A smart management system and method based on video privacy information identification

Abstract

The application discloses a kind of wisdom management system and method based on video privacy information identification, it is related to information processing identification technical field, including: time stamp matching test is carried out to audio of different length, matching test model is established, audio in court video is extracted, audio is divided and optimization division result is obtained, each segment audio after division is converted into text separately, private content in text is identified, the location of private content in audio is found by time stamp matching and is carried out mute, complete audio is synthesized after mute processing, and the division process of subsequent audio to be muted is planned, after frame extraction is carried out to court video image, private information in each frame image is identified and is carried out coding, audio-video is synthesized after coding video and mute processing audio, the probability that mute is not complete even part of private information in audio is not muted due to that time stamp matching deviation is larger is reduced, the reliability of court audio private information processing is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of information processing and recognition technology, specifically to a smart management system and method based on video privacy information recognition. Background Technology

[0002] Making some court hearing audio and video recordings publicly available on the internet can enhance the significance of legal publicity and education. However, privacy information needs to be processed before making them public. After identifying privacy information in the video, it is necessary to mute the privacy information and blur the image to meet the requirements of the security regulations for making court hearing audio and video recordings publicly available. When muting audio, the common practice is to extract the audio from the court hearing video, convert the audio into text, identify the private content in the text, and then use timestamp matching to find the location of the private content in the audio and mute it. However, for long audio clips, the deviation in timestamp matching is relatively large. Existing technology directly converts the entire audio clip into text, which may lead to incomplete muting or even failure to mute some private information in the audio due to large deviations in subsequent timestamp matching. This cannot guarantee the reliability of processing private information in court hearing audio. Summary of the Invention

[0003] The purpose of this invention is to provide a smart management system and method based on video privacy information recognition, so as to solve the problems raised in the prior art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a smart management method based on video privacy information recognition, the method comprising: S1: Perform timestamp matching tests on audio files of different durations and establish a matching test model; S2: Extract audio from the court hearing video, identify the speech rate in the audio, divide the audio according to the speech rate, and perform first-level optimization processing on the division results through audio fusion. After the first-level optimization processing, select whether to perform second-level segmentation optimization processing on the division results based on the matching test model. S3: Convert each segment of audio into text separately, identify private content in the text, find the location of the private content in the audio by timestamp matching and mute it, and combine several mute audio segments into a complete audio; S4: Establish an audio segmentation information database, match the audio to be processed by speech rate segmentation only with the database, where audio to be processed by speech rate segmentation only refers to audio that has not yet undergone segmentation optimization processing, and plan the segmentation process of audio to be processed by speech rate segmentation. S5: After extracting frames from the court hearing video images, identify the privacy information in each frame and blur it out. Then, synthesize the blurred video with the muted audio.

[0005] Preferably, in step S1: Randomly prepare n test audio files with different durations, and keep the speaking speed of the speaker unchanged in each audio file. The duration set of the n test audio files is counted as {t1, t2,..., t n}, record the timestamps of randomly selected m words in each audio file, convert each audio file into a text with timestamps, calculate the deviation values of the timestamps of the m words in the text and the corresponding words in the audio file, and count the maximum timestamp deviation value after each audio file is converted into a text. The maximum timestamp deviation value set is obtained as {T1, T2,..., T n}, T1 represents the maximum value among the deviation values of the timestamps of the m words in the first text and the corresponding words in the audio file, and the test model training data {(t1, T1), (t2, T2),..., (t n , T n )} is obtained. After linearly fitting the training data, a matching test model is established: y = ax + b, where a and b represent the fitting coefficients, x represents the variable in the model that represents the audio duration, and y represents the variable in the model that represents the maximum timestamp deviation value.

