离dwa、待审声音特征描述fv（vij）与第一声音特征知识kva的距离dva，待审图像特征描述fg（vij） 与第一图像特征知识kga的距离dga，综合三距离得到待审特征描述与第一样本特征知识ka的总距离da =(（dwa）^2 +（dva）^2 +（dga）^2 )^(1/2)））。同样地求出待审特征描述与第二样本特征知识的kb的总距离db=(（dwb）^2 +（dvb）^2 +（dgb）^2 )^(1/2)））。
步骤S722，视频数据类型模糊判别模块4比较待审特征描述与第一样本特征知识ka的总距离da和待审特征描述与第二样本特征知识kb的总距离db，如果da<db,则进行步骤S723，否则进行步骤S727。
步骤S723、判断是否待审文字特征描述fw（vij）落入第一文字特征知识取值范围（kwa_min,kwa_max）,且待审声音特征描述fv（vij）落入第一声音特征知识取值范围（kva_min,kva_max）, 且待审图像特征描述fg（vij）落入第一图像特征知识取值范围（kga_min,kga_max）；如果是则进行步骤S724，否则进行步骤S725。
步骤S724、待审视频帧vij属于与第一样本特征知识ka对应的第a类特定审查类别，如枪击类视频，本流程结束。
步骤S725、判断是否待审文字特征描述fw（vij）落入第二文字特征知识取值范围（kwb_min,kwb_max）,且待审声音特征描述fv（vij）落入第二声音特征知识取值范围（kvb_min,kvb_max）, 且待审图像特征描述fg（vij）落入第二图像特征知识取值范围（kgb_min,kgb_max）；如果是则进行步骤S726，否则进行步骤S622。
步骤S726、确定待审视频帧vij属于与第二样本特征知识kb对应的第b类特定审查类别，如抢砸类视频，本流程结束。
步骤S727、判断是否待审文字特征描述fw（vij）落入第二文字特征知识取值范围（kwb_min,kwb_max）,且待审声音特征描述fv（vij）落入第二声音特征知识取值范围（kvb_min,kvb_max）, 且待审图像特征描述fg（vij）落入第二图像特征知识取值范围（kgb_min,kgb_max）；如果是则进行步骤S726，否则进行步骤S729。
步骤S729、判断是否待审文字特征描述fw（vij）落入第一文字特征知识取值范围（kwa_min,kwa_max）,且待审声音特征描述fv（vij）落入第一声音特征知识取值范围（kva_min,kva_max）, 且待审图像特征描述fg（vij）落入第一图像特征知识取值范围（kga_min,kga_max）；如果是则进行步骤S724，否则进行步骤S622。
步骤S622、判断是否已经完成对视频感知知识库3中所有样本特征知识的查找，如果否则进行步骤S623，如果是则本流程结束。
步骤S623、感知特征模糊匹配模块3从视频感知知识库K中搜索对应分项与待审特征描述某一分项次接近的第一样本特征知识，及相应分项均与其他分项次接近的第二样本特征知识，然后进行步骤S721。
本发明又提出一实施例，参照图3（c）所示的对于第三种情况，确定待审视频帧vij所属的特定审查类别的方案包括：
步骤S631、感知特征模糊匹配模块3找到第一样本特征知识ka、第二样本特征知识kb和第三样本特征知识kc，三者分别有一分项是与待审特征描述某一分项最接近，则进行步骤S731；
即如果待审文字特征描述fw（vij）与第一文字特征知识kwa最接近，待审声音特征描述fv（vij）与第二声音特征知识kvb最接近，而待审图像特征描述 fg（vij）与第三图像特征知识kgc最接近。
步骤S731、视频数据类型模糊判别模块4分别求出待审特征描述与各样本特征知识的总距离；
即分别求出待审文字特征描述fw（vij）与第一文字特征知识kwa的距离dwa，待审声音特征描述fv（vij） 与第一声音特征知识kva的距离dva，待审图像特征描述fg（vij） 与第一图像特征知识kga的距离dga，得到待审特征描述与第一样本特征知识ka的总距离da。同样地分别求出待审特征描述与第二样本特征知识的kb的总距离db和待审特征描述与第三样本特征知识的kc的总距离dc。
步骤S732、视频数据类型模糊判别模块4比较待审特征描述与各样本特征知识的总距离，如果da<db<dc,则进行步骤S733。
步骤S733、判断是否待审文字特征描述fw（vij）落入第一文字特征知识取值范围（kwa_min,kwa_max）,且待审声音特征描述fv（vij）落入第一声音特征知识取值范围（kva_min,kva_max）范围, 且待审图像特征描述fg（vij）落入第一图像特征知识取值范围（kga_min,kga_max）；如果是则进行步骤S734，否则进行步骤S735。
步骤S734、待审视频帧vij属于与第一样本特征知识ka对应的第a类特定审查类别，如枪击类视频，则本流程结束。
步骤S735、判断是否待审文字特征描述fw（vij）落入第二文字特征知识取值范围（kwb_min,kwb_max）,且待审声音特征描述fv（vij）落入第二声音特征知识取值范围（kvb_min,kvb_max）范围, 且待审图像特征描述fg（vij）落入第二图像特征知识取值范围（kgb_min,kgb_max）；如果是则进行步骤S736，否则进行步骤S737。
