BPO video auditing system based on large model

Abstract

The invention discloses a BPO video auditing system based on a large model, and particularly relates to the technical field of video processing, and the system comprises a video processing server, and a data preprocessing module, a channel construction module, a video distribution module, a violation analysis module, a recheck module and a storage module which are in bidirectional connection with the video processing server. Based on the video distribution rule, the pain point that the long video occupies the auditing channel for a long time is avoided, the auditing speed of the long video is improved, management is facilitated, the user experience is improved, the time length of the problem sub-video can be expanded, on one hand, the content of the problem sub-video is more perfect, and the user experience is improved. And on the other hand, the problem of missing of the sub-video clip content based on the video distribution rule is avoided, the auditing precision is improved, the meaning deviation caused by missing of the defective sub-video content is avoided, auditing personnel can audit the video conveniently, and rechecking personnel do not need to repeatedly call the missing defective sub-video content.

Description

Technical Field

[0001] The invention relates to the technical field of video processing, and in particular to a BPO video auditing system based on a large model. Background Art

[0002] BPO is a business outsourcing some repetitive non-core or core business processes to suppliers in order to reduce costs and improve service quality. With the prosperous development of Internet technology, there are more and more videos on the Internet, especially videos uploaded by users themselves, so there are a lot of violations. Now many websites support and allow users to upload videos and show them to the public. Based on the increase in the number of uploaded videos, websites and other third parties adopt the method of outsourcing the uploaded videos for review and processing. Based on the BPO video review outsourcing business mode, the review efficiency of websites and other third parties has been greatly improved.

[0003] The present invention designs a BPO video review system based on a large model for outsourcing enterprises, which can directly connect and process the videos to be reviewed sent by third-party terminals. Outsourcing enterprises can directly apply the video review system of the present invention to complete comprehensive and efficient review processing of the videos to be processed by third-party terminals.

[0004] A video review method and system is disclosed in the patent application with reference to publication number CN107959863A. The review system constructs an original video screenshot preprocessing method for the same video to be reviewed, and adopts a multi-person and / or multi-level review method, so as to improve the review efficiency, enhance the review reliability, and reduce the review labor intensity.

[0005] The prior art has the following deficiencies:

[0006] The existing video review system can reduce the review intensity and improve the review efficiency by reviewing the same video by multiple people and / or multiple levels, and cooperating with the original video screenshot preprocessing method. However, the review by multiple people and / or multiple levels cannot allocate and process the videos in real time according to the actual situation such as the review efficiency, review time and waiting time, which makes it easy to have the problem of video review accumulation in the review process;

[0007] In addition, when the existing video review system conducts multi-person and / or multi-level review, the video is divided into multiple original video segments and transmitted to multiple people for review. Based on the video allocation rules, there is a problem of missing adjacent interval information of a single illegal video. In the multi-video allocation, the specific circumstances of adjacent original video segments are not covered, which is prone to deviation in meaning, making it impossible for reviewers to accurately judge review violations. Summary of the invention

[0008] The purpose of the present invention is to provide a BPO video audit system based on a large model to solve the above technical problems.

[0009] To solve the above technical problems, the present invention is implemented through the following technical solutions.

[0010] The present invention is a BPO video review system based on a large model, comprising:

[0011] A video processing server and a data preprocessing module, a channel construction module, a video distribution module, a violation analysis module, a re-examination module and a storage module which are bidirectionally connected to the video processing server respectively;

[0012] The data preprocessing module is used to aggregate the videos to be reviewed uploaded by the third-party terminal and identify the total duration of each video to be reviewed {T1, T2, T3...T n};

[0013] The channel construction module builds the review channel based on the videos to be reviewed uploaded by the third-party terminal, and the preset video processing time {S1, S2, S3...S i} to compare and generate a multi-channel review and evaluation path to determine whether to enter the video allocation state of the video allocation module;

[0014] The video allocation module is used to slice the videos to be reviewed that have entered the video allocation state through the analysis and comparison of the channel construction module into k sub-videos according to the total duration, and mark the durations of the k sub-videos as {t1, t2, t3...t k}, and allocate k sub-videos to the corresponding review channels through the video allocation algorithm to complete the review;

[0015] The violation analysis module reviews each video to be reviewed through a video review algorithm in the review channel, generates review rating levels P1, P2 and P3, and performs violation video analysis based on the review rating levels P1, P2 and P3.