[0006] Preferably, in step S2: After authorization, extract the audio from a randomly selected court trial video, identify the speaking speed of the audio, and divide the audio according to the speaking speed: The speaking speed of each segment of the divided audio remains unchanged. It is counted that a total of f segments of audio are divided, and the duration set of each segment of audio is {V1, V2,..., V f}, arrange the f segments of audio in ascending order of duration and randomly divide them into e groups. The duration of each segment of audio in the previous group is less than that in the next group. In a randomly obtained grouping result, the duration mean set of the e groups of audio is {L1, L2,..., L e}, calculate the dispersion M of the durations of the e groups of audio in a randomly obtained grouping result, , calculate the dispersion of the durations of the e groups of audio in different grouping results, and take the grouping result with the largest dispersion as the final grouping result. Perform a first-level optimization on the audio division result: Merge the audio in the first group of the final grouping result with the adjacent audio: The duration of the previous audio segment of a randomly selected audio segment in the first group is A, and the duration of the next audio segment of the randomly selected audio segment is B. If A < B, synthesize the corresponding audio with the previous audio segment; if A > B, synthesize the corresponding audio with the next audio segment; if A = B, synthesize the corresponding audio with the previous or next audio segment. The duration set of each segment of audio after the first-level optimization is obtained as C = {C1, C2,..., C kLet k represent the number of audio segments remaining after the first-level optimization. Input the values ​​from set C one by one into the matching test model: Let x equal the values ​​in set C, and predict the set of maximum timestamp deviation values ​​after converting k audio segments into text as D={C1*a+b,C2*a+b,...,C...}. k *a+b}, where * represents multiplication, sets the timestamp deviation threshold to U, and sets the value of U according to the ITU-T G.114 standard. The absolute value of U is generally set to 125ms. The values ​​in set D are compared with U one by one. If all the values ​​in set D are less than U, no secondary optimization is performed; otherwise, the audio corresponding to the value greater than or equal to U is selected, and the selected audio is further segmented into secondary optimization: the interval reference value is set to (Ub) / a, and for each selected audio segment, it is divided once every (Ub) / a until the division is completed. This invention first divides the extracted court hearing audio according to speech rate variations. Then, considering the randomness of audio length in speech rate-based division, which may include audio that is too short and therefore likely to have incomplete semantics, a first-level segmentation optimization process is performed: the shortest audio is filtered out and then combined with adjacent audio. This reduces the probability that semantic incompleteness due to short audio after only segmenting by speech rate, leading to the inability to identify complete privacy information after converting the corresponding audio into text, is a concern. Furthermore, considering the issue of some audio being too long after the first-level optimization process, the audio... Excessive audio length can lead to incomplete muting when the timestamp mismatch between the subsequent text and audio privacy content accumulates. Therefore, a matching test model is used to screen audio segments that may have excessive timestamp mismatch for further segmentation and optimization. This ensures that the final audio segment is neither too long nor too short, solving the problem that existing technologies, which directly convert the entire audio segment into text, may result in incomplete muting or even the omission of some privacy information in the audio due to large subsequent timestamp mismatches. This ensures the reliability of the processing of privacy information in court audio.

[0007] Preferably, in step S3: each segmented audio is individually converted into text with a timestamp, the feature extraction model is used to identify the privacy content of the text, the location of the privacy content in the audio is found by timestamp matching and muted, and the muted audio is synthesized into a complete audio. Using the feature extraction model to identify the privacy content of the text is an existing technology. For example, when identifying privacy information of the address type, the privacy elements are first defined, that is, the privacy data type to be identified is clearly defined as the address information type. Then, a pre-trained model, such as the BERT model, is used to identify the address in the text, which can realize the identification of privacy information of the address type.