步骤S736、确定待审视频帧vij属于与第二样本特征知识kb对应的第b类特定审查类别，如抢砸类视频，则本流程结束。
步骤S737、判断是否待审文字特征描述fw（vij）落入第三文字特征知识取值范围（kwc_min,kwc_max）,且待审声音特征描述fv（vij）落入第三声音特征知识取值范围（kvc_min,kvc_max）范围, 且待审图像特征描述fg（vij）落入第三图像特征知识取值范围（kgc_min,kgc_max）；如果是则进行步骤S738，否则进行步骤S632。
步骤S738、确定待审视频帧vij属于与第三样本特征知识kc对应的第c类特定审查类别，如裸臀类视频。
步骤S632、判断是否已经完成对视频感知知识库3中所有样本特征知识的查找，如果否则进行步骤S633，如果是则本流程结束。
步骤S633、感知特征模糊匹配模块3从视频感知知识库K中搜索对应一分项分别与待审特征描述各分项次接近的第一样本特征知识ka、第二样本特征知识kb和第三样本特征知识kc，然后进行步骤S731。
对于db<da<dc、dc<db<da等多种情况，按照从总距离最小的样本特征知识到总距离最大的顺序，调整比较顺序，采用与步骤S733至步骤S738类似的步骤，依次比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型，具体流程不再赘述。若比较完三个样本特征知识，待审特征描述的各分项没有均落入同一样本特征知识的取值范围，则进行步骤S632。
本发明采用的特征抽取、汇总、求距离、求总距离、求取值范围等数据处理均不限于上述各实施例中给出的示例方案，还可采用其他方案实现。
参照图2示出的总流程示意图，本发明还提出一种基于感知知识库的视频内容审查方法，包括：
步骤S1、视频样本数据库S包括至少一视频样本子库Si，分别预先存储并管理与至少一类特定审查类别对应的样本视频；
步骤S2、样本数据感知特征抽取模块1对视频样本子库Si中预先存储的样本视频，分别根据每一帧的至少一特征进行感知特征抽取，获得与每一帧分别对应的样本特征描述的至少一分项；
步骤S3、样本数据感知特征抽取模块2对样本特征描述的至少一分项分别加权平均，得到至少一分项样本特征知识；分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围；
步骤S4、视频感知知识库K将与特定审查类别对应的分项样本特征知识及其取值范围存入与至少一类特定审查类别对应的视频感知知识子库Ki中；
步骤S5、待审数据感知特征抽取模块2分别对待审视频每一帧的至少一特征进行感知特征抽取，产生与其每一帧分别对应的待审特征描述的至少一分项；
步骤S6、感知特征模糊匹配模块3分别从视频感知知识库K中搜索相应分项与待审特征描述的至少一分项最接近的样本特征知识；
步骤S7、视频数据类型模糊判别模块4分析找到的至少一样本特征知识，确定与其对应的待审视频帧所属的特定审查类别；
步骤S8、用户视频数据类型库U存储并管理待审视频帧及其所属的特定审查类别；
步骤S9、审查结果综合查询模块5根据用户指令查询用户视频数据类型库U中的待审视频帧及其所属的特定审查类别；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警。
本发明提出一实施例，基于上述实施例提出，步骤S6和步骤S7具体包括：
步骤S701、视频数据类型模糊判别模块4分析待审特征描述各分项分别与视频感知知识库K中哪些样本特征知识的对应分项最接近；
步骤S702、对于待审特征描述各分项均与一样本特征知识的对应分项最接近的，视频数据类型模糊判别模块4比较待审特征描述的各分项是否落入样本特征知识的相应分项的取值范围，由此确定待审特征描述属于样本特征知识对应的特定类型；
步骤S703、对于待审特征描述各分项分别与至少二样本特征知识的对应分项最接近的，视频数据类型模糊判别模块4分别求出待审特征描述与各样本特征知识的总距离；
步骤S704、视频数据类型模糊判别模块4比较待审特征描述与各样本特征知识的总距离，从总距离最小的样本特征知识到总距离最大的样本特征知识，视频数据类型模糊判别模块4分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；如果待审特征描述的各分项未能均落入找到的样本特征知识相应分项的取值范围，则
步骤S601、感知特征模糊匹配模块3从视频感知知识库K中搜索待审特征描述各分项与对应分项次接近的至少一样本特征知识，重复求出总距离并比较判断直到完成对视频感知知识库K中所有样本特征知识的查找。
以上实施例仅表达了本发明的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对本发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本发明构思的前提下，还可以做出若干变形和改进，这些都属于本发明的保护范围。因此，本发明专利的保护范围应以所附权利要求为准。