[0016] Furthermore, the method for receiving the audit evaluation path includes:

[0017] The total duration of each video to be reviewed uploaded by the third-party terminal {T1, T2, T3...T n} and the preset video processing time {S1, S2, S3...S i}Conduct comparative analysis;

[0018] When T n ≦S1, enter the S1 channel, when S1≦T n ≦S2, enter the S2 channel, when S2≦Tn ≦S3, enter the S3 channel, and so on. n >S i When , the video distribution module enters the video distribution state.

[0019] Furthermore, the video allocation algorithm includes:

[0020] Get the evaluation value U of the video review channel;

[0021] Preset video review channel comparison value bU;

[0022] Combine the evaluation value U with the total duration T of the video data n The corresponding allocation slices are obtained to obtain the appropriate video review channel. According to the preset threshold of the video review channel, the sub-video lengths {t1, t2, t3...t k}Send it to the video review channel for review.

[0023] Furthermore, the method for obtaining the evaluation value U includes:

[0024] Get the mean review time SD of the video review channel, get the number of videos waiting for review W in the video review channel, combine the mean review speed and the average review speed to get the evaluation value U, compare the evaluation value U with the comparison value bU, and determine whether to enter the review channel;

[0025] Intercept x audit videos in the audit channel, and obtain x video audit data {g1, g2, g3...g x The required time is accumulated and averaged to obtain the average review speed of the video review channel;

[0026] The expression of the mean review speed of the video review channel is:

[0027]

[0028] Where SD is the mean review speed of the video review channel, SD jz is the review duration of the zth video review channel;

[0029] The expression of the evaluation value U is:

[0030]

[0031] Where α is the evaluation percentage of the mean SD of the review duration, and β is the evaluation percentage of the number of videos waiting for review W in the video review channel;

[0032] When bU≧U, the sub-video duration {t1, t2, t3...t k}Send it to the video review channel;

[0033] When bU≦U, the sub-video duration {t1, t2, t3...t k}Send it to the video review channel.

[0034] Furthermore, the video review algorithm specifically includes the following steps:

[0035] Image data acquisition and digitization: Each video data to be reviewed is sent to the corresponding review channel according to the processing process of the channel construction module and the video allocation module. After receiving the video to be reviewed, the review channel uses the image recognition algorithm, text recognition algorithm and audio recognition algorithm to perform violation identification processing based on the violation feature database pre-created in the storage module;

[0036] Audit result rating processing: Use a digital analog-to-digital converter to convert digital signals into analog signals, and finally output them to a display or other device, and derive audit rating levels P1, P2 and P3, where P1 represents no violation, P2 represents suspected violation, and P3 represents violation. The suspected violation videos that are judged as P2 but have not entered the video allocation module will be directly re-inspected by the re-inspection module. If the suspected violation videos that are judged as P2 have entered the video allocation module, they will be processed by the sub-video re-inspection processing algorithm and then sent to the re-inspection module for re-inspection.

[0037] Furthermore, the specific processing process of the sub-video recheck processing algorithm is as follows:

[0038] Get the duration of the suspected illegal sub-videos {wt1, wt2, wt3...wt a};

[0039] Based on the duration of the suspected illegal sub-videos of P2 {wt1, wt2, wt3...wt a}Generate the total duration value zt of the problem sub-video;

[0040] The sub-videos corresponding to the total duration value zt of the summarized problematic sub-videos are integrated into a video data packet to be re-inspected and sent to the re-inspection module for re-inspection processing.