[0008] Preferably, in step S4: an audio segmentation information database is established. After each audio segmentation optimization process is completed, the audio segmentation information before optimization is stored in the audio segmentation information database. The audio segmentation information before optimization is the duration information of each audio segment obtained after segmenting the audio only according to the speech rate. The segmentation information of a random audio segment to be muted, which is completed solely by speech rate segmentation, is obtained: the duration of each audio segment after segmenting the corresponding audio according to the speech rate is obtained, and a first feature matrix is ​​generated based on the duration of each audio segment. g1 represents the duration of the first audio segment; obtain the duration of each segment of a random audio file from the audio segmentation information database before optimization, and generate a second feature matrix based on the duration of each audio segment. Let p represent the number of audio segments after a random audio in the audio segmentation information database is segmented only according to speech rate. The segmentation process for the audio to be censored is planned as follows: Subtract the first feature matrix and the second feature matrix. If the result is a zero matrix, then the audio corresponding to the second feature matrix is ​​determined to match the corresponding audio to be censored; otherwise, the two audios are determined not to match. If the two audios are determined to match, the segmentation process for the corresponding audio to be censored is planned as follows: directly segment the audio to be censored according to the segmentation result after optimization of the audio that matches the audio to be censored. If the two audios are determined to match, the segmentation process for the corresponding audio to be censored is planned as follows: segment the audio according to the process in step S2. After completing multiple audio segmentation optimization processes in step S2, an audio segmentation information database is established. The audio segments to be muted, which are only segmented by speech rate, are matched with the audio segmentation information in the database through matrix operations. When the duration of each audio segment after initial speech rate segmentation is equal to the duration of each audio segment in the database, the audio segments to be muted are directly segmented according to the segmentation results of the matched audio segments. There is no need to use the audio segmentation optimization process in step S2 again, which saves the time spent on audio segmentation and improves the efficiency of privacy information recognition.

[0009] A smart management system based on video privacy information recognition includes: a timestamp matching test module, an audio segmentation and optimization module, a muting module, an image processing module, and an audio-video synthesis module. The timestamp matching test module performs timestamp matching tests on audio of different durations to establish a matching test model. The audio segmentation and optimization module extracts audio from the court hearing video, segments the audio, and optimizes the segmentation results. The muting module identifies and mutes the privacy content in the audio, synthesizes several muted audio segments into a complete audio, and plans the segmentation process for subsequent audio to be muted. The image processing module extracts frames from the court hearing video images, identifies the privacy information in each frame, and redacts the privacy information. The audio-video synthesis module synthesizes the redacted video with the muted audio.

[0010] Preferably, the audio segmentation and optimization module includes an audio segmentation unit and a segmentation result optimization unit. The audio segmentation unit identifies the speech rate in the audio and segments the audio according to the speech rate. The segmentation result optimization unit performs a first-level optimization process on the segmentation result based on speech rate through audio fusion. After completing the first-level optimization process, it selects whether to perform a second-level segmentation optimization process on the segmentation result based on the matching test model.

[0011] Preferably, the noise reduction module includes an audio noise reduction unit and a segmentation process planning unit. The audio noise reduction unit converts the segmented audio into text, identifies private content in the text, finds the location of the private content in the audio through timestamp matching, and performs noise reduction. Several segments of noise-reduced audio are then combined into a complete audio. The segmentation process planning unit establishes an audio segmentation information database. After each audio segmentation optimization process is completed, the audio segmentation information before optimization is stored in the audio segmentation information database. Subsequent audio segments to be noise-reduced, which are segmented only by speech rate, are matched with the information database to plan the segmentation process for subsequent audio segments to be noise-reduced.