Claims (10)

  1. 一种基于感知知识库的视频内容审查系统，包括：
    视频样本数据库，包括至少一视频样本子库，分别存储并管理与至少一类特定审查类别对应的样本视频；
    样本数据感知特征抽取模块，对视频样本子库中的样本视频逐帧进行感知特征抽取，产生与其每一帧对应的样本特征描述；对样本特征描述进行汇总得到样本特征知识，定义样本特征知识的取值范围；
    视频感知知识库，存储并管理样本特征知识及其取值范围；
    待审数据感知特征抽取模块，对待审视频逐帧进行感知特征抽取，产生与其每一帧对应的待审特征描述；
    感知特征模糊匹配模块，分别从视频感知知识库中搜索与各待审特征描述最接近的样本特征知识；
    视频数据类型模糊判别模块，分析找到的样本特征知识，确定所述待审特征描述对应待审视频帧所属的特定审查类别；
    用户视频数据类型库，存储并管理待审视频帧及其所属的特定审查类别；
    审查结果综合查询模块，根据用户指令查询用户视频数据类型库中的待审视频帧及其所属的特定审查类别；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警。
  2. 根据权利要求1所述的基于感知知识库的视频内容审查系统，其特征在于：
    所述样本数据感知特征抽取模块分别对样本视频每一帧的至少一特征进行感知特征抽取，获得与每一帧分别对应的样本特征描述的至少一分项。
  3. 根据权利要求2所述的基于感知知识库的视频内容审查系统，其特征在于：
    所述样本数据感知特征抽取模块对样本特征描述进行汇总得到样本特征知识，是对样本特征描述的至少一分项分别加权平均，得到至少一分项样本特征知识；所述样本数据感知特征抽取模块分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围。
  4. 根据权利要求2所述的基于感知知识库的视频内容审查系统，其特征在于：
    视频感知知识库包括与至少一类特定审查类别对应的视频感知知识子库，分别存储与特定审查类别对应的分项样本特征知识及其取值范围。
  5. 根据权利要求2所述的基于感知知识库的视频内容审查系统，其特征在于：
    所述待审数据感知特征抽取模块分别对待审视频每一帧的至少一特征进行感知特征抽取，产生与其每一帧分别对应的待审特征描述的至少一分项。
  6. 根据权利要求2至5任意一项所述的基于感知知识库的视频内容审查系统，其特征在于：
    对于待审特征描述的所有分项均与视频感知知识库中某一样本特征知识相应分项最接近的，视频数据类型模糊判别模块分别比较待审特征描述的各分项是否均落入所述样本特征知识的相应分项的取值范围，是则确定待审特征描述属于所述样本特征知识对应的特定类型,否则
    感知特征模糊匹配模块从视频感知知识库中搜索所有分项均与所述待审特征描述的相应分项次接近的样本特征知识；
    视频数据类型模糊判别模块分别比较待审特征描述的各分项是否均落入所述次接近的样本特征知识的相应分项的取值范围，是则确定待审特征描述属于所述次接近的样本特征知识对应的特定类型,否则
    感知特征模糊匹配模块继续查找次接近的样本特征知识，直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
  7. 根据权利要求2至5任意一项所述的基于感知知识库的视频内容审查系统，其特征在于：
    对于待审特征描述某一分项与视频感知知识库中第一样本特征知识的对应分项最接近，且其他分项均与第二样本特征知识的相应分项最接近的，视频数据类型模糊判别模块分别求出待审特征描述与第一样本特征知识的总距离、待审特征描述与第二样本特征知识的总距离；
    视频数据类型模糊判别模块比较待审特征描述与第一样本特征知识的总距离和待审特征描述与第二样本特征知识的总距离，从总距离较小的样本特征知识到总距离较大的样本特征知识，视频数据类型模糊判别模块分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；
    如果待审特征描述的各分项既未均落入总距离较小的样本特征知识的相应分项的取值范围，也未均落入总距离较大的样本特征知识相应分项的取值范围，则
    感知特征模糊匹配模块从视频感知知识库中搜索待审特征描述某一分项与对应分项次接近的第一样本特征知识，且其他分项均与相应分项次接近的第二样本特征知识，重复求出所述总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