[0041] Furthermore, the method for generating the total duration value zt of the problematic sub-video includes:

[0042] Get the duration of the sub-videos of P2 suspected of violating the rules {wt1, wt2, wt3...wt a} any question sub-video duration wt b ;

[0043] Get the wt of any question sub-video bThe duration of the adjacent sub-videos in the set of k sub-videos in the entire slice to be reviewed is t b-1 and t b+1 ;

[0044] Calculate the total duration of the problematic sub-video zt. The specific expression is:

[0045] zt=t b-1 +wt b +t b+1 .

[0046] Furthermore, the wt of any of the problem sub-videos b The duration of the adjacent sub-videos in the set of k sub-videos in the entire slice to be reviewed is t b-1 and t b+1 The corresponding generation process of the suspected illegal video is as follows:

[0047] Get the sub-video duration t respectively b-1 and t b+1 The corresponding slice sub-video duration data is compared with the problem sub-video duration wt b The corresponding sub-videos suspected of violating regulations are integrated into a video data packet to be re-inspected through a video synthesis algorithm, and sent to the re-inspection module for re-inspection.

[0048] Furthermore, the recheck module is used to transmit the generated total duration value of the problematic sub-video to the recheck personnel for rechecking, obtain the review result, and feed back the data to the third-party terminal in combination with the review rating coefficient;

[0049] The method for generating the audit evaluation feedback form includes:

[0050] Get the review data of each video review channel;

[0051] Obtain the relevant audit data of P1 being non-violation, P2 being suspected of violation, and P3 being violation, mark the content involving violation, and generate an audit evaluation feedback form;

[0052] The generated audit evaluation feedback form is transmitted to the third-party terminal.

[0053] Furthermore, the storage module is used to store the pre-created violation feature database and the video audit log data to provide a basis for subsequent data tracing.

[0054] Technical effects and advantages of a BPO video audit system based on a large model of the present invention:

[0055] The present invention obtains the total duration of each video to be reviewed uploaded by a third-party terminal, generates a multi-channel review and evaluation path based on channel construction rules, can distribute the videos to be reviewed for review, avoid the problem of large-scale accumulation and uneven distribution in each review channel, and improve the review efficiency. At the same time, based on the video distribution rules, long videos can be sliced ​​and distributed to each review channel, avoiding the pain point of long videos occupying the review channel for a long time, improving the review speed of long videos, facilitating management and improving user experience, and based on the total duration value generation rule of the problem sub-video, it can extend the duration of the problem sub-video and extract the content of the adjacent sub-videos of the problem sub-video. On the one hand, the content of the problem sub-video is more complete, which is beneficial to subsequent review. On the other hand, it avoids the problem of missing sub-video clip content based on the video distribution rules, improves the review accuracy, avoids the meaning deviation caused by the missing problem sub-video content, facilitates the reviewer to review the video, reduces the difficulty of re-inspection, and does not need the re-inspector to repeatedly retrieve the missing problem sub-video content. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic diagram of a BPO video review system based on a large model provided by the present invention;

[0057] Figure 2 This is a schematic diagram of the BPO video review process provided by the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Example 1: Please refer to Figure 1 , Figure 2 As shown, the BPO video review system based on a large model described in this embodiment is applied to video processing technology, including:

[0060] It includes a video processing server and a data preprocessing module, a channel construction module, a video distribution module, a violation analysis module, a re-examination module and a storage module which are bidirectionally connected with the video processing server respectively;

[0061] The data preprocessing module is used to aggregate the videos to be reviewed uploaded by the third-party terminal. The third-party terminal and the video processing server are connected to each other through the communication module. The communication module specifically uses wireless communication to transmit data, and uniformly transcodes the received videos to be reviewed into a unified format, and adjusts the resolution of the received videos to be reviewed. At the same time, it identifies the total duration of each video to be reviewed {T1, T2, T3...T n}, where T n Indicates the nth video to be reviewed uploaded by the third-party terminal. Each video to be reviewed is sorted according to the total length {T1, T2, T3...T n} mark, which is conducive to the subsequent differentiation and comparison of each video to be reviewed, and convenient for each video to be reviewed to be sent to the corresponding review channel, so that each video to be reviewed can enter the review channel in an orderly manner, improving the review efficiency;