[0012] Compared with the prior art, the beneficial effects of the present invention are: This invention first divides the extracted court hearing audio according to speech rate variations. Then, considering the randomness of audio length in speech rate-based division, which may include audio that is too short and therefore likely to have incomplete semantics, a first-level segmentation optimization process is performed: the shortest audio is filtered out and then combined with adjacent audio. This reduces the probability that semantic incompleteness due to short audio after only segmenting by speech rate, leading to the inability to identify complete privacy information after converting the corresponding audio into text, is a concern. Furthermore, considering the issue of some audio being too long after the first-level optimization process, the audio... Excessive audio length can lead to incomplete muting when the timestamp matching deviation between the subsequent text privacy content and the audio privacy content accumulates to a certain extent. Therefore, a matching test model is used to screen out audio that may have excessive timestamp matching deviations and then further segment and optimize it. This ensures that the final audio segment is neither too long nor too short, solving the problem that existing technologies may directly convert the entire audio segment into text, which could lead to incomplete muting or even the failure to mutate some privacy information in the audio due to large subsequent timestamp matching deviations. This ensures the reliability of the processing of privacy information in court audio. After completing multiple audio segmentation optimization processes in step S2, an audio segmentation information database is established. The audio segments to be muted, which are only segmented by speech rate, are matched with the audio segmentation information in the database through matrix operations. When the duration of each audio segment after initial speech rate segmentation is equal to the duration of each audio segment in the database, the audio segments to be muted are directly segmented according to the segmentation results of the matched audio segments. There is no need to use the audio segmentation optimization process in step S2 again, which saves the time spent on audio segmentation and improves the efficiency of privacy information recognition. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a smart management method based on video privacy information recognition according to the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] Example 1: As Figure 1 As shown, this embodiment provides a smart management method based on video privacy information recognition, the method including: S1: Conduct timestamp matching tests on audios of different durations to establish a matching test model: Randomly prepare n test audios with different durations, where the speaking speed of the speaker in each audio remains unchanged. The set of durations of the n test audios is counted as {t1, t2,..., t n}, record the timestamps of randomly selected m words in each audio, convert each audio into text with timestamps, calculate the deviation values between the timestamps of the m words in the text and the corresponding words in the audio, and count the maximum timestamp deviation value after each audio is converted into text. The set of maximum timestamp deviation values is obtained as {T1, T2,..., T n}, where T1 represents the maximum value among the deviation values between the timestamps of the m words in the first text and the corresponding words in the audio, obtaining the training data for the test model {(t1, T1), (t2, T2),..., (t n , T n ). After performing linear fitting on the training data, establish a matching test model: y = ax + b, where a and b represent the fitting coefficients, x represents the variable in the model that represents the audio duration, and y represents the variable in the model that represents the maximum timestamp deviation value; S2: Extract the audio from the court trial video, identify the speaking speed in the audio, divide the audio according to the speaking speed, and perform a primary optimization process on the division result through audio fusion. After the primary optimization process, select whether to perform a secondary optimization process of re-segmentation on the division result according to the matching test model: After obtaining permission, extract the audio from a randomly selected court trial video, identify the speaking speed in the audio, and divide the audio according to the speaking speed: The speaking speed of each segment of the divided audio remains unchanged. It is counted that a total of f segments of audio are divided, and the set of durations of each segment of audio is {V1, V2,..., V f}. Arrange the f segments of audio in ascending order of duration and randomly divide them into e groups, where the duration of each segment of audio in the previous group is less than that in the next group. In a randomly obtained grouping result, the set of average