  8. 根据权利要求2至5任意一项所述的基于感知知识库的视频内容审查系统，其特征在于：
    对于待审特征描述各分项分别与视频感知知识库中至少一样本特征知识的对应一分项最接近的，视频数据类型模糊判别模块分别求出待审特征描述与各样本特征知识的总距离；
    视频数据类型模糊判别模块比较待审特征描述与各样本特征知识的总距离，从总距离最小的样本特征知识到总距离最大的样本特征知识，视频数据类型模糊判别模块分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；
    如果待审特征描述的各分项未均落入任何一样本特征知识相应分项的取值范围，则
    感知特征模糊匹配模块从视频感知知识库中搜索待审特征描述某一分项与对应分项次接近的至少一样本特征知识，重复求出所述总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
  9. 一种基于感知知识库的视频内容审查方法，包括：
    视频样本数据库包括至少一视频样本子库，分别预先存储并管理与至少一类特定审查类别对应的样本视频；
    样本数据感知特征抽取模块对视频样本子库中预先存储的样本视频，分别根据每一帧的至少一特征进行感知特征抽取，获得与每一帧分别对应的样本特征描述的至少一分项；
    样本数据感知特征抽取模块对样本特征描述的至少一分项分别加权平均，得到至少一分项样本特征知识；分别对各分项样本特征知识取最小值和最大值，作为样本特征知识的取值范围；
    视频感知知识库将与特定审查类别对应的分项样本特征知识及其取值范围存入与至少一类特定审查类别对应的视频感知知识子库中；
    待审数据感知特征抽取模块分别对待审视频每一帧的至少一特征进行感知特征抽取，产生与其每一帧分别对应的待审特征描述的至少一分项；
    感知特征模糊匹配模块分别从视频感知知识库中搜索相应分项与待审特征描述的至少一分项最接近的样本特征知识；
    视频数据类型模糊判别模块分析找到的至少一样本特征知识，确定与其对应的待审视频帧所属的特定审查类别；
    用户视频数据类型库存储并管理待审视频帧及其所属的特定审查类别；
    审查结果综合查询模块根据用户指令查询用户视频数据类型库中的待审视频帧及其所属的特定审查类别；和/或根据用户指定的特定类型视频告警设置，当待审视频中包含符合特定类型视频告警设置的帧，则向用户发出告警。
  10. 如权利要求9所述的一种基于感知知识库的视频内容审查方法，其特征在于，所述视频数据类型模糊判别模块分析找到的至少一样本特征知识，确定与其对应的待审视频帧所属的特定审查类别包括：
    视频数据类型模糊判别模块分析待审特征描述各分项分别与视频感知知识库中哪些样本特征知识的对应分项最接近；
    对于待审特征描述各分项均与一样本特征知识的对应分项最接近的，视频数据类型模糊判别模块比较待审特征描述的各分项是否均落入所述样本特征知识的相应分项的取值范围，由此确定待审特征描述属于所述样本特征知识对应的特定类型；
    对于待审特征描述各分项分别与至少二样本特征知识的对应分项最接近的，视频数据类型模糊判别模块分别求出待审特征描述与各样本特征知识的总距离；
    视频数据类型模糊判别模块比较待审特征描述与各样本特征知识的总距离，从总距离最小的样本特征知识到总距离最大的样本特征知识，视频数据类型模糊判别模块分别比较待审特征描述的各分项均落入哪一样本特征知识相应分项的取值范围，由此确定待审特征描述属于哪一样本特征知识对应的特定类型；
    如果待审特征描述的各分项未能均落入找到的样本特征知识相应分项的取值范围，则
    感知特征模糊匹配模块从视频感知知识库中搜索待审特征描述各分项与对应分项次接近的至少一样本特征知识，重复求出所述总距离并比较判断直到确定待审特征描述所属的特定类型，或完成对视频感知知识库中所有样本特征知识的查找。
PCT/CN2013/085484 2012-12-25 2013-10-18 一种基于感知知识库的视频内容审查系统及方法 Ceased WO2014101539A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201210572338.5A CN103064940B (zh) 2012-12-25 2012-12-25 一种基于感知知识库的视频内容审查系统及方法
CN201210572338.5 2012-12-25