[0062] The channel construction module builds the review channel based on the videos to be reviewed uploaded by the third-party terminal, and the preset video processing time {S1, S2, S3...S i} to generate a multi-channel audit evaluation path to determine whether to enter the video allocation state of the video allocation module, where S i represents the video processing time of the i-th audit channel pre-examination, and S1<S2<S3<......<S i , by setting the preset video processing time in each review channel to {S1, S2, S3...S i}, generate multi-channel audit evaluation paths, and improve audit efficiency;

[0063] The video allocation module is used to slice the videos to be reviewed that have entered the video allocation state through the analysis and comparison of the channel construction module into k sub-videos according to the total duration. The sliced ​​sub-videos adopt the slicing method to separate the long video into different sub-video units, which can be allocated to each review channel to avoid the problem of long waiting period for long video review, and the duration of the sliced ​​k sub-videos is marked as {t1, t2, t3...t k}, and allocate k sub-videos to the corresponding review channels through the video allocation algorithm to complete the review;

[0064] The violation analysis module reviews each video to be reviewed in the review channel through the video review algorithm, generates review rating levels P1, P2 and P3, and analyzes the violation videos based on the review rating levels P1, P2 and P3;

[0065] The methods of receiving the audit assessment path include:

[0066] The total duration of each video to be reviewed uploaded by the third-party terminal {T1, T2, T3...T n} and the preset video processing time {S1, S2, S3...S i}Conduct comparative analysis;

[0067] When T n ≦S1, enter the S1 channel, when S1≦T n ≦S2, enter the S2 channel, when S2≦T n ≦S3, enter the S3 channel, and so on. n >S i When , the video distribution module enters the video distribution state;

[0068] The video distribution algorithms include:

[0069] Get the evaluation value U of the video review channel;

[0070] Preset video review channel comparison value bU;

[0071] Combine the evaluation value U with the total duration T of the video data n The corresponding allocation slices are obtained to obtain the appropriate video review channel. According to the preset threshold of the video review channel, the sub-video lengths {t1, t2, t3...t k} is sent to the video review channel for review, and the duration of the sub-videos after slicing is {t1, t2, t3...t k}Can smoothly enter the preset video processing time {S1, S2, S3...S i} for review and distribution;

[0072] Methods for obtaining the evaluation value U include:

[0073] Get the mean review time SD of the video review channel, get the number of videos waiting for review W in the video review channel, combine the mean review speed and the average review speed to get the evaluation value U, compare the evaluation value U with the comparison value bU, and determine whether to enter the review channel;

[0074] Intercept x audit videos in the audit channel, and obtain x video audit data {g1, g2, g3...g x The required time is accumulated and averaged to obtain the average review speed of the video review channel;

[0075] The expression of the mean review speed of the video review channel is:

[0076]

[0077] Where SD is the mean review speed of the video review channel, SD jz is the review duration of the zth video review channel;

[0078] The expression of the evaluation value U is:

[0079]

[0080] Where α is the evaluation percentage of the mean SD of the review duration, and β is the evaluation percentage of the number of videos waiting for review W in the video review channel;

[0081] When bU≧U, the sub-video duration {t1, t2, t3...t k}Send it to the video review channel;

[0082] When bU≦U, the sub-video duration {t1, t2, t3...t k}Send it to the video review channel;

[0083] Since the evaluation value U in each review channel will change in real time due to the number of waiting times in each review channel, by obtaining the evaluation value U in each review channel, the expected waiting time of the review channel can be estimated, and the sub-video can be accurately assigned to the review channel, which reduces the review waiting time and enables each review channel to conduct the review work smoothly;

[0084] The video review algorithm specifically includes the following steps:

[0085] Image data acquisition and digitization: Each video data to be reviewed is sent to the corresponding review channel according to the processing process of the channel construction module and the video allocation module. After receiving the video to be reviewed, the review channel uses the image recognition algorithm, text recognition algorithm and audio recognition algorithm to perform violation identification processing based on the violation feature database pre-created in the storage module;

[0086] Audit result rating processing: Use a digital analog-to-digital converter to convert digital signals into analog signals, and finally output them to a display or other device, and obtain audit rating levels P1, P2, and P3. P1 means no violation, P2 means suspected violation, and P3 means violation. The suspected violation videos that are determined as P2 and have not entered the video distribution module are directly re-inspected by the re-inspection module. If the suspected violation videos that are determined as P2 and have entered the video distribution module are processed by the sub-video re-inspection processing algorithm, they are sent to the re-inspection module for re-inspection.

[0087] When analyzing violations on a video: first upload the illegal videos, pictures, texts, and audios that need to be screened to the violation feature database;

[0088] Feature extraction: Then, through the image recognition algorithm, text recognition algorithm and audio recognition algorithm, key frame video, text features and audio feature data are extracted from the video data to be reviewed;

[0089] The image recognition algorithm can use convolutional neural network (CNN) to effectively extract all image features in the video data to be reviewed through structures such as convolutional layer, pooling layer and fully connected layer;

[0090] The text recognition algorithm can adopt an OCR algorithm, such as CRNN (convolutional recurrent neural network). CRNN (convolutional recurrent neural network) adopts an end-to-end approach, extracts features through the convolution layer and models the timing information through the recurrent layer, so as to achieve high-precision text recognition and effectively extract all text features in the video data to be reviewed;

[0091] The audio recognition algorithm can use deep learning neural networks to convert sound vibrations into electrical signals, and then into digital signals, converting from time domain signals to frequency domain. Through tools such as Fourier transform, the sound is decomposed into different frequency components, effectively extracting all audio features in the video data to be reviewed;

[0092] Comparison analysis rating: The extracted key frame video, text features and audio feature data are analyzed with the violation feature database through similarity comparison method to determine whether the video content is in violation of the regulations and derive the review rating levels P1, P2 and P3.

[0093] The similarity value of P1 is ao, the similarity value of P2 is bo, the similarity value of P3 is co, and ao <bo<co;

[0094] The similarity value refers to: the number of identical image features, text features, and audio features obtained by comparing all image features, text features, and audio features in the audited video with the violation feature database, and the numerical value based on the proportion of the number of identical image features, text features, and audio features to all image features, text features, and audio features.

[0095] Similarity comparison methods include:

[0096] Obtain all image features in the key frame video, record them as the total number of key frame videos sq, compare all image features of the extracted key frame video with all features in the violation feature database, record the features with the same comparison results as the same features, and obtain the number of the same features sy by counting;

[0097] The key frame video similarity value is obtained based on the same feature quantity value sy and the total number of key frame videos sq:

[0098] The expression of key frame video similarity value is:

[0099]

[0100] Obtain all text features in the audit video, and record them as the total text feature value wq. Compare all text features of the extracted text feature data with all features in the violation feature database, record the features with the same comparison results as the same features, and obtain the number of the same features wy by counting;

[0101] The text feature similarity value is obtained based on the same feature quantity value wy and the total number of text features value wq:

[0102] The expression of text feature similarity value is:

[0103]

[0104] Obtain all audio features in the audit video and record them as the total number of audio features iq. Compare all audio features of the extracted audio feature data with all features in the violation feature database. Record the features with the same comparison results as the same features, and obtain the number of the same features iy.