durations of the e groups of audio is {L1, L2,..., L e}, calculate the dispersion M of the durations of the e groups of audio in a randomly obtained grouping result, , calculate the dispersions of the durations of the e groups of audio in different grouping results, and use the grouping result with the largest dispersion as the final grouping result to perform a primary optimization process on the audio division result: Fuse the audio in the first group of the final grouping result with the adjacent audio: Obtain that the duration of the previous audio of a randomly selected audio in the first group is A, and the duration of the next audio of a randomly selected audio is B. If A < B, synthesize the corresponding audio with the previous audio; if A > B, synthesize the corresponding audio with the next audio; if A = B, synthesize the corresponding audio with the previous or next audio; For example: if the duration of the first segment of a random audio segment in the first group is A=120ms and the duration of the second segment is B=50ms, and A>B, then the corresponding audio segment and the second segment of audio segment will be combined. After the first-level optimization, the duration set of each audio segment is C = {C1, C2, ..., C...}. k Let k represent the number of audio segments remaining after the first-level optimization. Input the values ​​from set C one by one into the matching test model: Let x equal the values ​​in set C, and predict the set of maximum timestamp deviation values ​​after converting k audio segments into text as D={C1*a+b,C2*a+b,...,C...}. k *a+b}, where * represents multiplication, sets the timestamp deviation threshold to U, and sets the value of U according to the ITU-T G.114 standard. The absolute value of U is generally set to 125ms. The values ​​in set D are compared with U one by one. If all the values ​​in set D are less than U, no secondary optimization is performed; otherwise, the audio corresponding to the value greater than or equal to U is selected, and the selected audio is further segmented into secondary optimization: the interval reference value is set to (Ub) / a, and for each selected audio segment, it is divided once every (Ub) / a until the division is completed. S3: Convert each segment of audio into text separately, identify the private content in the text, find the location of the private content in the audio through timestamp matching and mute it, and synthesize several mute audio segments into a complete audio; S4: Establish an audio segmentation information database. Match the audio segments to be processed by speech rate segmentation only with the database. Audio segments processed by speech rate segmentation only refer to those that have not yet undergone segmentation optimization. Plan the segmentation process for audio segments to be processed by speech rate segmentation: Establish the audio segmentation information database. After each audio segmentation optimization is completed, store the audio segmentation information before optimization in the database. The audio segmentation information before optimization is the duration of each audio segment obtained after segmenting the audio solely by speech rate. Obtain the segmentation information of a random audio segment to be processed by speech rate segmentation only: Obtain the duration of each audio segment after segmenting the corresponding audio by speech rate, and generate the first feature matrix based on the duration of each audio segment. g1 represents the duration of the first audio segment; obtain the duration of each segment of a random audio file from the audio segmentation information database before optimization, and generate a second feature matrix based on the duration of each audio segment. Let p represent the number of audio segments after a random audio in the audio segmentation information database is segmented only according to speech rate. The segmentation process for the audio to be processed by subsequent noise reduction is planned as follows: Subtract the first feature matrix and the second feature matrix. If the result is a zero matrix, then the audio corresponding to the second feature matrix is ​​determined to match the corresponding audio to be processed by noise reduction. Otherwise, the two audios are determined not to match. If the two audios are determined to match, the segmentation process for the corresponding audio to be processed by noise reduction is planned as follows: directly segment according to the segmentation result after optimization of the matched audio. If the two audios are determined to match, the segmentation process for the corresponding audio to be processed by noise reduction is planned as follows: segment the audio according to the process in step S2. S5: After extracting frames from the court hearing video images, identify and decrypt the privacy information in each frame. Then, synthesize the decrypted video with the muted audio and use the YOLOv8 algorithm to locate privacy regions in the image, such as privacy regions involving identity information, address information, etc.