Publications (1)

Publication Number Publication Date
WO2014101539A1 true WO2014101539A1 (zh) 2014-07-03

Family

ID=48107570

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2013/085484 Ceased WO2014101539A1 (zh) 2012-12-25 2013-10-18 一种基于感知知识库的视频内容审查系统及方法

Country Status (2)

Country Link
CN (1) CN103064940B (zh)
WO (1) WO2014101539A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113792600A (zh) * 2021-08-10 2021-12-14 武汉光庭信息技术股份有限公司 一种基于深度学习的视频抽帧方法和系统

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103064940B (zh) * 2012-12-25 2016-02-10 深圳先进技术研究院 一种基于感知知识库的视频内容审查系统及方法
CN109783689B (zh) * 2018-12-28 2021-05-21 广州华多网络科技有限公司 信息处理方法、装置及电子设备

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1445696A (zh) * 2002-03-18 2003-10-01 朗迅科技公司 自动检索图像数据库中相似图象的方法
CN101035281A (zh) * 2007-04-19 2007-09-12 鲍东山 分级内容审核系统
US20120128257A1 (en) * 2010-11-19 2012-05-24 Rovi Technologies Corporation Method and apparatus for identifying video program material or content via differential signals
CN103064940A (zh) * 2012-12-25 2013-04-24 深圳先进技术研究院 一种基于感知知识库的视频内容审查系统及方法