[0105] The audio feature similarity value is obtained based on the same feature quantity value iy and the total audio feature value iq:

[0106] The expression of audio feature similarity value is:

[0107]

[0108] Obtain the key frame video similarity value sp, the text feature similarity value wp, and the audio feature similarity value ip, add them up and average them to obtain the total similarity value ZP;

[0109] The expression of the total similarity value is:

[0110]

[0111] The total similarity value ZP is compared with the similarity value ao of P1, the similarity value bo of P2 and the similarity value co of P3;

[0112] When ZP≦ao, the output video to be reviewed is the review rating level P1;

[0113] When ao≦ZP≦bo, the video to be reviewed is output as review rating level P2;

[0114] When bo≦ZP≦co, the video to be reviewed is output as review rating level P3.

[0115] In this embodiment, based on the preset video processing time {S1, S2, S3...S i}, through channel construction rules, a multi-channel review and evaluation path is generated, which can distribute the review of videos to be reviewed, avoid the problems of large-scale accumulation and uneven distribution in each review channel, and improve the review efficiency. At the same time, based on the video allocation rules, long videos can be sliced ​​and allocated to each review channel, avoiding the pain point of long videos occupying the review channel for a long time, improving the review speed of long videos, facilitating management and improving user experience.

[0116] Example 2: Please refer to Figure 1 , Figure 2 As shown, the part not described in detail in this embodiment can be found in the description of the first embodiment, which provides a sub-video recheck processing method, including:

[0117] The specific processing process of the sub-video re-check processing algorithm is as follows:

[0118] Get the duration of the suspected illegal sub-videos {wt1, wt2, wt3...wt a};

[0119] Based on the duration of the suspected illegal sub-videos of P2 {wt1, wt2, wt3...wt a}Generate the total duration value zt of the problem sub-video;

[0120] The sub-videos corresponding to the total duration value zt of the problem sub-videos are integrated into a video data packet to be re-inspected and sent to the re-inspection module for re-inspection processing;

[0121] The method for generating the total duration value zt of the problematic sub-video includes:

[0122] Get the duration of the sub-videos of P2 suspected of violating the rules {wt1, wt2, wt3...wt a} any question sub-video duration wt b ;

[0123] Get the wt of any question sub-video b The duration of the adjacent sub-videos in the set of k sub-videos in the entire slice to be reviewed is t b-1 and t b+1 ;

[0124] Calculate the total duration of the problematic sub-video zt. The specific expression is:

[0125] zt=t b-1 +wt b +t b+1 ;

[0126] The wt of any question sub-video b The duration of the adjacent sub-videos in the set of k sub-videos in the entire slice to be reviewed is tb-1 and t b+1 The corresponding generation process of the suspected illegal video is as follows:

[0127] Get the sub-video duration t respectively b-1 and t b+1 The corresponding slice sub-video duration data is compared with the problem sub-video duration wt b The corresponding sub-videos suspected of violations are integrated into a video data packet to be re-inspected through a video synthesis algorithm, and sent to the re-inspection module for re-inspection;

[0128] The wt of a single question sub-video b During the subsequent re-examination, due to the wt of the problematic sub-video b Based on long video slices, the wt of the problem sub-video is b There is a problem of missing information. The missing information will interfere with the review of the re-inspection personnel, resulting in biased meanings and making it impossible to judge the violation factors. Therefore, the sub-video duration t is obtained. b-1 and t b+1 The corresponding slice sub-video duration data is compared with the problem sub-video duration wt b The corresponding suspected illegal sub-videos are integrated into a video data packet to be reviewed through the video synthesis algorithm, so that the wt of a single problematic sub-video is b The data is more complete and the re-inspection accuracy is improved. The video synthesis algorithm is specifically the V-ME video synthesis algorithm, which is based on the diffusion model and video interpolation algorithm to achieve high-quality video generation and integrate it into a video data packet to be re-inspected;

[0129] The re-inspection module is used to transmit the generated total duration value data packet of the problematic sub-video to the re-inspection personnel for re-inspection, obtain the review result, and feed back the data to the third-party terminal in combination with the review rating coefficient;

[0130] The method for generating the audit evaluation feedback form includes:

[0131] Get the review data of each video review channel;

[0132] Obtain the relevant audit data of P1 being non-violation, P2 being suspected of violation, and P3 being violation, mark the content involving violation, and generate an audit evaluation feedback form;

[0133] Transmit the generated audit evaluation feedback form to a third-party terminal;

[0134] The storage module is used to store the pre-created violation feature database and the video audit log data, providing a basis for subsequent data tracing and ensuring the availability, persistence and security of the data. At the same time, the storage module can back up the operation of the audit system to facilitate subsequent queries, and can temporarily store video data packets uploaded by third-party terminals.