[0016] Example 2: This example provides a smart management system based on video privacy information recognition. It is implemented based on the management method described in this example, specifically including: a timestamp matching test module, an audio segmentation and optimization module, a muting module, an image processing module, and an audio-video synthesis module. The timestamp matching test module performs timestamp matching tests on audio of different durations to establish a matching test model. The audio segmentation and optimization module extracts audio from the court hearing video, segments the audio, and optimizes the segmentation results. The muting module identifies and mutes the privacy content in the audio, synthesizes several muted audio segments into a complete audio, and plans the segmentation process for subsequent audio to be muted. The image processing module extracts frames from the court hearing video image and identifies the privacy information in each frame, then masks the privacy information. The audio-video synthesis module synthesizes the masked video with the muted audio.

[0017] The audio segmentation and optimization module includes an audio segmentation unit and a segmentation result optimization unit. The audio segmentation unit identifies the speech rate in the audio and segments the audio according to the speech rate. The segmentation result optimization unit performs a first-level optimization process on the segmentation result based on speech rate through audio fusion. After completing the first-level optimization process, it selects whether to perform a second-level segmentation optimization process on the segmentation result based on the matching test model.

[0018] The audio muting module includes an audio muting unit and a segmentation process planning unit. The audio muting unit converts the segmented audio into text, identifies private content in the text, finds the location of the private content in the audio through timestamp matching, and mutes it. Several segments of muted audio are then combined into a complete audio. The segmentation process planning unit establishes an audio segmentation information database. After each audio segmentation optimization process is completed, the audio segmentation information before optimization is stored in the audio segmentation information database. Subsequent audio segments to be muted, which are segmented only by speech rate, are matched with the information database to plan the segmentation process for subsequent audio segments to be muted.

[0019] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A smart management method based on video privacy information recognition, characterized in that: The method includes the following steps: S1: Perform timestamp matching tests on audio files of different durations and establish a matching test model; S2: Extract audio from the court hearing video, identify the speech rate in the audio, divide the audio according to the speech rate, and perform first-level optimization processing on the division results through audio fusion. After the first-level optimization processing, select whether to perform second-level segmentation optimization processing on the division results based on the matching test model. S3: Convert each segment of audio into text separately, identify private content in the text, find the location of the private content in the audio by timestamp matching and mute it, and combine several mute audio segments into a complete audio; S4: Establish an audio segmentation information database, match the audio to be processed by speech rate segmentation with the database, and plan the segmentation process of the audio to be processed by speech rate segmentation. S5: After extracting frames from the court hearing video images, identify the privacy information in each frame and blur it out. Then, synthesize the blurred video with the muted audio.

2. The intelligent management method based on video privacy information recognition according to claim 1, characterized in that: In step S1: Randomly prepare n test audio files of different durations, and count the durations of the n test audio files as {t1, t2, ..., t...} n In each audio file, timestamps of m random words are recorded. Each audio file is converted into timestamped text. The timestamps of the m words in the text are calculated to match the timestamps of the corresponding words in the audio file. The maximum timestamp deviation after converting each audio file into text is calculated, resulting in a set of maximum timestamp deviations {T1, T2, ..., T}. n }, where T1 represents the maximum value among the timestamp deviations between the timestamps of the m words in the first text and the corresponding timestamps in the audio, and the training data for the test model is obtained as {(t1,T1),(t2,T2),...,(t n ,T n After fitting the training data with a straight line, a matching test model is established: y = ax + b, where a and b represent the fitting coefficients, x represents the variable representing the audio duration in the model, and y represents the variable representing the maximum timestamp deviation value in the model.

3. The intelligent management method based on video privacy information recognition according to claim 2, characterized in that: In step S2: After being authorized, randomly extract the audio from a court trial video, identify the speaking speed in the audio, and divide the audio according to the speaking speed: The speaking speed of each segment of the divided audio remains unchanged. It is statistically found that a total of f segments of audio are divided, and the duration set of each segment of audio is {V1, V2,..., V f}, randomly divide the f segments of audio into e groups in ascending order of duration, and in a randomly obtained grouping result, the duration mean set of the e groups of audio is {L1, L2,..., L e}, calculate the dispersion M of the durations of the e groups of audio in a randomly obtained grouping result, , calculate the dispersions of the durations of the e groups of audio in different grouping results, and use the grouping result with the largest dispersion as the final grouping result to perform a first-level optimization on the audio division result: fuse the audio in the first group in the final grouping result with the adjacent audio: Obtain that the duration of the previous audio of a randomly selected audio in the first group is A, and the duration of the next audio of the randomly selected audio is B. If A < B, synthesize the corresponding audio with the previous audio; if A > B, synthesize the corresponding audio with the next audio; if A = B, synthesize the corresponding audio with the previous or next audio.