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101021857A (zh) * 2006-10-20 2007-08-22 鲍东山 基于内容分析的视频搜索系统

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1445696A (zh) * 2002-03-18 2003-10-01 朗迅科技公司 自动检索图像数据库中相似图象的方法
CN101035281A (zh) * 2007-04-19 2007-09-12 鲍东山 分级内容审核系统
US20120128257A1 (en) * 2010-11-19 2012-05-24 Rovi Technologies Corporation Method and apparatus for identifying video program material or content via differential signals
CN103064940A (zh) * 2012-12-25 2013-04-24 深圳先进技术研究院 一种基于感知知识库的视频内容审查系统及方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113792600A (zh) * 2021-08-10 2021-12-14 武汉光庭信息技术股份有限公司 一种基于深度学习的视频抽帧方法和系统
CN113792600B (zh) * 2021-08-10 2023-07-18 武汉光庭信息技术股份有限公司 一种基于深度学习的视频抽帧方法和系统

Also Published As

Publication number Publication date
CN103064940A (zh) 2013-04-24
CN103064940B (zh) 2016-02-10

Similar Documents

Publication Publication Date Title
WO2020164267A1 (zh) 文本分类模型构建方法、装置、终端及存储介质
WO2020258657A1 (zh) 异常检测方法、装置、计算机设备及存储介质
WO2020082766A1 (zh) 输入法的联想方法、装置、设备及可读存储介质
WO2014101539A1 (zh) 一种基于感知知识库的视频内容审查系统及方法
WO2015003480A1 (zh) 一种社交媒体中的信息推荐方法和装置
WO2018058959A1 (zh) Sql审核方法、装置、服务器及存储设备
WO2020207035A1 (zh) 骚扰电话拦截方法、装置、设备及存储介质
Roach et al. Recent Trends in Video Analysis: A Taxonomy of Video Classification Problems.
WO2010120101A2 (ko) 역 벡터 공간 모델을 이용한 키워드 추천방법 및 그 장치
WO2018205373A1 (zh) 人伤理赔定损费用测算方法、装置、服务器和介质
WO2014146432A1 (zh) 一种冰箱内食品存储控制方法、智能控制终端及冰箱
CN112198385A (zh) 一种非侵入式负荷监测方法
WO2013113236A1 (zh) 一种二次电池的充电装置
WO2017026638A1 (ko) 제품 유형 분류 장치 및 방법
EP3821378A1 (en) Apparatus for deep representation learning and method thereof
WO2012050252A1 (ko) 분류기의 동적 결합에 의한 대용량 분류기 자동 생성 시스템 및 방법
WO2022131490A1 (ko) 딥러닝 기반 이미지 복구 기술을 이용한 패션 이미지 검색 방법 및 장치
WO2022114639A1 (ko) 다차원의 부분집합 연관 분석에 기반한 인공지능 학습 데이터셋 공정성을 보장하는 장치 및 이를 이용한 인공지능 학습 데이터셋 공정성을 보장하는 방법
WO2011162444A1 (ko) 온톨로지 스키마와 결합된 개체명 사전 및 마이닝 규칙을 이용한 개체명 사전 또는 마이닝 규칙 데이터베이스 갱신 장치 및 방법
WO2021051557A1 (zh) 基于语义识别的关键词确定方法、装置和存储介质
WO2022177150A1 (en) Method of personalized image and video searching based on a natural language query, and an apparatus for the same
WO2016074257A1 (zh) 透明广告牌及其智能显示方法
WO2011068315A4 (ko) 최대 개념강도 인지기법을 이용한 최적의 데이터베이스 선택장치 및 그 방법
CN104167211A (zh) 基于分层事件检测和上下文模型的多源场景声音摘要方法
WO2013008978A1 (ko) 개체 식별 결과 검색 시스템 및 방법

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 13868824

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 13868824

Country of ref document: EP

Kind code of ref document: A1