[0135] In this embodiment, based on the sub-video recheck processing algorithm, the sub-video after the long video is sliced ​​can be rechecked. At the same time, in order to avoid the sub-video {t1, t2, t3...t k In case of missing information, the total duration of the problematic sub-video can be generated to expand the content of any problematic sub-video, that is, the surrounding sub-video content adjacent to the problematic sub-video, and integrated into a video data packet to be re-inspected through a video synthesis algorithm, and sent to the re-inspection module for re-inspection. On the one hand, it makes the content of the problematic sub-video more complete, which is conducive to subsequent review, and on the other hand, it avoids the problem of missing sub-video content based on video allocation rules, improves the review accuracy, avoids the meaning deviation caused by the missing problematic sub-video content, facilitates reviewers to review the video, reduces the difficulty of re-inspection, and eliminates the need for re-inspectors to repeatedly retrieve missing problematic sub-video content.

[0136] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A BPO video review system based on a large model, characterized in that: include: A video processing server and a data preprocessing module, a channel construction module, a video distribution module, a violation analysis module, a re-examination module and a storage module respectively connected to the video processing server in a bidirectional manner; The data preprocessing module is used to aggregate the videos to be reviewed uploaded by the third-party terminal and identify the total duration of each video to be reviewed {T1, T2, T3...T n }; The channel construction module builds the review channel based on the videos to be reviewed uploaded by the third-party terminal, and the preset video processing time {S1, S2, S3...S i } to compare and generate a multi-channel review and evaluation path to determine whether to enter the video allocation state of the video allocation module; The video allocation module is used to slice the videos to be reviewed that have entered the video allocation state through the analysis and comparison of the channel construction module into k sub-videos according to the total duration, and mark the durations of the k sub-videos as {t1, t2, t3...t k }, and allocate k sub-videos to the corresponding review channels through the video allocation algorithm to complete the review; The violation analysis module reviews each video to be reviewed through a video review algorithm in the review channel, generates review rating levels P1, P2 and P3, and performs violation video analysis based on the review rating levels P1, P2 and P3.

2. A BPO video audit system based on a large model according to claim 1, characterized in that: The method for receiving the audit evaluation path includes: The total duration of each video to be reviewed uploaded by the third-party terminal {T1, T2, T3...T n } and the preset video processing time {S1, S2, S3...S i }Conduct comparative analysis; When T n ≦S1, enter the S1 channel, when S1≦T n ≦S2, enter the S2 channel, when S2≦T n ≦S3, enter the S3 channel, and so on. n >S i When , the video distribution module enters the video distribution state.

3. A BPO video audit system based on a large model according to claim 2, characterized in that: The video distribution algorithm includes: Get the evaluation value U of the video review channel; Preset video review channel comparison value bU; Combined with the evaluation value U, the total length T of the video to be reviewed n The corresponding allocation slices are obtained to obtain the appropriate video review channel. According to the preset threshold of the video review channel, the sub-video duration {t1, t2, t3...t k }Send it to the video review channel for review.