4. The intelligent management method based on video privacy information recognition according to claim 3, characterized in that: After the first-level optimization, the duration set of each audio segment is C = {C1, C2, ..., C...}. k Let k represent the number of audio segments remaining after the first-level optimization. Input the values ​​from set C one by one into the matching test model: Let x equal the values ​​in set C, and predict the set of maximum timestamp deviation values ​​after converting k audio segments into text as D={C1*a+b,C2*a+b,...,C...}. k *a+b}, set the timestamp deviation threshold to U, compare each value in set D with U, and if all values ​​in set D are less than U, choose not to perform secondary optimization processing; Otherwise, select audio segments with values ​​greater than or equal to U, and perform secondary segmentation optimization on the selected audio segments: set the interval baseline value to (Ub) / a, and for each selected audio segment, perform a division once every (Ub) / a interval until the division is completed.

5. The intelligent management method based on video privacy information recognition according to claim 4, characterized in that: In step S3: Each segment of audio is converted into text with a timestamp, the feature extraction model is used to identify the private content in the text, the location of the private content in the audio is found by timestamp matching and then muted, and the muted audio is synthesized into a complete audio.

6. The intelligent management method based on video privacy information recognition according to claim 5, characterized in that: In step S4: An audio segmentation information database is established. After each audio segmentation optimization process is completed, the audio segmentation information before optimization is stored in the audio segmentation information database. A random audio segmentation information is obtained that needs to be censored and is only segmented by speech rate. The duration of each audio segment after segmenting the corresponding audio by speech rate is obtained, and a first feature matrix is ​​generated based on the duration of each audio segment. g1 represents the duration of the first audio segment; Obtain the duration of each segment of a random audio file from the audio segmentation information database before optimization, and generate a second feature matrix based on the duration of each audio segment. Let p represent the number of audio segments after a random audio in the audio segmentation information database is segmented only according to the speech rate. The segmentation process of the audio to be censored is planned according to the following method: Subtract the first feature matrix and the second feature matrix. If the result is a zero matrix, it is determined that the audio corresponding to the second feature matrix matches the audio to be censored. Otherwise, it is determined that the two audios do not match. If the two audios match, the segmentation process of the audio to be censored is planned as follows: directly segment according to the segmentation result after optimization of the matched audio. If the two audios match, the segmentation process of the audio to be censored is planned as follows: segment the audio according to the process in step S2.

7. A smart management system based on video privacy information recognition, applied to the smart management method based on video privacy information recognition as described in any one of claims 1-6, characterized in that: The system includes: a timestamp matching test module, an audio segmentation and optimization module, a noise reduction module, an image processing module, and an audio-visual synthesis module; The timestamp matching test module is used to perform timestamp matching tests on audio of different durations to establish a matching test model. The audio is extracted from the court hearing video through the audio segmentation and optimization module, and the audio is segmented and the segmentation results are optimized. The noise reduction module identifies and removes private content from the audio, combines several noise-reduced audio segments into a complete audio, and plans the segmentation process for subsequent audio segments to be noise-reduced. The image processing module extracts frames from the court hearing video images and identifies privacy information in each frame, then redacts the privacy information. The audio-video synthesis module combines the censored video with the muted audio.

8. The intelligent management system based on video privacy information recognition according to claim 7, characterized in that: The audio segmentation and optimization module includes an audio segmentation unit and a segmentation result optimization unit. The audio segmentation unit identifies the speech rate in the audio and segments the audio according to the speech rate. The segmentation result optimization unit performs a first-level optimization process on the segmentation result based on speech rate through audio fusion. After completing the first-level optimization process, it selects whether to perform a second-level segmentation optimization process on the segmentation result based on the matching test model.

9. A smart management system based on video privacy information recognition according to claim 8, characterized in that: The noise reduction module includes an audio noise reduction unit and a segmentation process planning unit. The audio noise reduction unit converts the segmented audio into text, identifies the private content in the text, finds the location of the private content in the audio through timestamp matching and performs noise reduction, and combines several noise-reduced audio segments into a complete audio. An audio segmentation information database is established through the segmentation process planning unit. After each audio segmentation optimization process is completed, the audio segmentation information before optimization is stored in the audio segmentation information database. The audio to be processed by speech rate segmentation is matched with the information database to plan the segmentation process of the audio to be processed by speech rate segmentation.

Images