4. A BPO video audit system based on a large model according to claim 3, characterized in that: Methods for obtaining the evaluation value U include: Get the mean review time SD of the video review channel, get the number of videos waiting for review W in the video review channel, combine the mean review speed and the average review speed to get the evaluation value U, compare the evaluation value U with the comparison value bU, and determine whether to enter the review channel; Intercept x audit videos in the audit channel, and obtain x video audit data {g1, g2, g3...g x The required time is accumulated and averaged to obtain the average review speed of the video review channel; The expression of the mean review speed of the video review channel is: Where SD is the mean review speed of the video review channel, SD jz is the review duration of the zth video review channel; The expression of the evaluation value U is: Where α is the evaluation percentage of the mean SD of the review duration, and β is the evaluation percentage of the number of videos waiting for review W in the video review channel; When bU≧U, the sub-video duration {t1, t2, t3...t k }Send it to the video review channel; When bU≦U, the sub-video duration {t1, k2, t3...t k }Send it to the video review channel.

5. The BPO video audit system based on a large model according to claim 1, characterized in that: The video review algorithm specifically includes the following steps: Image data acquisition and digitization: Each video data to be reviewed is sent to the corresponding review channel according to the processing process of the channel construction module and the video allocation module. After receiving the video to be reviewed, the review channel uses the image recognition algorithm, text recognition algorithm and audio recognition algorithm to perform violation identification processing based on the violation feature database pre-created in the storage module; Audit result rating processing: Use a digital analog-to-digital converter to convert digital signals into analog signals, and finally output them to a display or other device, and derive audit rating levels P1, P2 and P3, where P1 represents no violation, P2 represents suspected violation, and P3 represents violation. The suspected violation videos that are judged as P2 but have not entered the video allocation module will be directly re-inspected by the re-inspection module. If the suspected violation videos that are judged as P2 have entered the video allocation module, they will be processed by the sub-video re-inspection processing algorithm and then sent to the re-inspection module for re-inspection.

6. A BPO video audit system based on a large model according to claim 5, characterized in that: The specific processing process of the sub-video re-check processing algorithm is as follows: Get the duration of the suspected illegal sub-videos {wt1, wt2, wt3...wt a }; Based on the duration of the suspected illegal sub-videos of P2 {wt1, wt2, wt3...wt a }Generate the total duration value zt of the problem sub-video; The sub-video durations corresponding to the total duration value zt of the problematic sub-videos are integrated into a video data packet to be re-inspected and sent to the re-inspection module for re-inspection processing.

7. A BPO video audit system based on a large model according to claim 6, characterized in that: The method for generating the total duration value zt of the problematic sub-video includes: Get the duration of the sub-videos of P2 suspected of violating the rules {wt1, wt2, wt3...wt a } any question sub-video duration wt b ; Get the wt of any question sub-video b The duration of the adjacent sub-videos in the set of k sub-videos in the entire slice to be reviewed is t b-1 and t b+1 ; Calculate the total duration of the problematic sub-video zt. The specific expression is: zt=t b-1 +wt b +t b+1 。 8. A BPO video audit system based on a large model according to claim 7, characterized in that: The wt of any of the question sub-videos b The duration of the adjacent sub-videos in the set of k sub-videos in the current slice to be reviewed is t b-1 and t b+1 The corresponding generation process of the suspected illegal video is as follows: Get the sub-video duration t respectively b-1 and t b+1 The corresponding slice sub-video duration data is compared with the problem sub-video duration wt b The corresponding sub-videos suspected of violating regulations are integrated into a video data packet to be re-inspected through a video synthesis algorithm, and sent to the re-inspection module for re-inspection.

9. A BPO video review system based on a large model according to claim 8, characterized in that: The recheck module is used to transmit the total duration value of the generated problematic sub-video to the recheck personnel for rechecking, obtain the review result, and feed back the data to the third-party terminal in combination with the review rating coefficient; The method for generating the audit evaluation feedback form includes: Get the review data of each video review channel; Obtain the relevant audit data of P1 being non-violation, P2 being suspected of violation, and P3 being violation, mark the content involving violation, and generate an audit evaluation feedback form; The generated audit evaluation feedback form is transmitted to the third-party terminal.

10. The BPO video review system based on a large model according to claim 5, characterized in that: The storage module is used to store the pre-created violation feature database and the video audit log data to provide a basis for subsequent data tracing.

Images